An improved script learning method and device based on event evolution graph

Abstract

The present invention relates to an improved script learning method and device based on an event evolution graph. The method comprises: constructing an abstract event evolution graph using a specific event network composed of a specific event chain; mining the event evolution knowledge contained in the event evolution graph to learn the initial embedding representation of context events and candidate events; inputting the initial embedding representation of context events and candidate events into a bidirectional LSTM network to obtain updated embedding representations of context events and candidate events, which contain the time sequence information of the event chain and the event evolution knowledge contained in the event evolution graph; using the updated embedding representation to calculate the similarity of context events and candidate events, obtain the similarity score of each candidate event, probabilize the similarity score, and select the candidate event with the highest probability as the final predicted event. The present invention has a high accuracy for script event prediction and can be used in fields such as privacy theft attack event prediction.

Description

Technical Field

[0001] The present invention relates to a script learning method, and in particular to a script learning method and device based on an event evolution graph, and an application of the method in predicting privacy theft attack events. The method belongs to the subfield of natural language understanding in the field of natural language processing, and also overlaps with privacy protection related technologies. Background Art

[0002] In today's era, natural language understanding is one of the most important tasks in the field of artificial intelligence, among which the understanding of unstructured text is increasingly valued. This is because the real world is full of unstructured texts, the most typical of which is newswire texts. Accurately understanding newswire texts is not only important for dialogue generation, intent recognition, and decision making, but also has a wider range of uses if the large amount of knowledge contained in the text can be well organized and utilized. In specific scenarios, such as privacy theft attack scenarios, using sequence event information extracted from server logs, etc., can predict and perceive possible privacy theft attacks in advance, and take timely measures to avoid privacy data leakage.

[0003] Scripts were first introduced into the field of artificial intelligence by Schank and Abelson (Roger C Schank and Robert P Abelson. Scripts, plans, goals and understanding: an inquiry into human knowledge structures. Technical report, 1977.), who defined it as a set of events consisting of a series of complex events without probabilistic semantics. Later, Chambers and Jurafsky (Nathanael Chambers and Daniel Jurafsky. Unsupervised learning of narrative event chains. In ACL2008, Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics, June 15-20, 2008, Columbus, Ohio, USA, pages 789–797, 2008.) used a co-occurrence-based model to calculate the PMI (Pointwise Mutual Information) of the predicate-GR event pair to characterize the relationship between events. Jans et al. (Bram Jans, Steven Bethard, IvanVulic, and Marie-Francine Moens. Skip n-grams and ranking functions for predicting script events. In EACL 2012,13 th Conference of the European Chapterof the Association for Computational Linguistics, Avignon, France, April 23-27, 2012, pages 336–344, 2012.) used the maximum likelihood method to estimate the probability of Skip N-grams to improve algorithm performance. Pichotta and Mooney (Pichotta, K., and Mooney, R. 2014. Statistical script learning with multi-argument events. In Proceedings of the 14 thConference of the European Chapter of the Association for Computational Linguistics, 220–229.) and Rudinger et al. (Pichotta, K., and Mooney, R. 2014. Statistical script learning with multi-argument events. In Proceedings of the 14 th Conference of the European Chapter of the Association for Computational Linguistics, 220–229.) Script modeling based on the joint probability distribution of co-occurring computational events.

[0004] The above methods all use statistics or probability to characterize event relationships for script learning. Due to the sparsity of events, the effectiveness of their algorithm models is limited. Granroth-Wilding and Clark (Mark GranrothWilding and Stephen Clark. What happens next? Event prediction using a compositional neural network model. In Proceedings of the Thirtieth AAAI Conference on ArtificialIntelligence, February 12-17, 2016, Phoenix, Arizona, USA, pages 2727–2733, 2016.) began to introduce event embedding into the field of script learning. They used neural networks to simultaneously learn the embedding combination of event predicates and arguments as the embedding representation of events, and then used a siamese network to calculate the coherence score of event pairs. In the subsequent work of Pichotta and Mooney (Pichotta, K., and Mooney, RJ2016. Learning statistical scripts with LSTM recurrent neural networks. In AAAI, 2800–2806. and Pichotta, K., and Mooney, RJ2016. Using sentence-level LSTM language models for script inference. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers.), the long short-term memory network LSTM (Long Shor-Term Memory) was introduced to model the temporal relationship between event sequences to better encode script information, but it may be troubled by the overfitting problem.Wang et al. (Zhongqing Wang, Yue Zhang, and Ching-Yun Chang. Integrating order information and event relation for script event prediction. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, EMNLP 2017, Copenhagen, Denmark, September 9-11, 2017, pages 57–67, 2017.) combined LSTM and deep memory network (Deep Memory network) on its basis, and achieved better results by recursively modeling a more abstract representation of the scene to refine the event weight and event relationship calculation.

[0005] After fully mining the script information encoding based on event pairs and event chains, Li et al. (Zhongyang Li, Xiao Ding, and Ting Liu. Constructing narrative event evolutionary graph for script event prediction. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, July 13-19, 2018, Stockholm, Sweden, pages 4201-4207, 2018.) innovatively proposed using event evolutionary graph EEG (Event Evolutionary Graph) for script learning, and used a scalable graph neural network SGNN (Scaled Graph Neural Network) to model the rich semantic information in the graph to learn the embedding representation of events, fully mining the script information and greatly improving the model effect. Lv et al. (S. Lv, W. Qian, L. Huang, J. Han and S. Hu, SAM-Net: Integrating Event-Level and Chain-Level Attentions to Predict What Happens Next, In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, Honolulu, Hawaii, USA, January 27-February 1, 2019, page 6802-6809, 2019.) proposed the concept of event segment, defining it as a collection of continuous or discontinuous events in the script that are closely related to each other. They used the relationship between event segments and candidate events to learn and predict scripts, and used DenseNet to extract event chain features to remove redundant information, so as to avoid insufficient utilization of semantic information and introduce a large amount of noise. Therefore, using the EEG event evolution graph to organize knowledge, using complex networks to encode event chain information, and fully mining the interactive information between context events and between context events and candidate events are of great significance for script learning. Summary of the invention

[0006] The purpose of the present invention is to optimize the script learning method based on the event evolution graph through the deep fusion of event chain information. On the basis of the original script learning method based on the event evolution graph, the event chain information is fully utilized to better mine the interaction information between events and obtain better prediction results.

[0007] The technical solution adopted by the present invention is as follows:

[0008] An improved script learning method based on event evolution graph includes the following steps:

[0009] Using the concrete event network composed of concrete event chains, an abstract event evolution graph is constructed;

[0010] Mining the event evolution knowledge contained in the event evolution graph to learn the initial embedding representation of context events and candidate events (initial embedding representation);

[0011] Input the initial embedding representations of context events and candidate events into the bidirectional LSTM network to obtain updated embedding representations of context events and candidate events, which contain the temporal information of the event chain and the event evolution knowledge contained in the event evolution graph;

[0012] The updated embedding representation is used to calculate the similarity between the context event and the candidate event to obtain the similarity score of each candidate event. The similarity score is probabilistically converted to obtain the probability corresponding to each candidate event, and the candidate event with the highest probability is selected as the final predicted event.

[0013] Furthermore, the use of a specific event network composed of specific event chains to construct an abstract event evolution graph includes:

[0014] Event chain extraction: Use natural language processing tools to perform dependency parsing and coreference resolution on unstructured text to obtain all verbs and related dependencies related to a certain protagonist, and select the longest event chain related to a certain protagonist as the initial event chain, where each event is in the form of a four-tuple (predicate, subject, object, prepositional object).

[0015] Generalization of specific events: All verbs are converted to lowercase and morphologically restored, and the predicates in specific events are classified into synonyms using a semantic dictionary. For verbs with the same semantics, a representative word is selected as the representation;

[0016] Abstract event linking: Based on the network composed of the extracted specific event chains, the generalized event nodes are linked to obtain the event evolution graph.

[0017] Furthermore, the method of mining the event evolution knowledge contained in the event evolution graph to learn the initial embedding representation of context events and candidate events includes:

[0018] The event evolution subgraphs involved in context events and candidate events are serialized into event sequences using a topological sorting algorithm and input into a pre-trained language model. The vectors output by the pre-trained language model are decomposed to serve as the initial embedding representations of context events and candidate events.

[0019] Furthermore, the inputting of the initial embedding representations of the context events and the candidate events into the bidirectional LSTM network to obtain updated embedding representations of the context events and the candidate events includes:

[0020] The initial embedding representation of the context event is input into a multi-layer bidirectional LSTM, and the two hidden states corresponding to the same event are concatenated to obtain the updated embedding representation of the context event, that is, the hidden state vector matrix of the context event is obtained;

[0021] The initial embedding representation of each candidate event is added to the end of the bidirectional LSTM for encoding and concatenated with the original vector to obtain the hidden state vector matrix of the candidate event.

[0022] Furthermore, the similarity calculation of the context event and the candidate event using the updated embedding representation to obtain a similarity score for each candidate event includes:

[0023] The hidden state vector matrix of the context event is input into the information condensation network for information fusion. The obtained vector is used as the overall representation vector of the context. The similarity between it and the candidate event is calculated to obtain the score of each candidate event.

[0024] Furthermore, the probabilizing the similarity scores is to probabilize the similarity scores using a softmax function.

[0025] An improved script learning device based on event evolution graph using the method of the present invention comprises:

[0026] The event evolution graph construction module is used to construct an abstract event evolution graph using a specific event network composed of specific event chains;

[0027] The initial embedding representation acquisition module is used to mine the event evolution knowledge contained in the event evolution graph to learn the initial embedding representation of context events and candidate events;

[0028] The embedding representation update module is used to input the initial embedding representation of the context event and the candidate event into the bidirectional LSTM network to obtain the updated embedding representation of the context event and the candidate event, which contains the temporal information of the event chain and the event evolution knowledge contained in the event evolution graph;

[0029] The event prediction module is used to calculate the similarity between context events and candidate events using the updated embedding representation, obtain the similarity score of each candidate event, probabilize the similarity score to obtain the probability corresponding to each candidate event, and select the candidate event with the highest probability as the final predicted event.

[0030] The beneficial effects of the present invention are:

[0031] The improved script learning method based on event evolution graph of the present invention can fully mine the contextual events and the interaction information between them and the candidate events, taking into account the inherent logic of the event scene and the interaction information between the specific contextual context and the candidate events. It has a high accuracy in script event prediction and can be used in fields such as privacy theft attack event prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is the overall structure diagram of the model. DETAILED DESCRIPTION

[0033] The method of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0034] The present invention firstly builds the event evolution graph EEG based on the work of Li et al. and draws on the method in the work of Zhao et al. (Sendong Zhao, Quan Wang, Sean Massung, Bing Qin, Ting Liu, Bin Wang, and ChengXiang Zhai. Constructing and embedding abstract event causality networks from text snippets. In Proceedings of the Tenth ACM International Conference on Web Search and Data Mining, WSDM2017, Cambridge, United Kingdom, February 6-10, 2017, pages 335–344, 2017.). The processing flow includes the following steps:

[0035] 1) Event chain extraction: First, use natural language processing tools to perform dependency parsing and coreference resolution on unstructured text to obtain all verbs and related dependencies related to a certain protagonist. The related dependencies are assumed to be the subject, object, and object of the preposition. If there are any, they will be obtained. If there are no dependent elements, they will be marked as NULL. The longest event chain related to a certain protagonist is selected as the initial event chain, where each event is in the form of a four-tuple (predicate, subject, object, preposition object).

[0036] 2) Generalization of specific events: In order to avoid the problem of sparse events in extraction, generalization of events is performed in this step. All verbs are converted to lowercase and morphologically restored. The predicates that appear in the specific events extracted in the previous step are classified as synonyms using a semantic dictionary. For verbs with the same semantics, a representative word is selected as the representation. The predicate-GR form proposed by Granroth-Wilding and Clark can also be used for abstract representation.

[0037] 3) Abstract event link: Based on the network composed of the extracted specific event chains, the generalized event nodes are linked to obtain the final required event evolution graph. The weight in the event evolution graph is the probability that the event represented by the tail node in the event chain occurs after the event represented by the head node.

[0038] Next, we use the constructed EEG to represent events, that is, we mine the event evolution knowledge contained in the EEG to learn the initial embedding of the event. This step draws on the work of Lv et al. (Lv S, Guo D, Xu J, et al. Graph-Based Reasoning over Heterogeneous External Knowledge for Commonsense Question Answering. CoRR, abs / 1909.05311, 2019.). Using the topological sorting algorithm proposed by them, we serialize the EEG subgraphs involved in the context events and candidate events into event sequences using the topological sorting algorithm, input them into the pre-trained language model, and decompose the output vector as the initial embedding representation of the event.

[0039] After obtaining the initial embeddings of the context event and the candidate event, the embedding of the context event is input into a multi-layer bidirectional LSTM, and the two hidden states corresponding to the same event are concatenated to obtain the updated context event embedding representation, that is, the hidden state vector matrix of the context event is obtained to encode the sequence information of the context event. For the candidate event, each candidate event is added to the end of the bidirectional LSTM for encoding and concatenated with the original vector to obtain the hidden state vector matrix of the candidate event.

[0040] Finally, the hidden state vector matrix of the context event is input into the information condensation network (such as DenseNet, MLP and other network structures) for information fusion. The obtained vector is used as the overall representation vector of the context. The similarity between it and the candidate event is calculated to obtain the score of each candidate event. The score is probabilized using the softmax function to obtain the probability corresponding to each candidate event. The candidate event with the highest probability is selected as the final choice, that is, the final predicted event. The overall structure of the model is as follows: Figure 1 shown.

[0041] A specific application example is provided below. This example is an example analysis using English news line text as corpus. It is only the application of the improved script learning method based on event evolution graph proposed by the present invention in news line text, which can fully illustrate the effect of the method. However, the present invention proposes a general framework. In addition to this example, the method can be applied to script knowledge learning of other unstructured texts. Therefore, other examples proposed based on the framework of the present invention all fall within the protection scope of the present invention.

[0042] Assume that the English newswire text set we use is T. We use Stanford CoreNLP to perform dependency parsing and coreference resolution on it, and get all the verbs and related dependencies around the protagonist T, for example: {(walks,T,restaurant,null),(seats,T,null,null),(reads,T,menu,null),(orders,T,food,null),(serves,waiter,food,T),(eats,T,food,null),(pays,T,null,null),(leaves,T,restaurant,null)}, where T represents the protagonist, and the lack of corresponding dependent components is represented by null. Take the event (serves,waiter,food,T) as an example. Specifically, the predicate of the four-tuple is serves, the subject related to it is waiter, the object is food, and the prepositional object is the protagonist T of the event chain.

[0043] For specific event generalization, we also first use natural language processing tools to convert the predicates in the quadruple extracted in the previous step to lowercase and restore their morphology. For example, (walks, T, restaurant, null) becomes (walk, T, restaurant, null). Then, we can use semantic dictionaries such as FrameNet and VerbNet to cluster predicates with the same semantics. Then, we further abstract the quadruple into predicate-GR form.

[0044] Next, we use the concrete event network composed of concrete event chains to construct an abstract event evolution graph, with the focus on the connection between concrete events and abstract events. Specifically, we do this: when there is an edge connecting two nodes in the concrete event network (that is, there is a sequential relationship between the two concrete events), there is also an edge connecting the abstract event nodes corresponding to the two concrete event nodes, with the same direction. Formally, the concrete event network is The specific event network is formed by connecting the extracted specific event chains. It is a four-tuple (v, s, o, po) in form, where v, s, o, po represent the predicate, subject, object, and prepositional object respectively. Indicates that node i occurs before node k. Abstract event evolution graph Node is a tuple (v, dep), where dep represents the dependency relationship between the subject and the predicate, and the edge Exists if and only if exists, and p in i,q in j (that is, nodes i and j are generalization nodes of nodes p and q respectively).

[0045] After obtaining the event evolution graph, the context events and candidate events in the problem are converted into binary forms, and the involved event evolution subgraphs are found in the event evolution graph. The topological sorting algorithm proposed by Lv et al. is used for serialization to obtain a series of event sequences denoted as L A , according to the correspondence between concrete and abstract events, L A The event binary in is restored to a quaternary event sequence and recorded as L S , and input it into a pre-trained language model such as XLNet. The output vector is decomposed as the initial embedding representation of context events and candidate events, denoted as Ech = {e1, e2, …, e n} and Eca={e c1 ,e c2 ,…,e ck}, where n represents the number of context events and k represents the number of candidate events.

[0046] In order to model the temporal relationship between contextual events, Ech is input into a two-layer bidirectional LSTM network, and the hidden states are concatenated to obtain the hidden state vector matrix of the contextual event Ech′={h1,h2,…,h n}, each candidate event e cj Add to the end of the bidirectional LSTM network, perform the same encoding and concatenate with itself to obtain the hidden state vector matrix Eca′={h c1 ,h c2 ,…,h ck}, formalized as follows:

[0047]

[0048]

[0049]

[0050]

[0051]

[0052]

[0053] Among them, h i represents the hidden state representation of the i-th context event, and Respectively represent e i The output of LSTM in both directions, e i represents the i-th context event; h cj represents the hidden state representation of the jth candidate event, and Indicates h cj The two components of Indicates the nth and last context event e n The forward LSTM output, e cj represents the jth candidate event.

[0054] The updated embeddings of context events and candidate events are obtained above, which fully absorb the temporal information of the event chain and the knowledge contained in the event evolution graph.

[0055] Finally, the hidden state vector matrix Ech′ of the context event is input into the information condensation network for information fusion, and the vector Ech″ is obtained as the overall representation vector of the context event chain. Its relationship with each candidate event h is calculated. cj The correlation score, also known as the similarity score, is calculated as follows:

[0056] s(e i ,e cj )=W ch ·Ech″+W ca ·h cj +b

[0057] Among them, W ch , W ca represents the coefficient and b represents the bias term.

[0058] Finally, the softmax function is used to probabilize the scores and obtain the probability corresponding to each candidate, which is recorded as P(e cj |Ech).

[0059] The candidate event with the highest probability is selected as the final choice, that is, the final predicted event.

[0060]

[0061] J = arg j max{P(e cj |Ech)}

[0062] Therefore, the Jth candidate is the predicted answer, and the training function is:

[0063]

[0064] Among them, P(e cj |Ech) is the predicted probability corresponding to the correct answer, N is the number of training samples, λ is the L2 regularization parameter, and θ is the model parameter.

[0065] It can be seen from the examples in the above specific implementation methods that the improved script learning method based on event evolution graph of the present invention can fully mine the context events and the interaction information between them and the candidate events, and has a high accuracy in script event prediction.

[0066] Based on the same inventive concept, another embodiment of the present invention provides an improved script learning device based on event evolution graph using the method of the present invention, which comprises:

[0067] The event evolution graph construction module is used to construct an abstract event evolution graph using a specific event network composed of specific event chains;

[0068] The initial embedding representation acquisition module is used to mine the event evolution knowledge contained in the event evolution graph to learn the initial embedding representation of context events and candidate events;

[0069] The embedding representation update module is used to input the initial embedding representation of the context event and the candidate event into the bidirectional LSTM network to obtain the updated embedding representation of the context event and the candidate event, which contains the temporal information of the event chain and the event evolution knowledge contained in the event evolution graph;

[0070] The event prediction module is used to calculate the similarity between context events and candidate events using the updated embedding representation, obtain the similarity score of each candidate event, probabilize the similarity score to obtain the probability corresponding to each candidate event, and select the candidate event with the highest probability as the final predicted event.

[0071] Based on the same inventive concept, another embodiment of the present invention provides an electronic device (computer, server, smart phone, etc.), which includes a memory and a processor, the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing each step in the method of the present invention.

[0072] Based on the same inventive concept, another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, CD), which stores a computer program. When the computer program is executed by a computer, it implements the various steps of the method of the present invention.

[0073] The above examples are only used to illustrate the technical solutions of the present invention rather than to limit the same. A person skilled in the art may modify or make equivalent substitutions for the technical solutions of the present invention without departing from the spirit and scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. An improved script learning method based on event evolution graph, characterized in that: The following steps are involved: Using the concrete event network composed of concrete event chains, an abstract event evolution graph is constructed; Mining the event evolution knowledge contained in the event evolution graph to learn the initial embedding representation of context events and candidate events; Input the initial embedding representations of context events and candidate events into the bidirectional LSTM network to obtain updated embedding representations of context events and candidate events, which contain the temporal information of the event chain and the event evolution knowledge contained in the event evolution graph; The updated embedding representation is used to calculate the similarity between the context event and the candidate event to obtain the similarity score of each candidate event. The similarity score is probabilistically converted to obtain the probability corresponding to each candidate event, and the candidate event with the highest probability is selected as the final predicted event. The event evolution knowledge contained in the event evolution graph is mined to learn the initial embedding representation of context events and candidate events, including: The event evolution subgraphs involved in the context events and candidate events are serialized into event sequences using a topological sorting algorithm and input into a pre-trained language model. The vectors output by the pre-trained language model are decomposed to serve as the initial embedding representations of the context events and candidate events. The step of inputting the initial embedding representations of the context events and the candidate events into a bidirectional LSTM network to obtain updated embedding representations of the context events and the candidate events includes: The initial embedding representation of the context event is input into a multi-layer bidirectional LSTM, and the two hidden states corresponding to the same event are concatenated to obtain the updated embedding representation of the context event, that is, the hidden state vector matrix of the context event is obtained; The initial embedding representation of each candidate event is added to the end of the bidirectional LSTM for encoding and concatenated with the original vector to obtain the hidden state vector matrix of the candidate event.

2. The method according to claim 1, characterized in that: The specific event network composed of specific event chains is used to construct an abstract event evolution graph, including: Event chain extraction: Use natural language processing tools to perform dependency parsing and coreference resolution on unstructured text to obtain all verbs and related dependencies related to a certain protagonist, and select the longest event chain related to a certain protagonist as the initial event chain, where each event is in the form of a four-tuple, which is a predicate, a subject, an object, and a prepositional object; Generalization of specific events: All verbs are converted to lowercase and morphologically restored, and the predicates in specific events are classified into synonyms using a semantic dictionary. For verbs with the same semantics, one word is selected as the representation; Abstract event linking: Based on the network composed of the extracted specific event chains, the generalized event nodes are linked to obtain the event evolution graph.

3. The method according to claim 2, characterized in that The network formed based on the extracted specific event chain links the generalized event nodes to obtain an event evolution graph, including: When there is an edge connecting two nodes in a specific event network, there is also an edge connecting the abstract event nodes corresponding to the two specific event nodes, with the same direction; The specific event network is ,The specific event network is formed by connecting the extracted specific event chains, and the nodes It is a quadruple form, in which They represent predicate, subject, object, prepositional object, and side Represents a node Before the node occur; The abstract event evolution diagram is , node It is a binary ,in dep Indicates the dependency relationship between the protagonist and the predicate, and the edge Exists if and only if exists, and , i.e. node and Node and The generalization node of .

4. The method according to claim 1, characterized in that: The similarity calculation of the context event and the candidate event using the updated embedding representation to obtain a similarity score for each candidate event includes: The hidden state vector matrix of the context event is input into the information condensation network for information fusion. The obtained vector is used as the overall representation vector of the context. The similarity between it and the candidate event is calculated to obtain the score of each candidate event.

5. The method according to claim 1, characterized in that: The probabilistic conversion of the similarity scores is to use a softmax function to probabilistically convert the similarity scores.

6. An improved script learning device based on event evolution graph using the method described in any one of claims 1 to 5, characterized in that: include: The event evolution graph construction module is used to construct an abstract event evolution graph using a specific event network composed of specific event chains; The initial embedding representation acquisition module is used to mine the event evolution knowledge contained in the event evolution graph to learn the initial embedding representation of context events and candidate events; The embedding representation update module is used to input the initial embedding representation of the context event and the candidate event into the bidirectional LSTM network to obtain the updated embedding representation of the context event and the candidate event, which contains the temporal information of the event chain and the event evolution knowledge contained in the event evolution graph; The event prediction module is used to calculate the similarity between the context event and the candidate event using the updated embedding representation to obtain the similarity score of each candidate event, probabilize the similarity score to obtain the probability corresponding to each candidate event, and select the candidate event with the highest probability as the final predicted event; The event evolution knowledge contained in the event evolution graph is mined to learn the initial embedding representation of context events and candidate events, including: The event evolution subgraphs involved in the context events and candidate events are serialized into event sequences using a topological sorting algorithm and input into a pre-trained language model. The vectors output by the pre-trained language model are decomposed to serve as the initial embedding representations of the context events and candidate events. The step of inputting the initial embedding representations of the context events and the candidate events into a bidirectional LSTM network to obtain updated embedding representations of the context events and the candidate events includes: The initial embedding representation of the context event is input into a multi-layer bidirectional LSTM, and the two hidden states corresponding to the same event are concatenated to obtain the updated embedding representation of the context event, that is, the hidden state vector matrix of the context event is obtained; The initial embedding representation of each candidate event is added to the end of the bidirectional LSTM for encoding and concatenated with the original vector to obtain the hidden state vector matrix of the candidate event.

7. An electronic device, characterized in that: The invention comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program comprises instructions for executing the method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a computer, the method described in any one of claims 1 to 5 is implemented.

Images