Multi-intervention subject prediction method based on event evolution knowledge

Abstract

The invention relates to a multi-intervention subject prediction method based on event evolution knowledge, and belongs to the technical field of artificial intelligence and natural language processing. According to the method, event evolution information transfer matrixes, including an event predicate transfer matrix, an event subject-predicate incidence matrix and an event object-predicate incidence matrix, are obtained by tracing event cases with similar evolution trends by utilizing the capability of a large language model in the aspect of identifying key node structure information in a historical chain; and obtaining regularized event case evolution information as an event chain hierarchical support of a model generation prediction entity. According to the method, the historical entities highly related to the predicates of the event to be complemented are identified by obtaining the dependency relationship between the entities and the predicates in the data set, and the core function of the predicates in the event triad is played, so that the missing entities in the event chain are predicted more accurately.

Description

Technical Field

[0001] The present invention relates to a case knowledge graph multi-intervention subject prediction method that integrates event evolution trends and historical co-occurrence knowledge, and belongs to the technical field of artificial intelligence and natural language processing.

[0002] Background

[0003] As a collection of human knowledge, knowledge graphs hold great research potential in natural language processing. Traditional knowledge graphs are typically static knowledge bases that utilize a graph-structured data topology to aggregate facts (also known as events) in the form of triples (s, p, o), where s and o represent subject and object entities, respectively, and p, as a relation type, represents the predicate. In the real world, knowledge is constantly evolving, leading to the construction and application of temporal knowledge graphs (TKGs), which add a timestamp t to event triples.

[0004] Temporal Knowledge Graph Completion (TKGC) aims to automatically infer missing links in a knowledge graph by learning from existing facts. Specifically, it aims to infer relation p given a query q, or entity o or s given a query q. TKG provides new perspectives and insights for many downstream applications, such as policy making, stock prediction, and dialogue systems, sparking research interest in TKG reasoning.

[0005] Many researchers have modeled the structural and temporal features of TKG for future event prediction. For example, Ma introduced the first learning method based on embedding representations, which encodes the learned representations to embed quadruple into a hidden latent space to capture event evolution patterns; BoxTE introduced dynamic entity representations of relational dependencies, enabling individual relations to learn different temporal behaviors, thereby effectively capturing temporal reasoning patterns. However, embedding-based methods require separate models to be designed for training on different datasets, and lose information about the semantic aspects of events in TKG. Therefore, even slight dataset modification operations and temporal adaptation modifications are difficult for the model to handle.

[0006] Rule-based approaches use understandable temporal logic rules to predict missing links. For example, AnyBURL extends static rules to temporal logic rules, capturing the path from the query entity to the predicted entity, including timestamps. xERTE and TLogic provide understandable evidence that can explain predictions. They learn reliable rules by learning confidence scores from knowledge graphs. Confidence scores are determined based on the similarity between the rule header and the rule body. The obtained rules can be used for knowledge graph construction and quality assurance. However, rule-based approaches have limited scalability; only similar datasets share similar rules, and they cannot capture long-term dependencies.

[0007] Since timestamped knowledge graphs can be viewed as a series of knowledge graph snapshots / subgraphs, some studies have considered TKGs as time-evolving subgraphs with different relational connections. For example, RTFE uses different Markov models for relations and entities to capture the evolutionary patterns of the knowledge graph, RE-NET models the dynamic evolution of facts in an autoregressive manner, TANGO employs neural ordinary differential equations, and RE-GCN uses multiple graph convolutional layers on each graph snapshot to capture the dependencies of concurrent facts, employing a gated recursive component to effectively learn long-term temporal patterns from historical facts. Typically, facts that occurred before and are relevant to the query are considered its historical context. Methods such as xERTE and TGAP use an attention mechanism that selectively focuses on important aspects. By computing edge attention scores, sampling adjacent historical facts, and iteratively generating interpretable reasoning subgraphs, they automatically capture relevant facts. Another approach to utilizing historical context employs external / domain knowledge as a heuristic or guideline when measuring the relevance of historical facts. CyGNet employs a replication generation mechanism to identify frequently repeated events and aggregates historical facts based on trend scores to enable predictions using more relevant cues.

[0008] Common benchmark datasets for TKGC methods are the ICEWS and GDELT datasets, which contain time-stamped events extracted from the Integrated Crisis Warning System Repository and the Global Event, Language, and Tone Database. These datasets are discrete event quads, and existing TKGC methods are designed, learned, and implemented for these datasets. However, real and widespread real-world events often have relatively complete plots and a series of event development paths. In practical applications, the information that needs to be completed is the key entity node of an event tuple in the complete event chain. Existing methods are mainly suitable for discrete quads. In practical applications, due to the lack of learning event chain knowledge and obtaining event background information, they are affected by noise interference and knowledge gaps, resulting in poor performance. Summary of the Invention

[0009] The purpose of this invention is to address the problems and shortcomings of the existing technology, and to creatively propose a multi-intervention subject prediction method based on event evolution knowledge in order to effectively solve the technical problems of knowledge graph multi-intervention subject prediction.

[0010] This method utilizes the ability of the large language model (LLM) to identify the structural information of key nodes in the historical chain. By integrating the knowledge of event chain evolution and the knowledge of different dimensions of the historical dependency of event triples, it mines temporal logic rules and forms a rule base, improving the model's ability to efficiently obtain and utilize effective information in the dataset, thereby more accurately predicting missing entities in the event chain.

[0011] The innovative points of the present invention include: obtaining the event evolution information transfer matrix by tracing back event cases with similar evolution trends, including the event predicate transfer matrix, the event subject-predicate association matrix, and the event object-predicate association matrix, to obtain regularized event case evolution information as the event chain hierarchical support for the model to generate predictive entities. By obtaining the dependency relationship between entities and predicates in the data set, historical entities that are highly relevant to the event predicates to be completed are identified, and the core role of predicates in event triples is brought into play. By utilizing the ability of large language models to identify the structural information of key nodes in the historical chain, combined with a retrieval strategy based on temporal logic rules and efficient instruction tuning with a small number of parameters, the potential of large language models in the field of temporal knowledge graph completion is explored to obtain a complete entity prediction and completion model based on event evolution knowledge.

[0012] The present invention is implemented by adopting the following technical solutions.

[0013] A multi-intervention subject prediction method based on event evolution knowledge includes the following steps:

[0014] Step 1: Process the original event dataset to generate the event chain dataset.

[0015] For example, we can first use the large language model ChatGPT3.5 to automatically generate response policy texts of some relevant departments based on the security incident text news data in the event datasets EEOK and THUCNews. By integrating the original text information and generation strategies, and after manual screening and processing, we can obtain a reliable security incident case library with event process descriptions and response strategies.

[0016] Then, we extract event triples from each event case (event chain) in the event case library (this can be achieved using the OmniEvent open source event extraction toolkit). We use the relative order of events in the event chain as timestamps to construct a time series event chain dataset.

[0017] Step 2: Perform semantic fusion preprocessing on the event chain dataset.

[0018] To improve the accuracy of event representation in a dataset and reduce redundant representations in the set of entity and event predicate representations, it is necessary to unify the different representations of the same entity in the dataset. Specifically, similarity fusion and coreference resolution can be used to semantically fuse elements in an event chain.

[0019] The BERT model can be used to embed the entity and predicate sets in the dataset, using cosine similarity as a similarity metric to unify global entity representations that meet a threshold similarity. Furthermore, the INCEpTION open-source tool can be used for entity link annotation, unifying co-referencing entities within event chains and completing event chain semantic fusion preprocessing.

[0020] Step 3: Design an event chain evolution information transfer matrix generation module and generate evolution information representation by combining similar event chain information.

[0021] In the transition matrix generation module, in order to efficiently obtain and use the evolution information in the dataset, the top k similar event chains are selected as the knowledge source of the event chain to be completed.

[0022] In this process, the predicate of the triple of the entity to be completed in the event chain is used as the core information. Integrating the event chain structure information, the BERT model can be used to represent the event chain as a whole, using cosine similarity as the semantic similarity indicator. By integrating key predicates, structural similarity information, and semantic similarity information as similarity indicators, a corresponding set of similar event chains is generated.

[0023] Afterwards, the corresponding similar event chains are traversed, and the event-predicate transfer matrix, event subject-predicate association matrix, and event object-predicate association matrix are constructed according to the frequency of occurrence. They respectively represent the event development transfer information, the association information between the event entity as the event subject and the predicate, and the association information between the event entity as the event object and the predicate, thereby regularizing the event case evolution information.

[0024] Step 4: Design an event triplet historical dependency generation module.

[0025] First, the frequencies of all entities associated with the subject or object and the predicate are calculated to capture different kinds of dependencies from two aspects: one is the similarity score vector between the query and the entity set, and the other is the corresponding frequency information of the query.

[0026] Then, for isolated event predicates without historical dependencies, the non-historical information is used as a control for historical dependencies through the transformation of indicator functions to achieve triple-level dependency representation.

[0027] Step 5: Fusion generates logical rules and obtains an entity completion model that integrates event chain evolution knowledge based on the large language model.

[0028] First, a set of temporal logic rules is obtained based on the transition matrix and historical dependencies, and they are ranked according to their confidence scores related to the query. The learned rules are applied and the top k rule bodies are selected to add historical events to the given query.

[0029] The temporal knowledge graph prediction task is then adapted as an instruction task for a large language model, with different components adapted. In addition to the fixed task instructions, the task input is modeled as a chain of historical cases retrieved from a given query, with each fact filled in a template. The query content is similarly represented, allowing the large language model to generate predictions.

[0030] Step 6: Train the neural network and use cross entropy as the loss function to calculate the loss value to achieve self-generation of event chain completion entities.

[0031] Among them, within the set number of training rounds, cross entropy is used to calculate the gap loss between the predicted generated entities and the real entities, and the model parameters are adjusted through back propagation optimization to ultimately achieve the optimal model performance.

[0032] At this point, a model for completing intervention entities that integrates event chain evolution knowledge and historical dependencies has been completed, improving the model's ability to efficiently acquire and utilize effective information in the dataset. The model mines temporal logic rules and forms a rule base, thereby more accurately predicting missing intervention entities in event chains.

[0033] Beneficial effects

[0034] Compared with the prior art, the method of the present invention has the following advantages.

[0035] 1. This method conducts knowledge mining on event chains. By tracing back event cases with similar evolution trends, it obtains the event evolution information transfer matrix, including the event predicate transfer matrix, the event subject-predicate association matrix, and the event object-predicate association matrix. It obtains regularized event case evolution information as the event chain hierarchical support for the model to generate prediction entities.

[0036] 2. This method obtains the dependency relationship between entities and predicates in the dataset, identifies historical entities that are highly relevant to the event predicate to be completed, and gives full play to the core role of predicates in event triples.

[0037] 3. This method utilizes the ability of large language models to identify the structural information of key nodes in the historical chain, combines the retrieval strategy based on temporal logic rules and efficient instruction tuning with a small number of parameters, and explores the potential of large language models in the field of temporal knowledge graph completion. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Flowchart of the method of the present invention.

[0039] Figure 2 This is a diagram showing the implementation framework of the method of the present invention. DETAILED DESCRIPTION

[0040] The present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0041] Example

[0042] The dataset URL used in this example is:

[0043] https: / / github.com / RingBDStack / KGEvetPred

[0044] like Figure 1 As shown in FIG, a multi-intervention subject prediction method based on event evolution knowledge includes the following steps:

[0045] Step 1: Generate event chain dataset.

[0046] Since the existing TKGC dataset is in the form of isolated event quadruples, in order to address the situation in which real events in current practical problems often have complete plots and development paths, it is necessary to process the social security data in the event prediction dataset EEOK and the news text dataset THUCNews to generate an event chain dataset.

[0047] Specifically, the EEOK dataset and the social security data in THUCNews can be read in as a text set. For the text set D = (d1, d2, ..., d D ) Each event description d i , using the generative large language model ChatGPT3.5 to generate the application strategy set M, M = (m1, m2, ..., m D ), after manual review and adjustment, the text information is integrated to obtain a complete social security event case library text dataset Events including event descriptions and response strategies, where Events = (event1, event2, ..., event D ).

[0048] For each event text event in the dataset Events i , through the open source event extraction toolkit OmniEvent, the event chain information is extracted in the form of triples triple = (s, p, o), where s represents the event subject, p represents the event predicate (relation), and o represents the event object, forming the entity set E = (e1, e2, ..., e n ) and the relation set R=(r1,r2,…,r k ), where e n Represents an event entity consisting of an event subject s and an event object o, r k Represents the event predicate p. Finally, the event chain dataset EL is generated, EL=(el1,el2,…,el D), each event chain in the set i Including several event triples, el i =(triple1,triple2,…,triple in ), triple in Represents the triplets that make up an event chain.

[0049] In this embodiment, the constructed data set EEOK includes 221 event chains. The THUCNews security event data includes 2433 event chains.

[0050] Step 2: Perform semantic fusion preprocessing on the event chain dataset.

[0051] Since the original text of the events in the constructed event chain dataset is context-related, the event extraction process also has certain errors. There is a large amount of redundant information in the event entity set and relationship set. For example, "man" and "a man" are used to represent the same entity in the same event chain description, and "a person who treats diseases as a profession" and "doctor" are used in different event chains to represent the common entity "doctor". Noise information will be generated in the entity completion prediction and evaluation process. Therefore, it is necessary to perform semantic fusion preprocessing on the event data before use.

[0052] Specifically, we first embed the entity set E and the relationship set R (which can be implemented using the BERT model), use cosine similarity as the similarity judgment indicator, set a similarity threshold, and unify the global entity representations whose similarity reaches the threshold.

[0053] Then, the event chain is annotated with entity links (which can be achieved using the INCEpTION open source tool), and each event chain el i The internal co-referencing entities are unified to complete the event chain semantic fusion preprocessing.

[0054] Step 3: Design an event chain evolution information transfer matrix generation module and generate evolution information representation by combining similar event chain information.

[0055] Specifically, the method includes the following steps:

[0056] Step 3.1: Construct a set of similar event chains.

[0057] Through key predicate matching, semantic similarity, and structural similarity fusion of event chains, similar event chains in the dataset that match the evolutionary relationship of the event chain to be completed are mined to achieve efficient use of dataset information.

[0058] First, for the current query event chain el i =(triple1,triple2,…,triplein ), the query event triple to be completed is q = (s, r, ?), q = triple in The known subject s and predicate r are key information, and the indicator parameter a is set as flag information. A flag of 1 indicates that the key information is included in other event chains, and a flag of 0 indicates that it is not included.

[0059] Then, in order to obtain the semantic similarity of the event chain, the completed event chain el i The corresponding event original text event i Generate BERT semantic representation, use cosine similarity as the semantic similarity indicator, and set parameter b as the semantic similarity representation.

[0060] Next, we used GCN (Graph Convolutional Networks) to model the structure of the event chain. Using the MVGRL model, we converted entities into nodes and relationships into edges. By comparing and learning representations of nodes and edges, we calculated the structural similarity between event chains, using the parameter c as the structural similarity representation.

[0061] Finally, the three parts of knowledge are combined as the similarity representation of the event chain, and the similarity sim ij The calculation is as follows:

[0062] sim ij =c key a i +c sem b i +c st c i

[0063] Among them, a i is the sign information, b i is the semantic similarity representation, c i is the structural similarity representation; c key 、c sem 、c st They represent the weight parameters of the sign information, semantic similarity representation, and structural similarity representation respectively.

[0064] Query event chain el according to current and to-be-completed i The corresponding similarity ranking is taken and the TOP k similar event chains are recorded as Multiple similar event chains constitute a similar time chain set, recorded as

[0065] Step 3.2: Generate event chain evolution information transfer matrix.

[0066] Query the event chain el for the current completioni The corresponding similar chain set SIM i Mining is carried out to construct an event predicate transfer matrix that measures the development and evolution relationship of event predicates, and a subject / object-predicate association matrix that measures the co-occurrence association relationship between entities as subjects or objects and event predicates.

[0067] Specifically, for similar event chains Generate similar event chain entity sets for the entities and predicates it contains and relationship sets The dimension representing the set of similar entities is d1, The dimension of the similarity relationship set is d2, and its corresponding event predicate transfer matrix M r represents the frequency with which a predicate in a similar event chain and its next event predicate appear together as a context tuple, Subject-predicate association matrix and object-predicate association matrix They represent the frequency with which the entity co-occurs with the predicate as the subject or object of the predicate in similar event chains.

[0068] Then, the entire set of similar event chains is processed, the event evolution information transfer matrix is ​​constructed, and the event evolution pattern corresponding to the query event chain is mined.

[0069] Step 4: Design an event triplet historical dependency generation module.

[0070] In addition to knowledge at the event chain level, non-similar event chains also provide certain prior information at the event triple granularity. For most event triples, especially disposition strategy event tuples, their predicates frequently appear in other event chains. Therefore, common historical collocations contain reliable historical dependency information.

[0071] Specifically, in capturing the historical and non-historical information parts of the event triples, for the query q = (s, r, ?) of the entity to be completed, the historical event set is defined as D s,r , define the corresponding historical entity set as H s,r , not in H s,r Entities in are called non-historical entities:

[0072] D s,r =∪{(s,r,o)∈EL}

[0073] H s,r ={o|(s,r,o)∈D s,r}

[0074] Among them, s represents the event subject, p represents the event predicate (relation), o represents the event object, and EL represents the event chain dataset.

[0075] In most knowledge graphs, although many events often show recurring patterns, new events may have no historical events to refer to. To this end, it is necessary to consider not only historical entities but also non-historical entities. First, during data preprocessing, we study the triples of a given query. in =(s, r, ?) The frequency of historical entities. Specifically, the frequency of all entities that are related entities of subject s and predicate r is calculated. Since the frequency of non-historical entities cannot be calculated, the frequency representation is converted to

[0076] Z s,r (o)=λ(Φ ∑|{o|(s,r,o)∈EL}|>0 -Φ ∑|{o|(s,p,o)∈EL}|＝0 )

[0077] Among them, Φ β is an indicator function that returns 1 if β is true and 0 otherwise. |ε| represents the number of elements. λ represents the function parameter.

[0078] Adopting a learning strategy based on the copy mechanism, according to the input Z s,r (o) Capturing different kinds of dependencies from two aspects: one is the similarity score vector between the query and the entity set, and the other is the corresponding frequency information of the query.

[0079] For historical dependency, a latent context vector is generated for the query q It scores the historical dependencies of different object entities. The calculation formula is:

[0080]

[0081] Among them, tanh is the activation function, ⊕ represents the concatenation operator, W his and b his is a trainable parameter. A linear layer with tanh activation is used to aggregate the query information. The output of the linear layer is then multiplied by E to obtain an |ε|-dimensional vector, where each element represents the similarity score between the corresponding entity o and the query q.

[0082] After that, add Z according to the replication mechanism s,r , directly The index scores of historical entities in Z are improved to higher values ​​without participating in gradient updates. s,r Make Pay more attention to historical entities.

[0083] Similarly, for non-historical dependencies, the latent context vector Defined as:

[0084]

[0085] Among them, tanh is the activation function, E T represents the transpose of E, W nhis and b nhis are represented as trainable parameters.

[0086] The opposite of the historical dependency, minus Z s,r make Focus on non-historical entities.

[0087] By acquiring historical and non-historical dependency information, the dependency relationship representation at the triple granularity is achieved.

[0088] Step 5: Combine the event chain evolution information with the event history knowledge to generate logical rules, and implement intervention subject prediction and completion based on the large language model.

[0089] Step 5.1: In step 3, the event evolution information transfer matrix is ​​obtained, and a confidence score threshold P is set. The evolution information with a co-occurrence frequency higher than the threshold in the evolution matrix is ​​extracted as the evolution rule generation rule base.

[0090] Similarly, for the historical dependency information obtained in step 4, the high-frequency corresponding entities corresponding to each predicate are obtained as historical rules to form a set of temporal logic rules, and the co-occurrence frequency is used as the confidence score to sort the rule reliability.

[0091] Step 5.2: Use a large language model (such as ChatGLM3-6b) as the base model and adjust the time series knowledge graph prediction task to an instruction task of the large language model. During the implementation process, it is necessary to design task instructions and task inputs and specify the task output format.

[0092] The task instructions clearly define the missing event chain intervention subject completion task for the large language model, helping it understand the task definition and providing a clear solution for the large language model to follow and execute. The task input provides the large language model with the query event chain to be completed, and the corresponding event evolution information rules are input as constraints.

[0093] In the task output constraints, the evolution indicators required for the intervention subject completion task are set, namely entity semantic consistency, prediction logic consistency, entity role consistency, event chain context consistency and time consistency. At the same time, the output format of the large language model is restricted, and finally the normalized model prediction entity distribution is obtained.

[0094] Step 6: Use the cross entropy loss function to calculate the true entity distribution P i (Y) and predicted entity distribution The gap between

[0095]

[0096] in, Indicates the number of categories.

[0097] At this point, a model for completing intervention entities that integrates event chain evolution knowledge and historical dependencies has been completed, improving the model's ability to efficiently acquire and utilize effective information in the dataset. The model mines temporal logic rules and forms a rule base, thereby more accurately predicting missing intervention entities in event chains.

[0098] Figure 2 This is the implementation framework diagram of this method.

[0099] Table 1 Comparison of the effects of 10 generation methods on 2 datasets

[0100]

[0101] The performance of this method in completing the missing intervention subjects in the dataset on the HIT@1, HIT@3, HIT@10, and MRR indicators is shown in the last row of Table 1.

Claims

1. A multi-intervention subject prediction method based on event evolution knowledge, characterized in that: The following steps are involved: Step 1: Generate event chain dataset; Step 2: Perform semantic fusion preprocessing on the event chain dataset; Unify different representations of the same entity in the data set and semantically fuse elements in the event chain; Step 3: Design an event chain evolution information transfer matrix generation module, and generate evolution information representation by combining similar event chain information; In the transition matrix generation module, select the top k similar event chains as the knowledge sources of the event chain to be completed; In this process, the predicate of the triple of the entity to be completed in the event chain is used as the core information, the event chain structure information is integrated, the event chain is represented as a whole, and the cosine similarity is used as the semantic similarity indicator; By integrating key predicates, structural similarity information, and semantic similarity information as similarity indicators, a corresponding set of similar event chains is generated; Afterwards, the corresponding similar event chains are traversed, and the event predicate transfer matrix, event subject-predicate association matrix, and event object-predicate association matrix are constructed according to the frequency of occurrence. They represent the event development transfer information, the association information between the event entity as the event subject and the predicate, and the association information between the event entity as the event object and the predicate, respectively, and thus regularize the event case evolution information; Step 4: Design an event triple historical dependency generation module; First, the frequencies of all entities associated with the subject or object and the predicate are calculated to capture different kinds of dependencies from two aspects: one is the similarity score vector between the query and the entity set, and the other is the corresponding frequency information of the query; Then, for isolated event predicates without historical dependencies, the non-historical information is used as a control for historical dependencies through the transformation of indicator functions to achieve triple-level dependency representation. Step 5: Fusion generates logical rules and obtains an entity completion model that integrates event chain evolution knowledge based on the large language model; First, we obtain a set of temporal logic rules based on the transition matrix and historical dependencies, and sort them according to their confidence scores related to the query; we apply the learned rules and select the top k rule bodies to add historical events to the given query; Then, the temporal knowledge graph prediction task is adjusted to an instruction task for a large language model, and different parts are adapted; in addition to the fixed task instructions, the task input is modeled as a chain of historical cases retrieved from a given query, with each fact filled in the template; Step 6: Train the neural network and use cross entropy as the loss function to calculate the loss value, so as to generate event chain completion entities by itself; In particular, within the set number of training rounds, cross entropy is used to calculate the gap loss between the predicted generated entities and the real entities, and the model parameters are adjusted through back-propagation optimization to ultimately achieve the optimal model performance.

2. The multi-intervention subject prediction method based on event evolution knowledge as claimed in claim 1, characterized in that: In step 1, based on the event data set, a large language model is used to automatically generate the response policy texts of some relevant departments. After integrating the original text information and the generated strategy, manual screening and processing are performed to obtain a reliable security incident case library with event process descriptions and response strategies. Then, event triples are extracted for each event case in the event case library; The relative order of events in the event chain is used as the timestamp to construct a temporal event chain dataset.

3. The multi-intervention subject prediction method based on event evolution knowledge as claimed in claim 2, characterized in that: In step 1, the social security data in the event dataset is read in as a text collection; For a text set D = (d1, d2, ..., d D ) Each event description d i , using the generative large language model to generate the application strategy set M, M = (m1, m2, ..., m D ), after manual review and adjustment, the text information is integrated to obtain a complete social security event case library text dataset Events including event descriptions and response strategies, where Events = (event1, event2, ..., event D ); For each event text event in the data set Events i , through the open source event extraction toolkit OmniEvent, the event chain information is extracted in the form of triples triple = (s, p, o), where s represents the event subject, p represents the event predicate (relation), and o represents the event object, forming entity sets E = (e1, e2, ..., e n ) and the relation set R = (r1, r2, ..., r k ), where e n represents the event entity consisting of the event subject s and the event object o, r k Represents the event predicate p; Finally, the event chain dataset EL is generated, EL = (el1, el2, ..., el D ), each event chain in the set i Includes several event triplets, el i =(triple1, triple2,..., triple in ), triple in Represents the triplets that make up a chain of events.

4. The multi-intervention subject prediction method based on event evolution knowledge as claimed in claim 1, characterized in that: In step 2, semantic fusion preprocessing is performed on the event chain dataset through similarity fusion and coreference resolution methods; Firstly, the entity set and predicate set in the dataset are embedded and represented, cosine similarity is used as the similarity evaluation index, a similarity threshold is set, and the global entity representations whose similarity reaches the threshold are unified; Then, the event chains are annotated with entity links, the co-referencing entities within each event chain are unified, and the event chain semantic fusion preprocessing is completed.

5. The multi-intervention subject prediction method based on event evolution knowledge as claimed in claim 1, characterized in that: In step 3, a set of similar event chains is constructed using the following method: First, for the current query event chain el i =(triple1, triple2,..., triple in ), the query event triple to be completed is q = (s, r, ?), q = triple in The known subject s and predicate r are key information, and the indicator parameter a is set as the flag information. If the flag is 1, it means that the key information is included in other event chains, and if the flag is 0, it means that it is not included; Then, in order to obtain the semantic similarity of the event chain, the completed event chain el i The corresponding event original text event i Generate BERT semantic representation, use cosine similarity as the semantic similarity indicator, and set parameter b as the semantic similarity representation; Afterwards, the graph convolutional network uses GCN to model the structure of the event chain by converting entities into nodes and relationships into edges. It learns representations by comparing nodes and edges, calculates the structural similarity between event chains, and uses parameter c as the structural similarity representation. Finally, the three parts of knowledge are combined as the similarity representation of the event chain, and the similarity sim ij The calculation is as follows: sim ij =c key a i +c sem b i +c st c i Among them, a i is the sign information, b i is the semantic similarity representation, c i is the structural similarity representation; c key 、c sem 、c st Respectively represent the weight parameters of the sign information, semantic similarity representation, and structural similarity representation; Query the event chain el according to the current and to-be-completed i The corresponding similarity ranking is taken, and the TOP k similar event chains are recorded as Multiple similar event chains constitute a set of similar time chains, denoted as 6. The multi-intervention subject prediction method based on event evolution knowledge as claimed in claim 1, characterized in that: In step 3, the event chain evolution information transfer matrix is ​​generated using the following method: Query the event chain el for the current completion i The corresponding similar chain set SIM i Conduct mining to construct an event predicate transfer matrix that measures the evolutionary relationship of event predicates, and a subject / object-predicate association matrix that measures the co-occurrence association relationship between entities as subjects or objects and event predicates; For similar chains of events Generate similar event chain entity sets for the entities and predicates they contain and relationship set The dimension representing the set of similar entities is d1, The dimension of the similarity relationship set is d2, and its corresponding event predicate transfer matrix M r represents the frequency with which a predicate in a similar event chain and its next event predicate co-occur as a context tuple, Subject-predicate association matrix and the object-predicate association matrix They represent the frequency with which the entity co-occurs with the predicate as the subject or object of the predicate in similar event chains; Then, the entire set of similar event chains is processed, the event evolution information transfer matrix is ​​constructed, and the event evolution pattern corresponding to the query event chain is mined.

7. The multi-intervention subject prediction method based on event evolution knowledge as claimed in claim 1, characterized in that: In step 4, in capturing the historical and non-historical information parts of the event triples, for the query q = (s, r, ?) of the entity to be completed, the historical event set is defined as D s，r , and define the corresponding historical entity set as H s，r , not in H s，r Entities in are called non-historical entities: D s，r =∪{(s,r,o)∈EL} H s，r ={o|(s,r,o)∈D s，r } Among them, s represents the event subject, p represents the event predicate, o represents the event object, and EL represents the event chain dataset; First, during data preprocessing, we study the triple in = the frequency of the historical entity of (s, r,?); calculate the frequency of all entities that are associated entities of subject s and predicate r; convert the frequency representation into Z s，r (o)=λ(Φ ∑|{o|(s，r，o)∈EL}|＞0 -F ∑|{o|(s，p，o)∈EL}|＝0 ) Among them, Φ β is an indicator function that returns 1 if β is true, otherwise it returns 0; |ε| represents the number of elements; λ represents the function parameter; Adopting a learning strategy based on the copy mechanism, according to the input Z s，r (o) Capturing different kinds of dependencies from two aspects: one is the similarity score vector between the query and the entity set, and the other is the corresponding frequency information of the query; For historical dependency, generate a latent context vector for query q It scores the historical dependencies of different object entities. The calculation formula is: Among them, tanh is the activation function, Represents the concatenation operator, W his and b his is a trainable parameter; a linear layer with tanh activation is used to aggregate the query information; then, the output of the linear layer is multiplied by E to obtain an |ε|-dimensional vector, where each element represents the similarity score between the corresponding entity o and the query q; After that, add Z according to the replication mechanism s，r , directly The index scores of historical entities in Z are improved to higher values ​​without participating in gradient updates; s，r Make a greater focus on historical entities; Similarly, for non-historical dependencies, the latent context vector Defined as: Among them, tanh is the activation function, E T represents the transpose of E, W nhis and b nhis Represented as trainable parameters; The opposite of the historical dependency, minus Z s，r make Focus on non-historical entities; By acquiring historical and non-historical dependency information, the dependency relationship representation at the triple granularity is achieved.

8. The multi-intervention subject prediction method based on event evolution knowledge as claimed in claim 1, characterized in that: In step 6, the cross entropy loss function is used to calculate the true entity distribution P i (Y) and model predicted entity distribution The gap between in, Indicates the number of categories.

Images