Multi-perspective news creation method and system based on media viewpoints

Abstract

The present invention relates to the field of artificial intelligence technology. The present invention discloses a multi-perspective news creation method and system based on media viewpoints, including constructing an event evolution graph based on available news data, dividing the available news data into M stage news data according to the event evolution graph, respectively screening out M representative news data from the M stage news data, and when the semantic discrimination result is semantic consistency, reconstructing the M representative news data to obtain intermediate news data, inputting the intermediate news data into a pre-constructed factor extraction model to obtain N factor triplets, and generating target news data based on the N factor triplets. The present invention obtains available news data after screening, and considering that news content is staged, it can form comprehensive relevant content from representative news events at different stages.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more specifically, to a method and system for creating news from multiple perspectives based on media viewpoints. Background Art

[0002] Social media has become the main platform for the public to release information and exchange views. After an event occurs, people will quickly release and disseminate a large amount of relevant information on various social media, discuss derivative topics, and express their respective positions and opinions. Artificial intelligence is a cutting-edge technology that is committed to developing systems and algorithms that can simulate, expand and enhance human intelligence. With the application of artificial intelligence in all walks of life, the media industry is no exception. In existing technologies, there are already technologies that use artificial intelligence technology to collect relevant news on the Internet in real time and analyze this information to generate relevant content. However, there are still some challenges and problems in this process.

[0003] For example, the Chinese patent with publication number CN117493649A provides an information content generation method, device, equipment and storage medium. The patent extracts target key information closely related to the information background information from at least one first information content, and extracts target electronic resources closely related to the information background information from at least one second information content, so as to obtain effective information in a relatively short time. The Chinese patent with publication number CN111626049A provides a title correction method, device, electronic equipment and storage medium for multimedia information. The patent performs error recognition processing on the title of multimedia information, obtains the error position in the title, and searches a candidate correction database corresponding to the type based on the text of the error position.

[0004] Although the above patent can analyze the acquired news content based on artificial intelligence technology, the existing sources of news content are too complex, and direct analysis without screening is prone to produce erroneous related content. In addition, news content is phased, but the existing technology is less able to select representative news events according to different stages, making it difficult to form comprehensive related content.

[0005] In view of this, the present invention proposes a multi-perspective news creation method and system based on media viewpoints to solve the above problems. Summary of the invention

[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides a multi-perspective news creation method and system based on media viewpoints.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] First, a multi-perspective news creation method based on media perspectives, including:

[0009] S10: Acquire real-time news data and news source information, and divide the real-time news data into available news data and unavailable news data according to the news source information;

[0010] S20: constructing an event evolution graph based on the available news data, dividing the available news data into M stages of news data according to the event evolution graph, and selecting M representative news data from the M stages of news data, where M is an integer greater than 0;

[0011] S30: inputting the M representative news data into the pre-built semantic discrimination model in sequence to obtain a semantic discrimination result. When the semantic discrimination result is semantic inconsistency, the process returns to S20. When the semantic discrimination result is semantic consistency, the process proceeds to S40. The semantic discrimination result includes at least semantic inconsistency and semantic consistency.

[0012] S40: Reconstruct the M representative news data to obtain intermediate news data, input the intermediate news data into a pre-built feature extraction model to obtain N feature triplets, and generate target news data based on the N feature triplets, where N is an integer greater than 0.

[0013] Furthermore, the method for constructing the partitioning model includes:

[0014] Acquire Q groups of data, where Q is a positive integer greater than 1, and the data includes historical news source information, historical news data, and historical division results. Use the historical news source information, historical news data, and historical division results as a sample set, divide the sample set into a training set and a test set, build a classifier, use the historical news source information and historical news data in the training set as input data, use the historical division results in the training set as output data, train the classifier to obtain an initial classifier, test the initial classifier using the test set, and output a classifier that meets a preset accuracy as a division model.

[0015] Furthermore, the method of constructing an event evolution graph based on available news data includes:

[0016] Based on information extraction technology, the available news data is processed to obtain S sub-events. The temporal relationship between the S sub-events is determined according to the time labels. The causal relationship between the S sub-events is determined based on the graph neural network. An event evolution graph is constructed with the S sub-events as nodes and the temporal relationship and causal relationship between the S sub-events as edges.

[0017] Furthermore, the method of selecting M representative news data from the M stage news data includes:

[0018] Traverse the news data of M stages, calculate the consistency factor corresponding to each element in the news data of each stage, determine the representative news data according to the consistency factor, and thus screen out M representative news data.

[0019] Furthermore, the method for calculating the consistency factor includes:

[0020] ;

[0021] In the formula, For the The consistency factor corresponding to the elements is For the stage news data The vector representation of elements is For the stage news data The vector representation of elements is is the total number of elements in the stage news data, For the The element and The Euclidean distance between elements.

[0022] Furthermore, the method for constructing the semantic discrimination model includes:

[0023] The semantic discriminant model includes a generative adversarial network and a classifier, and the discriminator in the generative adversarial network is trained using the actual representative news data to obtain a trained discriminator;

[0024] The trained discriminator is used to guide the generator in the adversarial generative network to generate forged representative news data similar to the actual representative news data, where the forged representative news data represents news data with similar text content but different semantics;

[0025] The actual representative news data and the forged representative news data are input into the classifier for training to obtain the semantic discrimination results.

[0026] Furthermore, when the semantic determination result is semantic inconsistency, before returning to S20, the process further includes:

[0027] The marked news data in the semantic discrimination result is obtained, and the marked representative news data is removed from the corresponding stage news data.

[0028] Furthermore, the method of reconstructing the M representative news data to obtain the intermediate news data includes:

[0029] Each entity in the representative news data is labeled, and a relationship prompt sentence for the entity pair is added to each labeled representative news data.

[0030] In the second aspect, a multi-perspective news creation system based on media viewpoints is used to implement the multi-perspective news creation method based on media viewpoints, including:

[0031] Division module: used to obtain real-time news data and news source information, and divide the real-time news data into available news data and unavailable news data according to the news source information;

[0032] The first processing module is used to construct an event evolution graph based on available news data, divide the available news data into M stages of news data according to the event evolution graph, and select M representative news data from the M stages of news data, where M is an integer greater than 0;

[0033] Judgment module: used to input M representative news data into the pre-built semantic discrimination model in sequence to obtain semantic discrimination results. When the semantic discrimination result is semantic inconsistency, it returns to the first processing module. When the semantic discrimination result is semantic consistency, it transfers to the second processing module. The semantic discrimination result at least includes semantic inconsistency and semantic consistency.

[0034] The second processing module is used to reconstruct the M representative news data to obtain intermediate news data, input the intermediate news data into the pre-built feature extraction model, obtain N feature triplets, and generate target news data based on the N feature triplets, where N is an integer greater than 0.

[0035] A third aspect is a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the above-mentioned multi-perspective news creation method based on media viewpoints.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention constructs an event evolution graph based on available news data, divides the available news data into M stage news data according to the event evolution graph, screens out M representative news data from the M stage news data respectively, and sequentially inputs the M representative news data into a pre-constructed semantic discrimination model. When the semantic discrimination result is semantically consistent, the M representative news data are reconstructed to obtain intermediate news data, and the intermediate news data are input into a pre-constructed element extraction model to obtain N element triplets. Target news data is generated based on the N element triplets. The present invention obtains available news data after screening, and considering that news content is staged, it is possible to form comprehensive related content from representative news events at different stages. Similarly, due to the selection of representative news events at different stages, news creation can be carried out from multiple perspectives. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1A flowchart of the multi-perspective news creation method based on media viewpoints in the present invention;

[0039] Figure 2 It is a structural schematic diagram of the multi-perspective news creation system based on media viewpoints in the present invention;

[0040] Figure 3 is a schematic diagram of an event evolution diagram in the present invention;

[0041] Figure 4 is a schematic diagram of a computer-readable storage medium in the present invention;

[0042] In the attached figure:

[0043] 10, nodes; 20, edges; 30, event stage. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] Example 1

[0046] See also Figure 1 As shown, this embodiment discloses a multi-perspective news creation method based on media viewpoints, including:

[0047] S10: Acquire real-time news data and news source information, and divide the real-time news data into available news data and unavailable news data according to the news source information;

[0048] In this embodiment, the real-time news data may be news text content obtained from major platforms or major social media, but the above news text content all describes the same event, for example:

[0049] The first text content can be, "In mid-May 2024, a city officially announced that it would host an annual international marathon. Once the news was released, it attracted widespread attention. Runners from all over the world began to sign up, and preparations for the event were carried out in full swing. All parties were looking forward to the competition. The municipal government and event organizers actively prepared to ensure the safety and smooth progress of the event."

[0050] The second text could be, "As the race date approaches, the city enters an intense preparation stage. In early June, contestants arrive one after another, and the event atmosphere gradually heats up. On the day of the race, spectators gather early on both sides of the track and cheer constantly. The contestants go all out on the track, and the race is in full swing. Traffic control and security measures are in place in the city to ensure the smooth progress of the event."

[0051] The third text could be, "A week after the race, the impact of the event continued to ferment in the city. A Kenyan runner won the full marathon championship with a time of 2 hours and 10 minutes and set a new track record. The event not only attracted more than 10,000 runners to participate, but also promoted the development of related industries such as tourism and catering."

[0052] It should be noted that news source information includes but is not limited to publishing agency information, publishing platform information, publishing author information and publishing time information. The publishing agency information can be CNN, Xinhuanet, etc., and the publishing platform information can be Weibo, Douyin, etc.

[0053] Methods for dividing real-time news data into available news data and unavailable news data according to news source information include:

[0054] The real-time news data and the corresponding news source information are input into a pre-built partitioning model to obtain a partitioning result, and the real-time news data is partitioned into available news data and unavailable news data according to the partitioning result, wherein the partitioning result includes available and unavailable news data.

[0055] The construction methods of the partitioning model include:

[0056] Acquire Q groups of data, where Q is a positive integer greater than 1, and the data includes historical news source information, historical news data, and historical division results. Use the historical news source information, historical news data, and historical division results as a sample set, divide the sample set into a training set and a test set, build a classifier, use the historical news source information and historical news data in the training set as input data, use the historical division results in the training set as output data, train the classifier to obtain an initial classifier, test the initial classifier using the test set, and output a classifier that meets a preset accuracy as a division model. The classifier is preferably a naive Bayes model or a support vector machine model.

[0057] It is understandable that due to the excessive development of the Internet, the sources of news content are too complex. In order to avoid the subsequent generation of erroneous related content, it is necessary to screen the news content.

[0058] S20: constructing an event evolution graph based on the available news data, dividing the available news data into M stages of news data according to the event evolution graph, and selecting M representative news data from the M stages of news data, where M is an integer greater than 0;

[0059] Specifically, the method of constructing an event evolution graph based on available news data includes:

[0060] Based on information extraction technology, the available news data is processed to obtain S sub-events. The temporal relationship between the S sub-events is determined according to the time labels. The causal relationship between the S sub-events is determined based on the graph neural network. An event evolution graph is constructed with the S sub-events as nodes and the temporal relationship and causal relationship between the S sub-events as edges.

[0061] It should be noted that information extraction is a natural language processing (NLP) technology that is used to automatically extract structured information from unstructured text. Its goal is to convert large amounts of text data into structured data forms that can be processed by computers. Information extraction includes named entity recognition, relationship extraction, event extraction, and template filling. Taking the above-mentioned available news data including the first text content, the second text content, and the third text content as an example, three sub-events can be obtained through information extraction. The first sub-event is "In mid-May 2024, a city announced that it would host an annual international marathon." The second sub-event is "In early June 2024, contestants arrived one after another, and the audience gathered on both sides of the track early, cheering constantly. On the day of the race, the event was in full swing." The third sub-event is "Kenyan athletes set a record and won the championship. The impact of the event continued and promoted the development of local tourism and catering industries."

[0062] Specifically, the graph neural network also takes each sub-event as a node and defines relevant features for each sub-event. The relevant features include the description, timestamp, category, etc. of the event, and then updates the state of the node through a message passing mechanism. Message passing includes receiving information from neighboring nodes and aggregating it to the current node. This mechanism can help capture the causal relationship between nodes. The above-mentioned information extraction and graph neural network are both existing technologies, and this embodiment will not go into details.

[0063] It can be understood that, taking the first sub-event, the second sub-event and the third sub-event as examples, there is an obvious time relationship between the first sub-event and the second sub-event, and there is an obvious causal relationship between the second sub-event and the third sub-event. The above is just an example. Usually, there are several sub-events in each stage, such as Figure 3 As shown, Figure 3 Nodes 10, edges 20 and event stages 30 are shown, each node 10 represents a sub-event, and edges 20 can be a temporal relationship or a causal relationship.

[0064] It should be added that the reason why there are several sub-events in each stage is that there are different descriptions of the same thing, but they are essentially the same. For example, the first text content can be, "In mid-May 2024, a city officially announced that it would host an annual international marathon. Once the news was released, it attracted widespread attention. Runners from all over the world began to sign up, and preparations for the event were also carried out in full swing. All parties were looking forward to the competition. The municipal government and event organizers actively prepared to ensure the safety and smooth progress of the event." A similar text content could be, "On May 15, 2024, a city announced that it would host an annual international marathon. After the news was released, it quickly attracted widespread attention. Runners from all over the world signed up to participate, and preparations for the event began to accelerate. The municipal government and event organizers actively coordinated to ensure that the event could proceed smoothly and achieve the expected results."

[0065] The method of dividing the available news data into M stages of news data according to the event evolution graph includes:

[0066] M event stages in the event evolution graph are obtained, and available news data are divided into M stage news data according to the nodes in the M event stages.

[0067] It should be noted that the available news data includes several available news sub-data. Each node in the event evolution graph represents a sub-event. Each sub-event has corresponding available news sub-data. According to the position of the node in the event evolution graph, the corresponding available news sub-data can be divided into which stage. For example, following the above, the available news data includes 3 available news sub-data, namely the first text content, the second text content and the third text content. From the above content, it can be seen that there are 3 event stages in the event evolution graph, so the available news data is divided into 3 stages of news data accordingly.

[0068] The method of selecting M representative news data from M stage news data includes:

[0069] Traverse the news data of M stages, calculate the consistency factor corresponding to each element in the news data of each stage, determine the representative news data according to the consistency factor, and thus screen out M representative news data.

[0070] In this embodiment, there are multiple available news sub-data for each stage of news data. The above elements represent the available news sub-data. By calculating the consistency factor, the available news sub-data with the largest consistency factor is selected as the representative news data.

[0071] Methods for calculating consistency factors include:

[0072] ;

[0073] In the formula, For the The consistency factor corresponding to the elements is For the stage news data The vector representation of elements is For the stage news data The vector representation of elements is is the total number of elements in the stage news data, For the The element and The Euclidean distance between elements.

[0074] It should be noted that the Euclidean distance can be used to measure the similarity or difference between two vectors. In multidimensional space, the smaller the Euclidean distance of the vectors, the more similar they are. Conversely, the larger the distance, the greater the difference. Therefore, in this embodiment, the Euclidean distance between an element and other elements is calculated respectively to characterize the similarity between an element and other elements. In this embodiment, the larger the consistency factor, the more similar an element is to other elements. Therefore, this element can be used as the representative news data in the news data of a stage. This ensures that the selected representative news data is most similar to other news data in this stage. This means that the selected news data can more accurately reflect the overall content and theme of the news data in this stage, avoid selecting individual data points that deviate from the main information, and make subsequent analysis more efficient.

[0075] S30: inputting the M representative news data into the pre-built semantic discrimination model in sequence to obtain a semantic discrimination result. When the semantic discrimination result is semantic inconsistency, the process returns to S20. When the semantic discrimination result is semantic consistency, the process proceeds to S40. The semantic discrimination result includes at least semantic inconsistency and semantic consistency.

[0076] In this embodiment, the method for constructing the semantic discrimination model includes:

[0077] The semantic discriminant model includes a generative adversarial network and a classifier, and the discriminator in the generative adversarial network is trained using the actual representative news data to obtain a trained discriminator;

[0078] The trained discriminator is used to guide the generator in the adversarial generative network to generate forged representative news data similar to the actual representative news data, where the forged representative news data represents news data with similar text content but different semantics;

[0079] The actual representative news data and the forged representative news data are input into the classifier for training to obtain the semantic discrimination results.

[0080] It should be noted that Generative Adversarial Networks (GANs) is a generative model proposed by deep learning. Its goal is to generate realistic data samples through adversarial learning between two network models. GANs consists of two main components: the generator and the discriminator. These two components compete with each other and gradually improve the quality of samples generated by the generator through adversarial training.

[0081] It is understandable that due to the rapid development of the Internet, a large amount of ironic text content has appeared in commentary content or news content, resulting in inconsistency between the text content and its corresponding semantics. For example, someone reads a report that a city government decided to increase parking fees in the urban area to reduce traffic congestion. Then someone may leave a message on social media: "Great! I really can't wait to pay more for parking every day. What good news!" Since the ultimate goal of the present invention is to create news from multiple perspectives, the present invention filters out representative news data from the news data at each stage, and it is necessary to ensure that the semantics of the filtered representative news data are consistent. Therefore, the present invention needs to detect the semantics of the representative news data, so that the final created news will not have logical errors.

[0082] It should be added that since the training of the model requires a large amount of historical data, but there is not much ironic text content in the news field, it is necessary to generate the required historical data through adversarial training, so that the subsequent semantic discrimination model can be more robust.

[0083] When the semantic determination result is semantic inconsistency, before returning to S20, the following steps are also included:

[0084] The marked news data in the semantic discrimination result is obtained, and the marked representative news data is removed from the corresponding stage news data.

[0085] Specifically, the marked news data refers to representative news data whose semantics are inconsistent with other representative news data. For example, there are 6 representative news data that are input into the pre-built semantic discrimination model in sequence, of which 4 representative news data have the same semantics, while 2 representative news data have different semantics. The 2 representative news data need to be deleted and then return to S20 to be re-screened in the corresponding stage news data.

[0086] S40: Reconstruct the M representative news data to obtain intermediate news data, input the intermediate news data into a pre-built feature extraction model to obtain N feature triplets, and generate target news data based on the N feature triplets, where N is an integer greater than 0.

[0087] The method of reconstructing the M representative news data to obtain the intermediate news data includes:

[0088] Each entity in the representative news data is labeled, and a relationship prompt sentence for the entity pair is added to each labeled representative news data.

[0089] For example, representative news data could be "In mid-May 2024, a city officially announced that it would host an annual international marathon. Once the news was released, it attracted widespread attention. Runners from all over the world began to sign up, and preparations for the event were also carried out in full swing."

[0090] The reconstructed intermediate news data can be "In mid-May 2024, XX city XX officially announced that it will host the annual international XX marathon XX. Once the news was released, it attracted widespread attention. Runners from all over the world began to sign up, and preparations for the event were also carried out in full swing. The relationship between a city and the marathon is mask1, mask2 or mask3."

[0091] It should be noted that entities can represent keywords in news data, while masks refer to the relationship between entities. They will be replaced with specific relationship descriptions in actual use, for example:

[0092] mask1 may represent "host city", that is, "a certain city is the host city of a marathon".

[0093] mask2 may represent "location", i.e. "the marathon will be held in a certain city".

[0094] mask3 might stand for "event organization", i.e. "the organizing committee of a certain city is responsible for the preparation of the marathon".

[0095] In this way, by annotating entities and relationships, the reconstructed news data not only retains the main content of the original information, but also adds a clear description of the entity relationship, which helps to convey information more clearly and provides a structured information basis for subsequent feature extraction. This process enables the feature extraction model to more efficiently identify and extract important news elements. It can be understood that the above reconstruction of representative news data can be achieved through natural language processing technology, through entity recognition, relationship extraction, and template matching and filling. It can be understood that once the entities and relationships are identified, predefined templates (such as mask1, mask2, mask3) are used for placeholders. In actual applications, these placeholders will be replaced by specific relationship descriptions to generate structured intermediate news data.

[0096] The construction methods of the feature extraction model include:

[0097] A sample data set is obtained, wherein the sample data set includes historical intermediate news data and historical element triplets, the sample data set is divided into a sample training set and a sample test set, a regression network is constructed, the historical intermediate news data in the sample training set is used as input data of the regression network, and the historical element triplets in the sample training set are used as output data of the regression network, the regression network is trained to obtain an initial regression network for predicting real-time element triplets, the initial regression network is tested using the sample test set, and the initial regression network that satisfies a preset error value is output as a factor extraction model, and the initial regression network is preferably a deep neural network model.

[0098] In this embodiment, the element triplet generally includes a first entity, a second entity, and a relationship or action between the first entity and the second entity. Taking the above content as an example, "a city" can be the first entity, "a marathon race" can be the second entity, and "the venue" can be the relationship.

[0099] A method for generating target news data based on N element triplets may be to input the N element triplets into an existing large language model, such as GPT-4, BERT, or Wenyan Yixin, etc., so that the target news data can be generated.

[0100] This embodiment constructs an event evolution graph based on available news data, divides the available news data into M stages of news data according to the event evolution graph, selects M representative news data from the M stages of news data, and inputs the M representative news data into a pre-constructed semantic discrimination model in sequence. When the semantic discrimination result is semantically consistent, the M representative news data are reconstructed to obtain intermediate news data, and the intermediate news data are input into a pre-constructed feature extraction model to obtain N feature triplets, and target news data is generated based on the N feature triplets. This embodiment obtains available news data after screening, and considering that news content is staged, it is possible to form comprehensive relevant content from representative news events at different stages. Similarly, since representative news events at different stages are selected, news can be created from multiple perspectives.

[0101] Example 2

[0102] like Figure 2 As shown, this embodiment provides a multi-perspective news creation system based on media viewpoints on the basis of embodiment 1, including:

[0103] Division module: used to obtain real-time news data and news source information, and divide the real-time news data into available news data and unavailable news data according to the news source information;

[0104] In this embodiment, the real-time news data may be news text content obtained from major platforms or major social media, but the above news text content all describes the same event.

[0105] It should be noted that news source information includes but is not limited to publishing agency information, publishing platform information, publishing author information and publishing time information. The publishing agency information can be CNN, Xinhuanet, etc., and the publishing platform information can be Weibo, Douyin, etc.

[0106] Methods for dividing real-time news data into available news data and unavailable news data according to news source information include:

[0107] The real-time news data and the corresponding news source information are input into a pre-built partitioning model to obtain a partitioning result, and the real-time news data is partitioned into available news data and unavailable news data according to the partitioning result, wherein the partitioning result includes available and unavailable news data.

[0108] The construction methods of the partitioning model include:

[0109] Acquire Q groups of data, where Q is a positive integer greater than 1, and the data includes historical news source information, historical news data, and historical division results. Use the historical news source information, historical news data, and historical division results as a sample set, divide the sample set into a training set and a test set, build a classifier, use the historical news source information and historical news data in the training set as input data, use the historical division results in the training set as output data, train the classifier to obtain an initial classifier, test the initial classifier using the test set, and output a classifier that meets a preset accuracy as a division model. The classifier is preferably a naive Bayes model or a support vector machine model.

[0110] It is understandable that due to the excessive development of the Internet, the sources of news content are too complex. In order to avoid the subsequent generation of erroneous related content, it is necessary to screen the news content.

[0111] The first processing module is used to construct an event evolution graph based on available news data, divide the available news data into M stages of news data according to the event evolution graph, and select M representative news data from the M stages of news data, where M is an integer greater than 0;

[0112] Specifically, the method of constructing an event evolution graph based on available news data includes:

[0113] Based on information extraction technology, the available news data is processed to obtain S sub-events. The temporal relationship between the S sub-events is determined according to the time labels. The causal relationship between the S sub-events is determined based on the graph neural network. An event evolution graph is constructed with the S sub-events as nodes and the temporal relationship and causal relationship between the S sub-events as edges.

[0114] Specifically, the graph neural network also takes each sub-event as a node and defines relevant features for each sub-event. The relevant features include the description, timestamp, category, etc. of the event, and then updates the state of the node through a message passing mechanism. Message passing includes receiving information from neighboring nodes and aggregating it to the current node. This mechanism can help capture the causal relationship between nodes. The above-mentioned information extraction and graph neural network are both existing technologies, and this embodiment will not go into details.

[0115] The method of dividing the available news data into M stages of news data according to the event evolution graph includes:

[0116] M event stages in the event evolution graph are obtained, and available news data are divided into M stage news data according to the nodes in the M event stages.

[0117] It should be noted that the available news data includes several available news sub-data. Each node in the event evolution graph represents a sub-event. Each sub-event has corresponding available news sub-data. According to the position of the node in the event evolution graph, the corresponding available news sub-data can be divided into which stage. For example, following the above, the available news data includes 3 available news sub-data, namely the first text content, the second text content and the third text content. From the above content, it can be seen that there are 3 event stages in the event evolution graph, so the available news data is divided into 3 stages of news data accordingly.

[0118] The method of selecting M representative news data from M stage news data includes:

[0119] Traverse the news data of M stages, calculate the consistency factor corresponding to each element in the news data of each stage, determine the representative news data according to the consistency factor, and thus screen out M representative news data.

[0120] In this embodiment, there are multiple available news sub-data for each stage of news data. The above elements represent the available news sub-data. By calculating the consistency factor, the available news sub-data with the largest consistency factor is selected as the representative news data.

[0121] Methods for calculating consistency factors include:

[0122] ;

[0123] In the formula, For the The consistency factor corresponding to the elements is For the stage news data The vector representation of elements is For the stage news data The vector representation of elements is is the total number of elements in the stage news data, For the The element and The Euclidean distance between elements.

[0124] It should be noted that the Euclidean distance can be used to measure the similarity or difference between two vectors. In multidimensional space, the smaller the Euclidean distance of the vectors, the more similar they are. Conversely, the larger the distance, the greater the difference. Therefore, in this embodiment, the Euclidean distance between an element and other elements is calculated respectively to characterize the similarity between an element and other elements. In this embodiment, the larger the consistency factor, the more similar an element is to other elements. Therefore, this element can be used as the representative news data in the news data of a stage. This ensures that the selected representative news data is most similar to other news data in this stage. This means that the selected news data can more accurately reflect the overall content and theme of the news data in this stage, avoid selecting individual data points that deviate from the main information, and make subsequent analysis more efficient.

[0125] Judgment module: used to input M representative news data into the pre-built semantic discrimination model in sequence to obtain semantic discrimination results. When the semantic discrimination result is semantic inconsistency, it returns to the first processing module. When the semantic discrimination result is semantic consistency, it transfers to the second processing module. The semantic discrimination result at least includes semantic inconsistency and semantic consistency.

[0126] In this embodiment, the method for constructing the semantic discrimination model includes:

[0127] The semantic discriminant model includes a generative adversarial network and a classifier, and the discriminator in the generative adversarial network is trained using the actual representative news data to obtain a trained discriminator;

[0128] The trained discriminator is used to guide the generator in the adversarial generative network to generate forged representative news data similar to the actual representative news data, where the forged representative news data represents news data with similar text content but different semantics;

[0129] The actual representative news data and the forged representative news data are input into the classifier for training to obtain the semantic discrimination results.

[0130] It should be noted that Generative Adversarial Networks (GANs) is a generative model proposed by deep learning. Its goal is to generate realistic data samples through adversarial learning between two network models. GANs consists of two main components: the generator and the discriminator. These two components compete with each other and gradually improve the quality of samples generated by the generator through adversarial training.

[0131] The second processing module is used to reconstruct the M representative news data to obtain intermediate news data, input the intermediate news data into the pre-built feature extraction model, obtain N feature triplets, and generate target news data based on the N feature triplets, where N is an integer greater than 0.

[0132] The method of reconstructing the M representative news data to obtain the intermediate news data includes:

[0133] Each entity in the representative news data is labeled, and a relationship prompt sentence for the entity pair is added to each labeled representative news data.

[0134] For example, representative news data could be "In mid-May 2024, a city officially announced that it would host an annual international marathon. Once the news was released, it attracted widespread attention. Runners from all over the world began to sign up, and preparations for the event were also carried out in full swing."

[0135] The reconstructed intermediate news data can be "In mid-May 2024, XX city XX officially announced that it will host the annual international XX marathon XX. Once the news was released, it attracted widespread attention. Runners from all over the world began to sign up, and preparations for the event were also carried out in full swing. The relationship between a city and the marathon is mask1, mask2 or mask3."

[0136] It should be noted that entities can represent keywords in news data, while masks refer to the relationship between entities. They will be replaced with specific relationship descriptions in actual use, for example:

[0137] mask1 may represent "host city", that is, "a certain city is the host city of a marathon".

[0138] mask2 may represent "location", i.e. "the marathon will be held in a certain city".

[0139] mask3 might stand for "event organization", i.e. "the organizing committee of a certain city is responsible for the preparation of the marathon".

[0140] In this way, by annotating entities and relationships, the reconstructed news data not only retains the main content of the original information, but also adds a clear description of the entity relationship, which helps to convey information more clearly and provides a structured information basis for subsequent feature extraction. This process enables the feature extraction model to more efficiently identify and extract important news elements. It can be understood that the above reconstruction of representative news data can be achieved through natural language processing technology, through entity recognition, relationship extraction, and template matching and filling. It can be understood that once the entities and relationships are identified, predefined templates (such as mask1, mask2, mask3) are used for placeholders. In actual applications, these placeholders will be replaced by specific relationship descriptions to generate structured intermediate news data.

[0141] The construction methods of the feature extraction model include:

[0142] A sample data set is obtained, wherein the sample data set includes historical intermediate news data and historical element triplets, the sample data set is divided into a sample training set and a sample test set, a regression network is constructed, the historical intermediate news data in the sample training set is used as input data of the regression network, and the historical element triplets in the sample training set are used as output data of the regression network, the regression network is trained to obtain an initial regression network for predicting real-time element triplets, the initial regression network is tested using the sample test set, and the initial regression network that satisfies a preset error value is output as a factor extraction model, and the initial regression network is preferably a deep neural network model.

[0143] Example 3

[0144] This embodiment discloses an electronic device, including a power supply, an interface, a keyboard, a memory, a central processing unit, and a computer program stored in the memory and executable on the central processing unit. When the central processing unit executes the computer program, the multi-perspective news creation method based on media viewpoints provided by the above-mentioned methods is implemented. The interface includes a network interface and a data interface. The network interface includes a wired or wireless interface, and the data interface includes an input or output interface.

[0145] Since the electronic device introduced in this embodiment is an electronic device used to implement the multi-perspective news creation method based on media viewpoints in the embodiment of this application, based on the multi-perspective news creation method based on media viewpoints introduced in the embodiment of this application, the technical personnel of this field can understand the specific implementation method of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application will not be introduced in detail here. As long as the technical personnel of this field implement the electronic device used in the multi-perspective news creation method based on media viewpoints in the embodiment of this application, it belongs to the scope of protection of this application.

[0146] Example 4

[0147] like Figure 4 As shown, this embodiment discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed, the multi-perspective news creation method based on media viewpoints is implemented.

[0148] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters, weights and thresholds in the formula are set by technicians in this field according to actual conditions.

[0149] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server or data center to another website, computer, server or data center via a wired network or a wireless network. The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media sets. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD) or a semiconductor medium. The semiconductor medium may be a solid-state hard disk.

[0150] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the present invention can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0151] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0152] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only one, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0153] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0154] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0155] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

[0156] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-perspective news creation method based on media viewpoints, characterized in that: include: S10: Acquire real-time news data and news source information, and divide the real-time news data into available news data and unavailable news data according to the news source information; S20: constructing an event evolution graph based on the available news data, dividing the available news data into M stages of news data according to the event evolution graph, and selecting M representative news data from the M stages of news data, where M is an integer greater than 0; S30: inputting the M representative news data into the pre-built semantic discrimination model in sequence to obtain a semantic discrimination result. When the semantic discrimination result is semantic inconsistency, the process returns to S20. When the semantic discrimination result is semantic consistency, the process proceeds to S40. The semantic discrimination result includes at least semantic inconsistency and semantic consistency. S40: reconstructing the M representative news data to obtain intermediate news data, inputting the intermediate news data into a pre-built factor extraction model to obtain N factor triplets, and generating target news data based on the N factor triplets, where N is an integer greater than 0; Among them, the method of constructing an event evolution graph based on available news data includes: Based on information extraction technology, available news data is processed to obtain S sub-events. The temporal relationship between the S sub-events is determined based on the time tags. The causal relationship between the S sub-events is determined based on the graph neural network. The event evolution graph is constructed with the S sub-events as nodes and the temporal relationship and causal relationship between the S sub-events as edges. The method of selecting M representative news data from M stage news data includes: Traverse the news data of M stages, calculate the consistency factor corresponding to each element in the news data of each stage, determine the representative news data according to the consistency factor, and thus screen out M representative news data; Methods for calculating consistency factors include: ； In the formula, For the The consistency factor corresponding to the elements is For the stage news data The vector representation of elements is For the stage news data The vector representation of elements is is the total number of elements in the stage news data, For the The element and The Euclidean distance between elements.

2. The multi-perspective news creation method based on media viewpoints according to claim 1 is characterized in that: The construction methods of the partitioning model include: Acquire Q groups of data, where Q is a positive integer greater than 1, and the data includes historical news source information, historical news data, and historical division results. Use the historical news source information, historical news data, and historical division results as a sample set, divide the sample set into a training set and a test set, build a classifier, use the historical news source information and historical news data in the training set as input data, use the historical division results in the training set as output data, train the classifier to obtain an initial classifier, test the initial classifier using the test set, and output a classifier that meets a preset accuracy as a division model.

3. The multi-perspective news creation method based on media viewpoints according to claim 1 is characterized in that: The construction method of the semantic discrimination model includes: The semantic discriminant model includes a generative adversarial network and a classifier, and the discriminator in the generative adversarial network is trained using actual representative news data to obtain a trained discriminator; The trained discriminator is used to guide the generator in the adversarial generative network to generate forged representative news data similar to the actual representative news data, where the forged representative news data represents news data with similar text content but different semantics; The actual representative news data and the forged representative news data are input into the classifier for training to obtain the semantic discrimination results.

4. The multi-perspective news creation method based on media viewpoints according to claim 3 is characterized in that: When the semantic determination result is semantic inconsistency, before returning to S20, the following steps are also included: The marked news data in the semantic discrimination result is obtained, and the marked representative news data is removed from the corresponding stage news data.

5. The multi-perspective news creation method based on media viewpoints according to claim 4 is characterized in that: The method of reconstructing the M representative news data to obtain the intermediate news data includes: Each entity in the representative news data is labeled, and a relationship prompt sentence for the entity pair is added to each labeled representative news data.

6. A multi-perspective news creation system based on media viewpoints, which is used to implement the multi-perspective news creation method based on media viewpoints as described in any one of claims 1 to 5, characterized in that: include: Division module: used to obtain real-time news data and news source information, and divide the real-time news data into available news data and unavailable news data according to the news source information; The first processing module is used to construct an event evolution graph based on available news data, divide the available news data into M stages of news data according to the event evolution graph, and select M representative news data from the M stages of news data, where M is an integer greater than 0; Judgment module: used to input M representative news data into the pre-built semantic discrimination model in sequence to obtain semantic discrimination results. When the semantic discrimination result is semantic inconsistency, it returns to the first processing module. When the semantic discrimination result is semantic consistency, it transfers to the second processing module. The semantic discrimination result at least includes semantic inconsistency and semantic consistency. The second processing module is used to reconstruct the M representative news data to obtain intermediate news data, input the intermediate news data into the pre-built feature extraction model, obtain N feature triplets, and generate target news data based on the N feature triplets, where N is an integer greater than 0.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed, implements the multi-perspective news creation method based on media viewpoints as described in any one of claims 1 to 5.

Images