An event context link generation method fusing news occurrence time and semantic similarity

Abstract

The application discloses a method for generating event context links by fusing news occurrence time and semantic similarity, and comprises the following steps: carrying out embedded representation on news text content, digitally representing news occurrence time, and completing news set preprocessing; dividing the preprocessed news set into several news sub-sets according to occurrence time; combining the occurrence time and text content information of news, and using a clustering algorithm to calculate event cluster results and evaluate the event cluster results in each window; selecting clusters belonging to the same category from the event cluster results calculated under each window, fusing the clusters, and selecting event cluster representative nodes from the fused cluster results to form a new round of news set; repeating the above process for the newly formed news set until the final event context result is obtained. The application improves the accuracy, interpretability and efficiency of the generation of news event context results.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, specifically involving a method for generating event context links by integrating the time of news occurrence and semantic similarity. Background Technology

[0002] With the rapid development of the internet, all kinds of news events are emerging one after another, making it increasingly difficult to mine and analyze news events from the ever-increasing amount of data. Therefore, sorting out the context of events from massive amounts of news data is very important for the analysis and research of news events.

[0003] In existing technologies [1,2,3,4], the methods for generating event contexts for news articles, whether calculating the connections between news items based on words or the interactions between news items based on sentences, essentially only analyze from the perspective of textual semantic information. By calculating the similarity between news events and then simply sorting semantically similar news items according to their occurrence time, the context of the news events is obtained.

[0004] Generally, news events on the same topic follow a predictable pattern in timeline. However, current event context generation methods only consider semantic information, treating the occurrence time of news events merely as a sorting tool, neglecting the impact of occurrence time on similarity measurement and event evolution. Furthermore, existing context generation schemes lack interpretability; the models only provide the final generated result, lacking the reasoning chain within the context generation process. Moreover, for large sets of news texts, existing methods are prone to inefficiency.

[0005] [1] Zhou Xiaomin, Ying Honghui, Diao Zeming, Nie Qinqin, Shi Yi, Wang Yujie, Zhang Zhen, Wu Fei, Zhuo Caibiao, Fang Si'an, Li Bo. An event context generation method and system integrating deep semantic relation classification [P]. Guangdong Province: CN114265932A, 2022-04-01.

[0006] [2] Zhao Chongshuai, Dai Xudong, Gu Chengmin, Zhou Wei, Li Baoshan, Chen Zhigang. Event clustering / network construction method and related devices, equipment and storage media [P]. Anhui Province: CN114357159A, 2022-04-15.

[0007] [3] Jiao Mengshu, Yao Shijie, Luo Jia, Lei Yuling, Du Lei. Event context generation method, equipment and medium [P]. Hunan Province: CN115878761B, 2023-05-09.

[0008] [4] Lin Zhengyu, Shen Zhigang, Tang Zhongzhu, Zhou Ziyu, Cui Junjiao. A method and system for sorting out the context of news events [P]. Jiangsu Province: CN115964495A, 2023-04-14. Summary of the Invention

[0009] The purpose of this invention is to propose an event context generation method that integrates the time of news occurrence and semantic similarity, aiming to improve the accuracy, interpretability, and efficiency of news event context generation.

[0010] The technical solution to achieve the purpose of this invention is: a method for generating event context links that integrates news occurrence time and semantic similarity, comprising the following steps:

[0011] Step 1: News Data Preprocessing: Embedded representation of news text content, digital representation of news occurrence time, and completion of news collection preprocessing;

[0012] Step 2: News Collection Window Division: Divide the preprocessed news collection into windows according to the occurrence time to form several news subsets;

[0013] Step 3: News Event Cluster Calculation: Combining the occurrence time and text content information of the news, the clustering algorithm is used to calculate the event cluster results for each news subset within the window and then evaluated.

[0014] Step 4: Event Cluster Fusion in News Windows: Among the event cluster results calculated under each window, select clusters belonging to the same category for fusion, and select representative nodes of the event clusters in the fused cluster results to form a new round of news sets;

[0015] Step 5: Iterative update of news event context results: For the newly formed news set, repeat steps 2 to 4 until the final event context results are obtained.

[0016] Further, in step 1, news data preprocessing, the specific method is as follows:

[0017] Use TF-IDF to embed the content of news text into a representation.

[0018] Using timestamps, the time when a news event occurred is parsed into a timestamp according to a fixed format.

[0019] Further, in step 2, the news collection window is divided, and the specific method is as follows:

[0020] The order in which news events occurred can be obtained by sorting the timestamps of the news events from smallest to largest.

[0021] The sorted news collection is then segmented according to the window to form several news subsets.

[0022] Further, in step 3, the news event cluster is calculated, and the specific method is as follows:

[0023] Step 3.1: Calculate the semantic distance matrix between news event sets: After the preprocessing in Step 1, the text content of the news sets is represented as vectors, denoted here as... The semantic distance matrix of a set of news events is calculated using cosine distance, and the specific calculation method is as follows:

[0024]

[0025]

[0026] in, It is a semantic matrix of news text represented using TF-IDF, where each row of the matrix represents the text vector of a news article. Represents the normalized result matrix, A semantic distance matrix representing a set of news events;

[0027] Step 3.2: Calculate the time distance feature matrix between news events: using days as the smallest time granularity, after the preprocessing in Step 1, the set of timestamps of the news events is denoted as... ;

[0028] First, construct the time matrix of the news event collection:

[0029]

[0030] Use threshold Remove the time matrix Elements with a long time span:

[0031]

[0032] Calculate the time distance feature matrix of a news event set :

[0033]

[0034] in, Represents the natural constant. Indicates matrix subscripts, To adjust the coefficients and control the dispersion of the time feature distribution, the default setting is 0.1;

[0035] Step 3.3: Fusing the semantic distance matrix and the temporal distance matrix: The semantic distance matrix and the temporal distance feature matrix calculated in Steps 3.1 and 3.2 are fused to obtain the fused feature matrix. Here, a feature fusion weight is used for calculation, denoted as . The default value is set to 0.5, and the calculation formula is as follows:

[0036]

[0037] Step 3.4: Use a clustering algorithm to divide the event clusters and evaluate the division results: based on the fusion feature matrix obtained in Step 3.3. The event clustering algorithm is used to divide the data into clusters. This algorithm requires setting the number of cluster centers; a pre-calculated distance matrix is ​​input to obtain the clustering results. First, multiple values ​​within a range are selected as the number of cluster centers for the algorithm; the default number is set to [2, 10]. Then, depending on the set number of cluster centers, the fusion feature matrix is ​​input. This process generates multiple event cluster partitioning results. Finally, the contour coefficient is used to evaluate the multiple news event cluster partitioning results, and the result with the highest score is selected as the event cluster output for this window.

[0038] Furthermore, in step 4, the news window event clusters are merged, specifically using the following method:

[0039] For the event cluster results calculated under each window, clusters belonging to the same category are selected for fusion. Then, a representative node is chosen from the fused cluster results to replace the event cluster itself, forming a new news set. The selection strategy for the representative node is based on a distance density calculation method, assuming the number of nodes in the event cluster is... The formula for calculating the node selection strategy is as follows:

[0040]

[0041]

[0042]

[0043] in, Indicates the row and column numbers of the fused feature matrix. Indicates the first The score of each node, This indicates the index of the selected node in the current event cluster. After calculation, each event cluster will have a representative node. The representative nodes of all event clusters are merged to form a new news set.

[0044] An event context generation system that integrates news occurrence time and semantic similarity, based on the aforementioned event context generation method that integrates news occurrence time and semantic similarity, realizes the generation of event contexts that integrate news occurrence time and semantic similarity, specifically including the following modules:

[0045] News data preprocessing module: Embedded representation of news text content and digital representation of news occurrence time;

[0046] News Collection Window Segmentation Module: Divides the preprocessed news collection into windows according to the occurrence time, forming several news subsets, thus completing the news collection preprocessing;

[0047] News event cluster calculation module: Combining the occurrence time and text content information of the news, the module uses a clustering algorithm to calculate the event cluster results for each news subset within a window and evaluates the output results;

[0048] News Window Event Cluster Fusion Module: Among the event cluster results calculated under each window, clusters belonging to the same category are selected for fusion, and representative nodes of the event clusters are selected from the fused cluster results to form a new round of news collection;

[0049] The news event context result iterative update module: For newly formed news sets, it repeats the process of dividing the news set window, calculating news event clusters, and merging news window event clusters, retaining the intermediate process links, until the final event context result is obtained.

[0050] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it generates an event context link that integrates news occurrence time and semantic similarity based on the aforementioned event context link generation method.

[0051] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it generates an event context link that integrates news occurrence time and semantic similarity based on the aforementioned event context link generation method.

[0052] Compared with existing technologies, the present invention has the following significant advantages: (1) It proposes a method for dividing the time window of news events, which supports parallel computation of a large amount of news data and improves the efficiency of generating news event contexts; (2) It proposes a strategy for iteratively updating the event context generation results, which, while integrating multi-window output results and optimizing the event context structure, also preserves the evidence link of model reasoning during the iterative update process, thereby enhancing the interpretability of the context generation results; (3) In addition to considering textual semantic information as the basis for event association calculation, it also introduces the time characteristics of news events. News events belonging to the same theme generally have similar occurrence times. Therefore, using the occurrence time to assist in calculating the relationship between news events helps to filter news events with the same theme and improve the accuracy of generating event contexts. Attached Figure Description

[0053] Figure 1 It is a flowchart for generating the event timeline;

[0054] Figure 2 This is the result of the event context generation example. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0056] This invention proposes an event context generation framework that supports window segmentation of news collections based on news occurrence time. It employs a parallel approach to process the computation of multiple windows simultaneously, improving model efficiency. The framework uses an iterative method to integrate multi-window generation results and optimize the event context structure while preserving the model's inference chain, thus improving both the accuracy of the generated context and the interpretability of the model's inference process. Furthermore, this invention proposes an event cluster extraction method that combines news occurrence time and textual semantic information. It embeds news occurrence time and content information into a distance matrix, measuring the aggregation and dispersion of news items in the collection from both temporal and semantic dimensions, providing a computational basis for generating the event context.

[0057] Step 1: News Data Preprocessing. Preprocessing mainly includes embedded representation of the news text content and digital representation of the news occurrence time. Embedded representation of the text refers to converting the news text into vectors. Common methods include using the word frequency statistical method TF-IDF, word vectors Word2Vec, and the language model BERT. Generally, news on different topics will show significant changes in word frequency statistics. The TF-IDF-based embedding method can capture the semantic change features of news on different topics. Moreover, the TF-IDF-based method is more computationally efficient and less costly than word vectors and language models. Therefore, this invention uses TF-IDF as the text embedding representation method. For the news occurrence time, this invention uses a timestamp method, parsing the news occurrence time into a timestamp according to a fixed format "YYYY-MM-DD HH:mm:SS", providing a calculation basis for subsequent time distance matrix operations. Here, "YYYY" represents the year, "MM" represents the month, "DD" represents the day, "HH" represents the hour, "mm" represents the minute, "SS" represents the second, and "-" and ":" represent the separators for year, month, day and hour, minute, and second, respectively.

[0058] Step 2: News Collection Window Division. The preprocessed news collection is divided into windows according to the occurrence time, forming several news subsets. Step 3 is executed in parallel computing to infer the event cluster results of each sub-event. First, based on the converted news occurrence timestamps, the news is sorted in ascending order to obtain the chronological order of occurrence. The sorted news collection is then segmented according to windows, with the default window size set to 10 news items.

[0059] Step 3: News Event Cluster Calculation. For each news subset within a window, combining the news timestamps and textual semantic information, calculate the temporal and semantic distance feature matrices of the news subsets within the window. Use a clustering algorithm to obtain the event cluster partitioning results for the news subsets. The calculation process is specifically divided into the following four sub-steps:

[0060] Step 3.1: Calculate the semantic distance matrix between the news event sets. After the preprocessing in Step 1, the text content of the news sets is represented as vectors, denoted here as... The semantic distance matrix of a set of news events is calculated using cosine distance, as follows:

[0061]

[0062]

[0063] in, It is a semantic matrix of news text represented using TF-IDF, where each row of the matrix represents the text vector of a news article. Represents the normalized result matrix, The semantic distance matrix represents a set of news events.

[0064] Step 3.2: Calculate the temporal distance feature matrix between news events. Since the granularity of news occurrence times varies, this invention uses days as the smallest time granularity for calculation. After the preprocessing in Step 1, the set of timestamps for the news events is denoted as... .

[0065] First, construct the time matrix of the news event collection:

[0066]

[0067] It is generally believed that news with a long time span has low relevance. To reduce noise from irrelevant news, this invention introduces a threshold variable. The default threshold is set to 7, which filters out time-distance information with a time span exceeding 7 days. Using the threshold... Remove the time matrix Elements with a long time span:

[0068]

[0069] Calculate the time distance feature matrix of a news event set :

[0070]

[0071] in, Represents the natural constant. Indicates matrix subscripts, To adjust the coefficient and control the dispersion of the time feature distribution, the default setting is 0.1.

[0072] Step 3.3: Fusing the semantic distance matrix and the temporal distance matrix. The semantic distance matrix and the temporal distance feature matrix calculated in Steps 3.1 and 3.2 are fused to obtain the fused feature matrix. Here, a feature fusion weight is used for calculation, denoted as . The default value is set to 0.5. The calculation formula is as follows:

[0073]

[0074] Step 3.4: Use a clustering algorithm to divide the event clusters and evaluate the division results. Based on the fusion feature matrix obtained in Step 3.3... Hierarchical clustering algorithm is used to divide event clusters. To enable the hierarchical clustering algorithm to consider multiple scenarios, multiple values ​​are selected within a range as the number of cluster centers, generating various event cluster division results. The default setting is between [2, 10]. The silhouette coefficient is used to evaluate the multiple news event cluster division results, and the result with the highest score is selected as the event cluster output result for this window.

[0075] Step 4: Event Cluster Fusion in News Windows. From the event cluster results calculated under each window, clusters belonging to the same category are selected for fusion. Then, representative nodes of the event clusters are selected from the fused cluster results to form a new round of news sets. To improve the efficiency of event context generation, this invention uses representative nodes of event clusters instead of the event clusters themselves in the model calculation process. The selection strategy for representative nodes of event clusters adopts a distance density-based calculation method, assuming that the number of nodes in an event cluster is... The formula for calculating the node selection strategy is as follows:

[0076]

[0077]

[0078]

[0079] in, Indicates the row and column numbers of the fused feature matrix. Indicates the first The score of each node, Indicates the index of the selected node in the current event cluster.

[0080] After calculation, each event cluster will have a representative node, and the representative nodes of all event clusters will be merged to form a new news collection.

[0081] Step 5: Iterative update of news event context results. For the newly formed news set, repeat steps 2 to 4, retaining the intermediate process links, until the final event context result is obtained. The overall framework iterative update process is as follows: (1) In the first round, input the original news set, each news item is an event cluster, and the representative node of each event cluster is the news itself. After calculation in steps 2 to 4, a new round of event cluster division results are formed, and the original news is merged according to the division results. Since each node in the news set input in the first round represents the news itself, the news is merged according to the new round of division results to obtain a new event cluster, and each event cluster contains several news items. Select a representative node in each event cluster to form a new news set. (2) In the second round, input the new set, and repeat steps 2 to 4 to obtain the event cluster division results of the second round. Since each node in the set input in the second round represents the event cluster results of the previous round, the clusters represented by the nodes will be merged when merging according to the new round of event cluster division results. After selecting representative nodes from the merged clusters, a new round of news sets is formed. This process is repeated to obtain the final event context and the process chain of each round of model inference.

[0082] To enhance model stability, the maximum number of iterations in the update process is set to 100 by default. The model will be forced to exit if the maximum number of iterations is exceeded. Additionally, to improve model efficiency, a maximum stable iteration count parameter is added, set to 3 by default. If the network results generated by the model tend to stabilize, and the number of consecutively unchanged iterations exceeds the maximum stable iteration count, the model will exit the loop process.

[0083] Example

[0084] To verify the effectiveness of the present invention, the following embodiments are conducted.

[0085] Input: News about the XX conflict and related news about the XXXXX joint military exercise in 2023 were collected from the internet. To verify the effectiveness of semantic information in generating the model's context, news data about the XX conflict in August 2022 and June 2023 were selected. Additionally, to verify the effectiveness of temporal features, news about the XXXXX joint military exercise in April 2023 was added. The total number of data entries is 531. Entries 0 to 61 represent news about the XX conflict that occurred around June 2023, entries 62 to 376 represent news about the military exercise in April 2023, and entries 377 to 530 represent news about the XX conflict that occurred around August 2022.

[0086] Step 1: News Data Preprocessing. The input news collection is first segmented using the jieba word segmentation tool, then TF-IDF is used to convert the news content text into vectors. For the occurrence time of the news, the occurrence time is parsed into a timestamp according to the format "YYYY-MM-DD HH:mm:ss". The vector results and timestamp results are combined and organized into a news collection.

[0087] Step 2: News Collection Window Division. Based on the news collection obtained in Step 1, first sort the news collection according to the chronological order of their occurrence. Set the time window size to 10, and then divide the sorted collection into several news subsets.

[0088] Step 3: News Event Cluster Calculation. Based on the timestamps of the news events, calculate the distance between the occurrence times of different news events, using days as the smallest time granularity. Results exceeding a threshold are removed, resulting in a time distance matrix. The time feature coefficient is set to 0.1. Based on the text vectors obtained in Step 1, calculate the semantic distance matrix using cosine similarity. Adjust the semantic distance matrix and the time distance matrix using a scaling factor, then sum them to obtain the fused feature matrix. The scaling factor is set to 0.5. Calculate the fused feature matrix using a hierarchical clustering algorithm. The number of cluster centers is between 2 and 10. Multiple results are evaluated using the silhouette coefficient, and the result with the highest score is selected as the event cluster partitioning result for that window.

[0089] Step 4: Event Cluster Fusion in the News Window. Based on the event cluster division results obtained in Step 3, nodes belonging to the same cluster are merged to obtain new event clusters. Representative nodes are selected for each new cluster to form a new round of news collections.

[0090] Step 5: Iterative Update of News Event Context Results. The new set obtained in Step 4 is used as input to iteratively calculate the model again, retaining intermediate results during the iteration process as the basis for the inference chain. After the iteration is complete, the final news event context is output.

[0091] Output: Provides the main outline of the event, as shown below. Figure 2 As shown. From Figure 2 As can be seen, after 17 rounds of iteration, the model outputs an event timeline, while also preserving the link results of the iteration process from round 1 to round 17, providing the basis for the reasoning process. In the final result of round 17, it can be seen that the first and second of the three timeline results generated by the model basically conform to the original data theme division, and the conflict data and military exercise data can be well distinguished. Furthermore, the news of conflict events from different time slices can also be well integrated.

[0092] This invention also proposes an event context generation system that integrates news occurrence time and semantic similarity. Based on the aforementioned event context generation method that integrates news occurrence time and semantic similarity, the system realizes the generation of event contexts that integrate news occurrence time and semantic similarity, specifically including the following modules:

[0093] News data preprocessing module: Embedded representation of news text content and digital representation of news occurrence time;

[0094] News Collection Window Segmentation Module: Divides the preprocessed news collection into windows according to the occurrence time, forming several news subsets, thus completing the news collection preprocessing;

[0095] News event cluster calculation module: Combining the occurrence time and text content information of the news, the module uses a clustering algorithm to calculate the event cluster results for each news subset within a window and evaluates the output results;

[0096] News Window Event Cluster Fusion Module: Among the event cluster results calculated under each window, clusters belonging to the same category are selected for fusion, and representative nodes of the event clusters are selected from the fused cluster results to form a new round of news collection;

[0097] The news event context result iterative update module: For newly formed news sets, it repeats the process of dividing the news set window, calculating news event clusters, and merging news window event clusters, retaining the intermediate process links, until the final event context result is obtained.

[0098] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it generates an event context link that integrates news occurrence time and semantic similarity based on the aforementioned event context link generation method.

[0099] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it generates an event context link that integrates news occurrence time and semantic similarity based on the aforementioned event context link generation method.

[0100] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0101] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for generating event context links by integrating news occurrence time and semantic similarity, characterized in that, Includes the following steps: Step 1: News Data Preprocessing: Embedded representation of news text content, digital representation of news occurrence time, and completion of news collection preprocessing; Step 2: News Collection Window Division: Divide the preprocessed news collection into windows according to the occurrence time to form several news subsets; Step 3: News Event Cluster Calculation: Combining the occurrence time and text content information of the news, the clustering algorithm is used to calculate the event cluster results for each news subset within the window and then evaluated. Step 4: Event Cluster Fusion in News Windows: Among the event cluster results calculated under each window, select clusters belonging to the same category for fusion, and select representative nodes of the event clusters in the fused cluster results to form a new round of news sets; Step 5: Iterative update of news event context results: For the newly formed news set, repeat steps 2 to 4 until the final event context results are obtained; in: Step 1, news data preprocessing, the specific method is as follows: Use TF-IDF to embed the content of news text into a representation. Use timestamps to parse the time of news occurrence into timestamps according to a fixed format; Step 2, dividing the news collection window, the specific method is as follows: The order in which news events occurred can be obtained by sorting the timestamps of the news events from smallest to largest. The sorted news collection is then divided into several news subsets according to the window. Step 3, Calculation of news event clusters, the specific method is as follows: Step 3.1: Calculate the semantic distance matrix between news event sets: After the preprocessing in Step 1, the text content of the news set is represented as a vector, denoted as X. The semantic distance matrix of the news event set is calculated using cosine distance, as follows: D x =1-X'X' T (2) Where X is the semantic matrix of the news text represented using TF-IDF, with each row of the matrix representing the text vector of a news item, X' representing the normalized X matrix, and D... x A semantic distance matrix representing a set of news events; Step 3.2: Calculate the time distance feature matrix between news events: Using days as the smallest time granularity, after the preprocessing in Step 1, the set of timestamps of the news events is denoted as {t1, t2…t}. n }; First, construct the time matrix of the news event collection: Use a threshold to remove elements in the time matrix T with excessively long time spans: Calculate the time distance feature matrix D of the news event set. t : Where e represents the natural constant, ij represents the matrix subscript, and w is the adjustment coefficient, which controls the dispersion of the time feature distribution, and is set to 0.1 by default; Step 3.3: Fusing the semantic distance matrix and the temporal distance matrix: Fusing the semantic distance matrix and the temporal distance feature matrix calculated in Steps 3.1 and 3.2 to obtain the fused feature matrix D. mix Here, a feature fusion weight is used for calculation, denoted as w. t The default value is set to 0.5, and the calculation formula is as follows: D mix =w t D x +(1-w t )D t (6) Step 3.4: Use a clustering algorithm to divide the event clusters and evaluate the division results: Based on the fusion feature matrix D obtained in Step 3.3 mix Hierarchical clustering algorithm is used to divide event clusters. First, multiple values ​​are selected within a range for the hierarchical clustering algorithm to be used as the number of cluster centers. The default number of cluster centers is set to be between [2, 10]. Then, the fusion feature matrix D is input. mix Based on different numbers of cluster centers, various event cluster partitioning results are generated; finally, the silhouette coefficient is used to evaluate the multiple news event cluster partitioning results, and the result with the highest score is selected as the event cluster output result of this window. Step 4, news window event cluster merging, the specific method is as follows: For the event cluster results calculated under each window, clusters belonging to the same category are selected for fusion. Then, a representative node is selected from the fused cluster results to replace the event cluster itself, forming a new round of news sets. The selection strategy for the representative node of the event cluster uses a distance density-based calculation method. Assuming the number of nodes in an event cluster is N, the calculation formula for the representative node selection strategy is as follows: s i ∈S,S={s1,s2…,s N } (8) k = argmax(S) (9) Where i,j represent the row and column numbers of the fused feature matrix, and s i Let represent the score of the i-th node, and k represent the index of the selected node in the current event cluster. After calculation, each event cluster will have a representative node. The representative nodes of all event clusters are merged to form a new news set.

2. A system for generating event context links that integrates news occurrence time and semantic similarity, characterized in that, The event context generation method based on the fusion of news occurrence time and semantic similarity as described in claim 1, specifically includes the following modules: News data preprocessing module: Embedded representation of news text content and digital representation of news occurrence time; News Collection Window Segmentation Module: Divides the preprocessed news collection into windows according to the occurrence time, forming several news subsets, thus completing the news collection preprocessing; News event cluster calculation module: Combining the occurrence time and text content information of the news, the module uses a clustering algorithm to calculate the event cluster results for each news subset within a window and evaluates the output results; News Window Event Cluster Fusion Module: Among the event cluster results calculated under each window, clusters belonging to the same category are selected for fusion, and representative nodes of the event clusters are selected from the fused cluster results to form a new round of news collection; The news event context result iterative update module: For newly formed news sets, it repeats the process of dividing the news set window, calculating news event clusters, and merging news window event clusters, retaining the intermediate process links, until the final event context result is obtained.

3. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it generates an event context link that integrates news occurrence time and semantic similarity based on the event context link generation method of claim 1.

4. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it realizes the generation of an event context link that integrates news occurrence time and semantic similarity based on the event context link generation method of claim 1.

Images