Data processing method and device, electronic equipment and readable medium

Abstract

The application provides a data processing method and device, electronic equipment and readable medium, comprising: obtaining a content generation request for a target topic, the request containing description information of the target topic; performing media retrieval according to the description information of the target topic to obtain first candidate information; querying a target event node related to the target topic from a preset event knowledge graph, the event knowledge graph containing a plurality of event nodes and event association edges, the event node containing second candidate information related to the event; merging the first candidate information and the second candidate information contained in the target event node to obtain a candidate information set; and according to the correlation degree between the target topic and the candidate information in the candidate information set, screening and combining the candidate information matched by the target topic from the candidate information set to obtain content data of the content generation request. The method can improve the richness of recalled information and reduce the correlation deviation, which is beneficial to improving the result accuracy.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more particularly to a data processing method, apparatus, electronic device, and readable medium. Background Technology

[0002] "Special features" are an important content format that aggregates and structures fragmented information. When faced with breaking news or long-standing social issues, operators need to quickly create special features to meet users' needs for a comprehensive understanding of the event. Accurately collecting information related to the special feature's theme is one of the core elements in building a high-quality feature.

[0003] In related technologies, the basic document slices selected by the user are converted into vectors. Then, based on the query statement triggered by the user, several information fragments most similar to the vectors are retrieved. Thematic content is obtained by summarizing these information fragments.

[0004] However, in such methods, information retrieval is mainly based on vector similarity retrieval from the vectors converted from documents. The scope of information retrieval is limited to a finite range directly related to the original text. Insufficient information retrieval scope leads to reduced richness of recalled information and high relevance deviation, affecting the accuracy and reliability of the generated results. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a data processing method, apparatus, electronic device, and readable medium to improve the richness of recall information and reduce the relevance bias of recall information, thereby enhancing the accuracy and reliability of the generated results.

[0006] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0007] According to one aspect of the embodiments of this application, a data processing method is provided, including:

[0008] Obtain a content generation request for a target topic, wherein the content generation request contains descriptive information about the target topic; Media retrieval is performed based on the description information of the target topic to obtain the first candidate information; The target event nodes related to the target topic are queried from a preset event knowledge graph. The event knowledge graph contains multiple event nodes and event association edges. The event nodes are used to represent events and contain second candidate information related to the events. The event association edges are used to represent the association relationship between events. The first candidate information is merged with the second candidate information contained in the target event node to obtain a candidate information set; Based on the degree of correlation between the target topic and the candidate information in the candidate information set, candidate information matching the target topic is filtered and combined from the candidate information set to obtain the content data of the content generation request.

[0009] According to one aspect of the embodiments of this application, a data processing apparatus is provided, comprising: The content acquisition module is configured to acquire a content generation request for a target topic, wherein the content generation request includes descriptive information of the target topic; The media retrieval module is configured to perform media retrieval based on the description information of the target topic to obtain first candidate information; The graph query module is configured to query target event nodes related to the target topic from a preset event knowledge graph. The event knowledge graph contains multiple event nodes and event association edges. The event nodes are used to represent events and contain second candidate information related to the events. The event association edges are used to represent the association relationships between events. The information merging module is configured to merge the first candidate information with the second candidate information contained in the target event node to obtain a candidate information set; The information filtering module is configured to filter and combine candidate information matching the target topic from the candidate information set based on the degree of relevance between the target topic and the candidate information in the candidate information set, so as to obtain the content data of the content generation request.

[0010] In the embodiments of this application, based on the above technical solution, the graph query module is specifically configured to: locate a first event node corresponding to the target topic in the preset event knowledge graph according to the description information of the target topic; traverse the event nodes located downstream of the first event node in the event knowledge graph according to the event association edge with the first event node to obtain a second event node; and use the first event node and the second event node together as target event nodes related to the target topic.

[0011] In the embodiments of this application, based on the above technical solution, the information filtering module is specifically configured to: obtain a first event node corresponding to the target topic in the event knowledge graph; if the first event node contains a predetermined information sequence set for the target topic, obtain candidate information corresponding to the predetermined information sequence from the candidate information set, and combine them to obtain a first content chapter, wherein the candidate information in the predetermined information sequence is selected based on its relevance to the target topic; fill the first content chapter into the corresponding position in the predetermined content structure to obtain the content data of the content generation request.

[0012] In the embodiments of this application, based on the above technical solution, the information filtering module is further configured to: if the first event node does not contain the predetermined information sequence, generate a first prompt word based on the description information of the target topic, the summary information of the candidate information in the candidate information set, and the predetermined prompt information template, wherein the first prompt word is used to guide the large language model to generate a topic information sequence; input the first prompt word into the large language model so that the large language model can filter and sort the candidate information according to the relevance of the candidate information to the target topic to obtain the topic information sequence output by the language model; obtain the candidate information corresponding to the topic information sequence from the candidate information set and combine them to obtain a first content chapter.

[0013] In the embodiments of this application, based on the above technical solution, the information filtering module is further configured to: if the content generation request contains a deep analysis identifier, then select the remaining candidate information whose similarity is lower than a predetermined threshold according to the similarity between other candidate information in the candidate information set and candidate information in the first content chapter, wherein the other candidate information is candidate information in the candidate information set that does not belong to the first content chapter; generate a second prompt word according to the remaining candidate information and a predetermined prompt information template, the second prompt word being used to guide the large language model to generate multiple sub-topics according to the remaining candidate information; input the second prompt word into the large language model to obtain multiple sub-topics output by the large language model and the remaining candidate information corresponding to each sub-topic; combine the remaining candidate information corresponding to each sub-topic to obtain multiple second content chapters, the second content chapters being used to fill the corresponding positions in the predetermined content structure.

[0014] In the embodiments of this application, based on the above technical solution, the information filtering module is further configured to: obtain information editing records of the client for the generated content data, the information editing records containing modification operations on candidate information contained in the content data; and add constraint information based on the modification operations in the information editing records to the prompt information template, the constraint information being used to indicate the large language model.

[0015] In the embodiments of this application, based on the above technical solution, the information merging module is specifically configured to: obtain second candidate information corresponding to each of the target event nodes; calculate the union between the first candidate information and the second candidate information according to the information identifiers of the first candidate information and the second candidate information to obtain a union set; score the relevance between the description information of the target topic and the summary information of each candidate information in the union set using a large language model to obtain a relevance score corresponding to each candidate information; and delete candidate information whose corresponding relevance score is lower than a predetermined score threshold from the union set according to the relevance score corresponding to each candidate information to obtain a candidate information set.

[0016] In the embodiments of this application, based on the above technical solution, the information merging module is specifically configured to: delete candidate information whose corresponding relevance score is lower than a predetermined score threshold from the union set according to the relevance score corresponding to each candidate information, to obtain a filtering result set; cluster the candidate information in the filtering result set according to the semantics of each candidate information, to obtain multiple information clusters; for each information cluster, select candidate information whose quality score meets a predetermined score condition according to the quality score of each candidate information in the information cluster, to obtain an information subset corresponding to each information cluster; merge the information subsets corresponding to each information cluster to obtain a candidate information set.

[0017] In the embodiments of this application, based on the above technical solution, the information merging module is further configured to: extract meta-information of each candidate information in the union set, wherein the meta-information includes at least one of information source, quality score and information status; and delete candidate information that does not meet predetermined filtering conditions from the union set according to the meta-information, wherein the predetermined filtering conditions include constraints set for the meta-information.

[0018] In the embodiments of this application, based on the above technical solution, the content acquisition module is specifically configured to: receive a generation instruction sent by a client, the generation instruction including setting a topic name, topic description and system parameters for a target topic, the system parameters being used to filter candidate information of the target topic according to predetermined filtering conditions; concatenate the topic name and the topic description to form the description information of the target topic; and generate a content generation request for the target topic based on the description information and the system parameters.

[0019] In the embodiments of this application, based on the above technical solution, the information filtering module is further configured to: display a content draft interface through a client, the content draft interface including a first component, a second component, and a third component, the first component displaying the names of each content chapter in the content data, the second component displaying an information preview interface corresponding to the content chapter selected in the first component, and the third component displaying an editing interface for the content chapter selected in the first component; updating the content data and sending it to the client according to the information editing instruction sent by the client, so as to display the editing result in the third component, wherein the information editing instruction is triggered by an information editing operation in the third component; generating topic information containing the content data and sending it to the client according to the confirmation instruction sent by the client, so as to display the information display page of the target topic in the client.

[0020] In the embodiments of this application, based on the above technical solution, the description information includes the topic name and topic description information of the target topic; the media retrieval module is specifically configured to: extract keywords from the topic name and the topic description information to obtain topic keywords; retrieve first recall information associated with the topic keywords based on the topic keywords; concatenate and embed the topic name and the topic description information to obtain an embedding vector corresponding to the target topic; retrieve similar information vectors similar to the embedding vector in the vector database based on the embedding vector to obtain second recall information corresponding to the similar information vector, wherein the vector database contains multiple information vectors and corresponding event information; and merge the first recall information and the second recall information as first candidate information.

[0021] According to one aspect of the embodiments of this application, an electronic device is provided, the electronic device comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to perform a data processing method as described above by executing the executable instructions.

[0022] According to one aspect of the embodiments of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the data processing method as described in the above technical solutions.

[0023] According to one aspect of the embodiments of this application, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the data processing methods provided in the various optional implementations described above.

[0024] In an embodiment of this application, the solution obtains a content generation request for a target topic, the content generation request containing descriptive information of the target topic. Subsequently, the solution performs media retrieval based on the descriptive information of the target topic to obtain first candidate information. The solution queries a preset event knowledge graph for target event nodes related to the target topic. The event knowledge graph contains multiple event nodes and event-related edges. The event nodes represent events and contain second candidate information related to the events, and the event-related edges represent the relationships between events. Next, the solution merges the first candidate information with the second candidate information contained in the target event nodes to obtain a candidate information set. Then, based on the relevance between the target topic and the candidate information in the candidate information set, the solution filters and combines candidate information matching the target topic from the candidate information set to obtain the content data of the content generation request. By introducing event knowledge graph queries, the solution obtains candidate information related to the target topic based on the logical relationships between events in the knowledge graph. This avoids incomplete information retrieval and missing logical relationships caused by relying on single semantic retrieval, thereby improving the richness of the retrieved information and reducing the relevance bias of the retrieved information, which is conducive to improving the accuracy and reliability of the generated results.

[0025] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0026] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0027] Figure 1 The data processing method described in this application is applied to the system architecture of a data processing system.

[0028] Figure 2 This is an overall framework diagram of the data processing system in the embodiments of this application.

[0029] Figure 3 This is a flowchart of a data processing method according to an embodiment of this application.

[0030] Figure 4 This is a schematic flowchart illustrating the topic information retrieval process in the embodiments of this application.

[0031] Figure 5This is a schematic flowchart illustrating the content construction process in the embodiments of this application.

[0032] Figure 6 This is a schematic diagram illustrating the screening of candidate information through clustering in an embodiment of this application.

[0033] Figure 7 This is a schematic flowchart illustrating the prompt word correction process in an embodiment of this application.

[0034] Figure 8 This is a schematic diagram illustrating the overall process of generating content data in the embodiments of this application.

[0035] Figure 9 A schematic block diagram of the data processing apparatus in an embodiment of this application is shown.

[0036] Figure 10 A schematic diagram of the structure of a computer system suitable for implementing the electronic device of the present application is shown. Detailed Implementation

[0037] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0038] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0039] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve the predetermined function, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0040] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0041] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0042] First, let me introduce the relevant terms used in this application.

[0043] Candidate information: The candidate information in this application is usually publicly released information, including various forms such as text, images, videos, and short articles; in this application, the candidate information is added to the topic in the form of links, citations, or direct display, and becomes part of the topic content.

[0044] An event can be represented by a single text description, which typically includes key elements such as the time, place, people, and actions of the event. An event usually has several related candidate information, including various forms such as images, videos, short texts, and Q&A. In this application, an event is a component of a topic. A topic usually contains one or more core events to form the core development path of the topic, and may also include related events as supplementary information to the topic theme.

[0045] A topic, in essence, is also a form of textual description, but it doesn't necessarily have specific key elements such as time, place, people, or actions. Instead, it's a topic or event that can spark discussion. Each topic also has several related candidate information, including various forms such as text and images, videos, short texts, and Q&A. In this application, a topic or topic draft is a form of content generated by the application's solution. The application's solution collects information on relevant events based on the required topic and combines it to generate the specific content of the topic.

[0046] Event knowledge graph: A semantic network that describes the interrelationships between events using a graph structure. The graph in this application is a hierarchical knowledge graph structure, containing the hierarchy of topic-medium event-small event-candidate information. Each medium event node and small event node in the graph usually carries multiple candidate information related to the event.

[0047] It should be understood that the solution presented in this application can be applied to topic management systems based on information publishing platforms such as news platforms, especially in the process of generating topic content on the server. Specifically, the solution presented in this application can be applied in the following scenarios: I. Research Report Generation. In rapidly changing information environments such as financial markets, relevant personnel often need to quickly gather information, announcements, and news from across the internet to write in-depth investment analysis reports on macroeconomic events (such as "bank interest rate hikes"), industry policy changes, or sudden corporate situations (such as "executive resignations" or "supply chain disruptions"). For example, when dealing with new policy adjustments, key personnel changes, or sudden major events announced by a major authoritative institution, it is necessary to quickly understand relevant information. In this scenario, the system receives a topic generation request from the user. First, it uses text and vector dual-path retrieval to retrieve basic materials from online information and corporate announcement databases. Simultaneously, based on a pre-built "financial macro-micro event knowledge graph," the system locates the node that triggers the current event and retrieves related upstream and downstream supply chain reaction events along causal and sequential correlation edges, such as the pressure on downstream profits caused by upstream raw material price increases. After cross-scoring and semantic clustering using a large language model to remove illegal or low-relevance noise information, the system automatically assembles a draft content containing standard chapters such as "Event Summary" and "Market Impact" using preset templates. Furthermore, through in-depth analysis of the remaining candidate information, the system automatically extracts long-tail "potential investment opportunities or risk warnings" as supplementary chapters, which are then presented to analysts for review and publication in an interactive draft interface. This allows for the rapid generation of relevant reports, thereby improving efficiency and reducing labor costs. Moreover, the generated reports include indirectly related information such as related industries, thus enhancing the ability to identify risks transmitted through the industrial chain.

[0048] II. Breaking News Features. When faced with breaking news events, news media need to grasp the development of the news, analyze the event's timeline, and output objective and accurate news features to quickly provide news broadcasts. Upon detecting a breaking event and automatically triggering a news feature generation request, the system uses the union of information identifiers to integrate mainstream media reports with the evolution of similar historical events. During the data cleaning phase, the system utilizes metadata filtering conditions, such as limiting the authority of information sources and removing "unverified" information, supplemented by relevance scoring from a large language model, to extract objective facts. If no existing template exists for the news event, the system will generate prompts to guide the large language model to automatically plan the event's timeline and situation outline, piecing together fragmented materials into briefing sections such as "Event Cause," "Current Situation," and "Responses from All Parties." The final generated briefing can be fine-tuned and annotated by editors in a multi-perspective interactive interface, thereby shortening the response cycle for news events.

[0049] III. Insight Report Writing. When exploring cutting-edge technologies, industry research institutions, university think tanks, or consulting firms often face a massive and fragmented collection of Chinese and foreign academic papers, patent documents, and industry trends. They need to clarify the technological evolution and generate comprehensive insight reports. The system performs vector retrieval of the scientific literature database based on the topic description and filtering parameters input by researchers to obtain technical discussions. It also extracts key nodes and breakthroughs from each technological iteration from the "Knowledge Graph of Technological Evolution Events." The system performs semantic clustering and integration on these two types of data to generate standard overview content on the main line of technological development. Furthermore, the system conducts in-depth analysis, performing secondary clustering and summarizing of cutting-edge literature that did not enter the main line of content but had low similarity but was still relevant, generating independent chapters such as "Discussion of Edge Technology Routes" or "Analysis of Technological Bottlenecks," expanding the research perspective and breadth of thinking in the think tank report.

[0050] "Special topics" are an important content format that aggregates and structures fragmented information. Faced with sudden trending events or long-standing social issues, operators need to quickly create special topics to meet users' needs for a comprehensive understanding of the event. Accurately collecting information related to the topic's theme is one of the core elements in building high-quality special topics. In related technologies, user-selected document slices are converted into vectors, and then the most similar information fragments to the vectors are retrieved based on the user's query. The topic content is obtained by summarizing these information fragments. However, in this method, information retrieval is mainly based on vector similarity retrieval from the converted documents. The scope of information retrieval is limited to a finite range directly related to the original text. Insufficient information retrieval scope leads to reduced richness of retrieved information and higher relevance bias, affecting the accuracy and reliability of the generated results.

[0051] Based on this, the technical solution of this application embodiment proposes a data processing scheme. Specifically, please refer to... Figure 1The data processing method according to the embodiments of this application is applied to the system architecture of a data processing system. As can be seen, the system mainly includes a server 110 and a client 120. The server 110 is equipped with a data processing system used to recall relevant articles based on topic titles and descriptions provided by manual operations, and to structure them using large models and knowledge graphs, generating an initial draft according to operational needs. The client 120 is used to collect topic titles and descriptions provided by users, as well as relevant topic settings parameters, and to display the generated initial draft to users through an interactive interface. Users modify and confirm the publication of the draft through the client. The server 110 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. The client 120 can include smartphones, tablets, laptops, smart voice interaction devices, smart home appliances, vehicle terminals, aircraft, etc. The communication link between server 110 and client 120 can be a communication medium connection of various connection types, such as a wired communication link or a wireless communication link. Depending on the implementation requirements, the system architecture in this embodiment can have any number of terminal devices and servers.

[0052] Specifically, server 110 can receive a content generation request for a target topic from client 120. The content generation request includes descriptive information input by the user in client 120, typically including the topic name and a brief introduction. The user can also select system parameters for the topic in client 120, such as topic type (e.g., dynamic topic, discussion topic), information source (e.g., official or self-media), and topic template (e.g., in-depth analysis or quick analysis). Server 110 performs media retrieval based on the descriptive information of the target topic, obtaining related articles containing the same keywords or articles with sufficiently high vector similarity as first-order candidate information. Furthermore, server 110 queries a preset event knowledge graph for target event nodes related to the target topic. The event knowledge graph contains numerous relationships between topics, events, and articles. Relationships between events can be established according to chronological order, causality, and event decomposition. Each event has multiple related articles attached; articles attached to events related to the target topic in the event knowledge graph are considered first-order candidate information. Server 110 merges the first-order candidate information with the second-order candidate information contained in the target event nodes to obtain a candidate information set. During the merging process, server 110 deduplicates and filters the candidate information, removing low-quality information. Then, server 110 selects candidate information for the target topic from the candidate information set based on their relevance to the target topic, obtaining the content data for the content generation request. The generated content data can be displayed to the user as a draft of the topic through client 120, where the user can edit the draft content and publish the topic.

[0053] Please see Figure 2 , Figure 2 This is an overall framework diagram of the data processing system in an embodiment of this application. Figure 2As shown, the data processing system 200 comprises four modules: an information aggregation module 210, an information optimization module 220, a content construction module 230, and a human-computer collaboration module 240. The information aggregation module 210 is used to retrieve articles through multiple channels and aggregate them into an information set. Specifically, in step 211, the information aggregation module 210 retrieves articles through media retrieval; in step 212, it searches for related events in the knowledge graph; and then in step 213, it retrieves articles associated with these events. Furthermore, in step 214, the information aggregation module 210 searches for related sub-events based on the events, and in step 215, it retrieves articles associated with these sub-events. These retrieved articles are aggregated into an article data pool 216. The information optimization module 220 is used to filter the articles in the article data pool 216. Specifically, in step 221, the information optimization module 220 filters articles based on their quality, and in the process, it can filter out articles that do not meet the predetermined hard rules. In step 222, the information optimization module 220 filters out articles with low relevance to the topic using a large model. Finally, in step 223, semantic clustering is performed on the articles, and a final subset of articles from each cluster is selected as the result of the information optimization module 220. Based on the selection result of the information optimization module 220, the content construction module 230 transforms the cleaned unstructured articles into structured topic modules. Specifically, the content construction module 230 can fill in relevant articles according to predetermined chapters and use the large model to generate multiple chapters based on the remaining articles, combining them into a topic draft. The human-computer collaboration module 240 can modify, confirm, and publish the generated topic draft. The records of its operations can be used to adjust the prompts input into the large model when constructing the topic draft in the content construction module 230, thereby optimizing the accuracy of the generated draft.

[0054] The implementation details of the technical solutions in the embodiments of this application are described in detail below: Figure 3 A flowchart of a data processing method according to an embodiment of this application is shown. This data processing method can be executed by a device with computing capabilities, such as an enterprise-level server, a cloud computing cluster, or a user's local smart terminal device. This application uses a server as an example for description. Figure 3 As shown, this data processing method includes at least steps S310 to S350, which are described in detail below: Step S310: Obtain a content generation request for the target topic, wherein the content generation request contains descriptive information of the target topic.

[0055] Content generation requests typically originate from client-triggered requests. The descriptive information of the target topic in the content generation request can include text or other forms of data input by the user through the client. In practical applications, content generation requests can be triggered by user input through a human-computer interaction interface (such as a client's search box, dialog window, or form), or they can be data call instructions automatically initiated by upstream business systems through application programming interfaces (APIs). The target topic usually indicates the core intent or subject of discussion for this content generation task, such as a specific news event, market dynamics of a certain industry, or the life story of a certain person. The descriptive information is a specific portrayal of the target topic, and its data form can include, but is not limited to: long sentences in natural language, short text keyword combinations, query statements, and constraints related to the generation task (such as time range, geographical restrictions, etc.). As an example implementation, suppose the current system is a financial information platform. A user wants to generate an analysis report on "Company A's production capacity decline in the third quarter". The content generation request obtained at this time is a request to generate a financial analysis report. The target topic is "Company A's abnormal production capacity", and the description information may be a specific natural language input: "Please collect and organize the specific situation, relevant background and possible reasons for Company A's production capacity decline in the third quarter of this year".

[0056] Step S320: Perform media retrieval based on the description information of the target topic to obtain the first candidate information.

[0057] Upon receiving the aforementioned descriptive information, the server performs coarse and fine ranking searches within a pre-defined media corpus based on this information. Specific search techniques may include keyword-based literal matching and deep learning-based vector retrieval. In vector retrieval, the descriptive information can be mapped into multi-dimensional semantic feature vectors using a pre-defined algorithm or pre-trained language model. Simultaneously, the vector similarity between these vectors and various media content in the corpus is calculated, thereby recalling one or more semantically most similar media content. These documents, paragraphs, sentence fragments, images, or audio / video content recalled through media retrieval serve as the first candidate information for the target topic. In the example of "Company A's abnormal production capacity," the system may retrieve relevant original news reports, analyst commentary screenshots, and other first candidate information. Typically, this information is unstructured or semi-structured, encompassing the background context of the target topic.

[0058] Step S330: Query target event nodes related to the target topic from a preset event knowledge graph. The event knowledge graph contains multiple event nodes and event association edges. The event nodes are used to represent events and contain second candidate information related to the events. The event association edges are used to represent the association relationship between events.

[0059] Furthermore, the server queries a pre-defined event knowledge graph for target event nodes related to the target topic. The event knowledge graph represents the relationships between multiple events. It is typically pre-built; for example, it can be constructed manually or by a system for dynamic events occurring in the real or virtual world. The server then queries the event knowledge graph for event nodes corresponding to the target topic. The event knowledge graph contains multiple event nodes and event-related edges. Event nodes represent dynamic events occurring in the real or virtual world, such as "price increases," "factory shutdowns," and "policy releases." Each event node contains secondary candidate information related to the event, such as articles from different sources, including structured or semi-structured attribute descriptions such as trigger words, occurrence time, participating entities, and location. Event-related edges represent the logical relationships between events, including but not limited to causal relationships, temporal relationships, sequential relationships, and sub-event relationships. In practice, the server performs named entity recognition and event extraction on the descriptive information, extracting core event trigger words and arguments. These are then mapped to a pre-defined event knowledge graph to locate the corresponding topic node. Based on event association edges, one or more matching target event nodes are retrieved. Furthermore, the server performs subgraph walks or multi-hop queries along the event association edges of these target event nodes to recall logically connected adjacent event nodes. In the example above, the server matched the target event node "Company A suspended operations in the third quarter" in the event knowledge graph and, through causal relationship association edges, retrieved another connected event node, "Global core component supply chain disruption." The system extracts the structured summaries and argument attributes recorded in these event nodes and uses them as second-stage candidate information.

[0060] Step S340: Merge the first candidate information with the second candidate information contained in the target event node to obtain a candidate information set.

[0061] Subsequently, the server merges the first candidate information with the second candidate information contained in the target event node to obtain a candidate information set. Since the first candidate information is the basic context obtained from a wide-ranging media retrieval, while the second candidate information is usually core information that is more closely related to the corresponding event and has been filtered, the system can employ various strategies for fusion during the merging process. For example, besides simply adding the information to the same set, the structured fields of the second candidate information can be transformed into natural language sentences through template filling, such as transforming "node 1-causal edge-node 2" into "because of event A, event B occurred," and then added to the candidate information set along with the media content of the first candidate information; alternatively, the two types of information can be mapped to the same multimodal or multidimensional vector space for unified expression. Simultaneously, data cleaning and deduplication operations can be performed during the merging process to remove redundant data that completely overlaps in the descriptions of the first and second candidate information, thereby forming a candidate information set with high information richness and diversity in perspectives and sources.

[0062] Step S350: Based on the degree of correlation between the target topic and the candidate information in the candidate information set, filter and combine the candidate information that matches the target topic from the candidate information set to obtain the content data of the content generation request.

[0063] Finally, the server filters candidate information related to the target topic from the candidate information set based on their relevance to the target topic, obtaining the content data for the content generation request. Specifically, since the merged candidate information set inevitably contains some noisy data or edge information, the server can use a re-ranking model to score each candidate information in the set, for example, through a cross-attention mechanism model. The scoring is based on the semantic and logical fit between the information and the target topic description. Based on the scoring results, the system sets a preset relevance threshold or directly selects the top N candidate information by score for retention. The high-quality information retained after rigorous screening is the final content data. This content data can be directly returned to the user as the final result, or it can be used as background context input to a large language model (LLM) to drive the language model to generate the final coherent text. In the example above, the server filters out noise interference related to "other irrelevant product lines of Company A", combines the extracted original news background with the extracted causal logic chain of "supply chain disruption leading to work stoppage", and outputs high-quality content data to help generate a well-reasoned financial analysis report.

[0064] In an embodiment of this application, the solution obtains a content generation request for a target topic, the content generation request containing descriptive information of the target topic. Subsequently, the solution performs media retrieval based on the descriptive information of the target topic to obtain first candidate information. The solution queries a preset event knowledge graph for target event nodes related to the target topic. The event knowledge graph contains multiple event nodes and event-related edges. The event nodes represent events and contain second candidate information related to the events, and the event-related edges represent the relationships between events. Next, the solution merges the first candidate information with the second candidate information contained in the target event nodes to obtain a candidate information set. Then, based on the relevance between the target topic and the candidate information in the candidate information set, the solution filters and combines candidate information matching the target topic from the candidate information set to obtain the content data of the content generation request. By introducing event knowledge graph queries, the solution obtains candidate information related to the target topic based on the logical relationships between events in the knowledge graph. This avoids incomplete information retrieval and missing logical relationships caused by relying on single semantic retrieval, thereby improving the richness of the retrieved information and reducing the relevance bias of the retrieved information, which is conducive to improving the accuracy and reliability of the generated results.

[0065] In the embodiments of this application, based on the above technical solution, during the process of querying target event nodes related to the target topic from a preset event knowledge graph, the server can locate a first event node corresponding to the target topic in the preset event knowledge graph according to the description information of the target topic. Next, the server traverses the event nodes located downstream of the first event node in the event knowledge graph according to the event association edges with the first event node to obtain a second event node. Then, the server uses the first event node and the second event node together as the target event node related to the target topic. In this embodiment, the server extracts logically coherent event information from the event knowledge graph through a directed traversal process based on the graph topology. Specifically, in the preset event knowledge graph, the data structure is usually presented as a large-scale directed graph, which not only stores the fact that the event occurred, but also implies the start and end of the event development. Specifically, the server can extract core events that can serve as "anchors" from unstructured description information input by the user or transmitted upstream. The server can invoke a natural language processing model to perform lexical analysis, syntactic dependency analysis, and semantic role labeling to accurately extract action trigger words and related arguments such as event participants, time, and location from the descriptive information. After extracting these features, the server can use entity disambiguation and event alignment algorithms to compare the extracted features with the attribute labels carried by each node in the event knowledge graph, thereby anchoring one or more initial matching nodes in the graph. These successfully anchored nodes, which directly correspond to the user's original input intent, can be topic nodes specifically corresponding to a particular topic in the event knowledge graph, or event nodes at the level of a topic. When the matching node is an event node, it can be directly used as the first event node; when the matching node is a topic node, the subordinate event nodes associated with the topic node are used as the first event node. As an example implementation, suppose the topic of the content generation request currently received by the server is a macroeconomic event, and its descriptive information is "summarizing the specific actions of recent interest rate hikes and their subsequent impacts". The system uses semantic parsing to identify "interest rate hike" as the core triggering event and locates the node representing this specific action—"Increase in benchmark interest rate in Month X, 202X"—in a pre-defined event knowledge graph. This node serves as the starting anchor point for graph traversal, i.e., the first event node. Subsequently, the server traverses the event knowledge graph downstream of the first event node based on the event-related edges associated with it, obtaining the second event node. In the event knowledge graph, because events themselves possess dynamic development and mutual triggering characteristics, event-related edges typically have a clear directionality. This directionality constitutes the "upstream" and "downstream" topological relationship.Downstream event nodes typically refer to sequential events that occur later than the first event node in chronological order, or consequential events that are directly or indirectly caused by the first event node in logical terms. Starting from the first event node, the server executes a graph traversal algorithm along event-related edges with specific directions (such as time edges pointing to the future or causal edges pointing to the result). Specific traversal strategies can employ breadth-first search (BFS) to obtain all direct subsequent effects of the first event node, or depth-first search (DFS) to uncover deeper subsequent events along a specific evolutionary path. In some embodiments, to prevent graph combinatorial explosion or circular dependencies during traversal, the server can set traversal constraints. For example, a maximum hop count limit can be set, traversing only downstream nodes within 3 hops of the first event node; or a weight threshold can be set based on the association weights attached to the event-related edges, representing the probability or strength of the correlation between two events, allowing traversal only along edges with association strength greater than a preset threshold. Through this controlled directed graph traversal, the server can automatically uncover related events that, while not directly mentioned in the user's original description, are objectively inevitable or will occur subsequently. See also. Figure 4 , Figure 4 This is a schematic flowchart illustrating the topic information retrieval process in an embodiment of this application. For example... Figure 4As shown, the information retrieval process is based on the topic title and topic indirection 401. The retrieval process is divided into two paths: keyword and semantic retrieval path 402 and knowledge graph retrieval path 409. In keyword and semantic retrieval 402, the input topic title and introduction are first parsed and keywords are extracted in 403, and keyword retrieval is performed based on the extracted keywords in 404. At the same time, the topic title and introduction are concatenated in 405 and converted into a vector in 406. The server retrieves the vector in the vector database in 407. The keyword and semantic retrieval path 402 finally outputs articles 408 that match the literal and semantic meaning based on 404 and 407. In the knowledge graph retrieval path 409, the server maps the input information to the knowledge graph in 410, thereby locating the event node in 411. Subsequently, in 412, the server traverses the articles attached to the event node based on the event node, thereby outputting event-related articles 413. The outputs of the two paths are combined in 414 to form the candidate article pool 415. Continuing with the example of interest rate hikes, the system uses "the benchmark interest rate was raised in Month X of 202X" as the first event node and traverses downstream along the directed edges labeled "caused (causal)" and "subsequently occurred (chronological)". The system might find the second event node "commercial banks raise loan interest rates" in the first hop, "corporate financing costs rise" and "real estate market transaction volume shrinks" in the second hop, and the cross-domain second event node "demand for building materials declines" in the third hop. These nodes at the back end of the logical and temporal evolution chain are collectively referred to as second event nodes. Finally, the server uses both the first and second event nodes as the target event node related to the target topic. The first event node represents the source or current aspect of the target topic, while the second event node represents the development trajectory and subsequent impact of the topic. The system combines and encapsulates these two types of nodes, outputting a logically coherent sequence of events. This sequence not only includes the basic facts explicitly queried by the user but also the event development path and related peripheral information. In the example of the economic analysis report generated above, the system combines "interest rate hike" with a series of nodes obtained through traversal, such as "increased loan interest rates," "rising corporate financing costs," and "shrinking real estate transaction volume," as the target event node set. When the candidate information from these nodes is used to generate the final content data, the generated Large Language Model (LLM) directly obtains a complete event development chain from policy release to its impact on the real economy, thereby enhancing the professionalism of the generated content.By using the above method, the solution takes the first event node directly corresponding to the user's intent and the second event node derived from, caused by, or sequentially following the initial event as the target input. This allows subsequent processing to obtain data containing the evolutionary chain, which helps to overcome the limitations of vision caused by isolated event information. It also allows the subsequent impact or related content of the event to be incorporated into the generated content, thereby improving the richness of the generated content.

[0066] In the embodiments of this application, based on the above technical solution, in the process of filtering and combining candidate information matching the target topic from the candidate information set according to the relevance between the target topic and the candidate information in the candidate information set to obtain the content data of the content generation request, the server first obtains the first event node corresponding to the target topic in the event knowledge graph. If the first event node contains a predetermined information sequence set for the target topic, the server obtains the candidate information corresponding to the predetermined information sequence from the candidate information set and combines them to obtain the first content chapter. The candidate information in the predetermined information sequence is selected based on its relevance to the target topic. Subsequently, the server fills the first content chapter into the corresponding position in the predetermined content structure to obtain the content data of the content generation request. In this embodiment, the server determines the core chapter based on the event context information provided in the event knowledge graph, which serves as the first content chapter in the generated content. Specifically, the server can locate the first event node in the event knowledge graph through natural language processing and entity alignment technology. In this embodiment, the first event node may contain a predetermined information sequence related to the event. The candidate information in the predetermined information sequence is selected based on its relevance to the target topic. Specifically, this involves pre-defined "event timeline" data within the event. This data can be high-confidence data generated by algorithms and manually verified, containing information identifiers for a series of candidate information representing the core timeline of the event's development. The predetermined information sequence can contain a set of information slots with a clear chronological order and specific semantic references. Each slot contains one or more candidate information items. The information slot set is a one-dimensional logical sequence pre-defined based on the narrative logic of historically selected information or expert experience, targeting a specific theme or event type. For example, for the theme of "traffic accident," the corresponding predetermined information sequence might be "time and location of occurrence -> vehicles and personnel involved -> accident details -> liability determination and handling." In actual execution, the server first checks whether the current first event node is configured with such a predetermined information sequence. If it exists, the server can directly obtain candidate information related to the core content of the current event or target theme from the predetermined information sequence. The server can smoothly transition and splice the candidate information in each slot of the sequence according to the preset chronological order, thereby combining them into an information block with complete local logic, serving as the first content chapter of the generated content. Finally, the server populates the corresponding position of the first content chapter in the predefined content structure, obtaining the content data from the content generation request. Here, the predefined content structure is a global layout or document formatting framework.This structure can be represented as a formatted file with placeholders, such as a Word document template with bookmarks, an XML / JSON structure file with a specific tag hierarchy, or a webpage frame containing standard HTML tags. A complete predefined content structure typically includes multiple parts, such as a "title area," "summary area," "core details area," and "disclaimer area." After assembling the first content chapter, the server embeds the data object of that chapter into a specific placeholder area reserved in the predefined content structure according to the mapping rules of the predefined content structure. It can be understood that in some embodiments, the server can directly use the result after only filling in the first content chapter as the generated result without further supplementing the associated information. If supplementary information is needed, or if there are other second or third content chapters generated by downstream event nodes, the system will also fill them into their respective content blocks according to the same logic. In an example embodiment, suppose a user inputs a content generation request to the system, requesting the writing of a safety topic on "X brand new energy vehicle battery recall." The system locates the first event node "X brand battery recall event" in the event knowledge graph and detects that the node contains a predefined information sequence. The sequence order is as follows: "Defect Description -> Safety Hazard Assessment -> Recall Scope and Batch -> User Guidance and Suggestions". Subsequently, the system extracts candidate information from each slot in the sequence, including "Cell Separator Process Defect", "Risk of Thermal Runaway under Extreme Weather", "30,000 Vehicles in Specific Batches Produced in 202X", and "Recommendation to Immediately Stop Fast Charging and Contact Stores". The system merges this information according to the sequence order 1 to 4 above, generating a "Recall Details Explanation" section as the core first content section. Finally, the system calls a corresponding template for a safety-related topic, fills this "Recall Details Explanation" section into the "Core Text Area" of the template, and simultaneously fills in the extraction time and title. A fixed disclaimer is retained at the bottom, thus ultimately outputting content data that can be directly used for external release. Through this method, the solution utilizes a pre-defined information sequence based on prior expert logic to construct the first content article, which helps avoid the omission of key elements or logical confusion, ensuring the accuracy and logical coherence of the output content.

[0067] In the embodiments of this application, based on the above technical solution, before filling the first content chapter into the corresponding position in the predetermined content structure to obtain the content data of the content generation request, if the first event node does not contain the predetermined information sequence, a first prompt word is generated according to the description information of the target topic, the summary information of the candidate information in the candidate information set, and the predetermined prompt information template. The first prompt word is used to guide the large language model to generate the topic information sequence. Subsequently, the server inputs the first prompt word into the large language model so that the large language model can filter and sort the candidate information according to the relevance of the candidate information to the target topic to obtain the topic information sequence output by the language model. Then, the server obtains the candidate information corresponding to the topic information sequence from the candidate information set and combines them to obtain the first content chapter. In this embodiment, in the absence of a preset predetermined information sequence, the server can construct a core information list based on the collected candidate information. Specifically, the server performs lightweight dimensionality reduction processing on each candidate text in the candidate information set and extracts its "summary information". This extraction process can be implemented by natural language processing algorithms or by using a lightweight local small model to generate the core idea. Each candidate message is compressed into a core summary and assigned a unique identifier. The server then invokes a predefined prompt template, a meta-prompt template, to integrate these three core elements. Specifically, this prompt template typically includes a task role setting (e.g., "You are a professional content editor"), task instructions (e.g., "Based on the given topic, select the most relevant items from the following summary list and arrange them in a logical order"), target topic constraints (i.e., injecting descriptive information about the target topic), and an input data area (i.e., summaries of the extracted candidate messages). Through this structured assembly, the system generates a clearly defined first prompt. The fundamental purpose of this first prompt is to guide the large language model to act as a scheduling center, dynamically generating a logical skeleton, i.e., a "topic information sequence." Next, after receiving the first prompt, the large language model continues to evaluate the numerous input summary messages. For example, the model can first perform "noise filtering," that is, determine whether each summary message matches the target topic's descriptive information in a deep semantic sense, eliminating marginal information or irrelevant noise that was mistakenly recalled during the retrieval stage. After preserving highly relevant summaries, the large language model further performs "logical ordering," automatically identifying the inherent potential connections between these pieces of information, such as the chronological order of events, causal chains, a general-to-specific argumentative structure, or an analytical path from phenomenon to essence. The model determines the order in which these relationships appear in the article. Finally, the large language model outputs a sequence of topic information. This sequence may be a simple JSON array or list, containing the identifiers (IDs) of the selected summaries or short topic phrases in a pre-sorted order.After obtaining this dynamic logical skeleton output by the large model, the server performs a reverse lookup of entity content. Based on the identifier mapping relationships in the sorted sequence of the large model, the server returns to the candidate information set and retrieves the corresponding complete candidate information. The server can then combine these complete candidate information pieces in the order indicated by the topic information sequence to form the first content chapter. Please refer to [link / reference]. Figure 5 , Figure 5 This is a schematic flowchart illustrating the content construction process in an embodiment of this application. Figure 5As shown, in step 501, the server begins constructing topical content based on the filtered articles. The construction process mainly consists of two parts: hotspot module construction and extended module construction. The hotspot module is the module containing the core content of the event, equivalent to the first content chapter. First, in step 502, the server determines whether the current node contains a predetermined information sequence. If it does, the server directly uses the predetermined information sequence in step 504, outputting the core article list 505 based on the information in the sequence. If it doesn't contain it, the server assembles prompt words in step 503, instructing the large model to filter and sort the articles input in step 501, generating the core article list, and thus outputting the core article list 505. Finally, in step 509, the server further checks whether the user requests the generation of other modules, i.e., whether in-depth analysis is required. If not, structured data, i.e., the topic containing the first content chapter, can be directly output based on the core article list 505. If in-depth analysis is required, the server removes articles already selected from the core article list from the input article set in step 506. Then, it calculates the semantic similarity between the remaining articles and the core content. Based on this semantic similarity, articles with high content overlap with the core article list are removed. For the remaining articles, the server inputs the cleaned remaining articles into the large model in step 507. Prompt words guide the model to plan sub-topics, generate sub-topics, and classify the corresponding articles. Based on the sub-topics and article classification results, the server constructs structured chapter data, i.e., the second content chapter. Subsequently, the server outputs structured thematic data based on the first and second content chapters. To illustrate with a specific example, using the "X brand new energy vehicle battery recall" scenario from the previous embodiment, suppose the user also requests a chapter in the report about "the social media public relations crisis and consumer rights protection controversy triggered by the battery recall." The "public relations crisis and consumer rights protection controversy" are divergent secondary events, and the "X brand battery recall crisis" node in the system graph does not have a predetermined information sequence attached. At this point, the server first compresses the more than 100 social media posts, forum articles, and consumer rights protection statements collected about this rights protection incident, extracting more than 100 short "summary information". Then, the server fills the user's intent "analyze the social media public relations crisis and consumer rights protection incident triggered by the recall" and these 100 summaries into a preset prompt word framework, generates the first prompt word, and inputs it into the large language model. After receiving the first prompt word, the large language model filters out irrelevant summaries such as spam posts, and arranges the selected real events according to dynamic logic and time sequence, outputting a topic information sequence, which contains candidate information arranged in logical order and time sequence. For example, the topic information sequence contains four candidate information, in the following order: [the first car owner releases a rights protection video] -> [the article that caused the incident to ferment on the short video platform and become a hot search] -> [the manufacturer's early strong response] -> [the news of the consumer association's intervention and the manufacturer's apology].Subsequently, the server fills the dynamically generated chapter into the corresponding position in the report's predetermined content structure. In this way, when a graph node lacks a manually preset logical template, the solution can extract candidate information summaries and combine them with a predetermined format to construct prompt words, guiding the large language model to generate a corresponding topic information sequence. The model determines the order, and the server assembles the original text. This eliminates the need to directly input large amounts of original text as context, avoiding the illusions that can easily occur when large models directly generate long texts, leading to inaccurate information. This improves the accuracy and reliability of the output chapters. Furthermore, the absence of a pre-set information sequence allows for faster response to unexpected events and improved information transmission efficiency.

[0068] In the embodiments of this application, based on the above technical solution, before filling the first content chapter into the corresponding position in the predetermined content structure to obtain the content data of the content generation request, if the content generation request contains a deep analysis identifier, the server selects the remaining candidate information whose similarity is lower than a predetermined threshold based on the similarity between other candidate information in the candidate information set and the candidate information in the first content chapter. The other candidate information refers to candidate information in the candidate information set that does not belong to the first content chapter. Subsequently, the server generates a second prompt word based on the remaining candidate information and a predetermined prompt information template. The second prompt word guides the large language model to generate multiple sub-topics based on the remaining candidate information. Next, the server inputs the second prompt word into the large language model to obtain multiple sub-topics output by the large language model and the remaining candidate information corresponding to each sub-topic. Finally, the server combines the remaining candidate information corresponding to each sub-topic to obtain multiple second content chapters, which are used to fill the corresponding positions in the predetermined content structure. In this embodiment, if the configuration requires the generation of other content chapters, the server executes incremental logic to further generate second content chapters. Specifically, the server first checks whether the content generation request contains a deep analysis identifier. In the system's interactive front-end or API interface definition, the "deep analysis identifier" can be represented as a Boolean parameter, a specific enumeration value, or a specific prompt word entered by the user through natural language, such as the request text containing words like "deep analysis," "comprehensive assessment," or "surrounding impact." The server can parse specific prompt words into a deep analysis identifier through intent recognition. When the server detects that the current content generation task carries this identifier, it triggers a deep expansion processing branch. The server can use difference operations to remove candidate information that has already been extracted and used to combine the first content chapter, thereby obtaining other unused candidate information. This provides physical isolation at the data level, avoiding the reuse of consumed basic data and preventing homogenization of chapter content. Subsequently, the server performs divergent filtering based on low similarity. Specifically, the server uses a pre-trained semantic representation model to map the candidate information contained in the first content chapter to one or more baseline semantic vectors, while also mapping other candidate information obtained in the previous step to contrastive semantic vectors. Then, the system calculates the spatial distance between the contrastive semantic vector and the baseline semantic vector, such as cosine similarity or Euclidean distance. If the similarity of another candidate piece of information is higher than a predetermined threshold, such as a similarity greater than 0.7, it means that the information is semantically just a synonym or paraphrase of the core facts of the first content chapter, such as a press release report on the same event from another media outlet.The server discards such data as redundant noise. Conversely, only when the similarity is below a predetermined threshold is it considered that the candidate information, although belonging to the same candidate set of the "target topic," has sufficiently significant differences and orthogonality from the core candidate information, and thus belongs to extended information with incremental value. The server retains this low-similarity data, defining it as residual candidate information. For the retained residual candidate information, the server uses it as contextual cues, along with specific instruction templates, and inputs it into the large model. For example, the instruction could be: "Please read the following peripheral information about the target topic, identify the multiple implicit discussion directions, extract a sub-topic for each direction, and finally classify and summarize the original information under the corresponding sub-topic." The large language model outputs a structured classification result, that is, it sorts out multiple general sub-topics and divides the residual candidate information summarized by each sub-topic. After obtaining the structured output of the large model, the server combines the residual candidate information under each divided sub-topic, such as adjusting the word order, transitioning words, and merging paragraphs, to obtain the combined output result, which serves as the second content chapter. In the final layout and assembly stage, the system will fill these second content chapters into the corresponding positions in the pre-reserved content structure, such as the "In-Depth Analysis Area," "Surrounding Impact Area," or "Market Feedback Area," thereby completing the assembly of the entire content data. Continuing with the example implementation of the battery recall described above, assume the user selected the "Generate In-Depth Industry Analysis Report" option, triggering the in-depth analysis flag. The server has already generated the first content chapter containing core facts such as "cell defects, 30,000 batches, and a recommendation to stop charging." At this point, the server retrieves the remaining candidate information from the candidate article library. When calculating similarity, if the server finds press releases that repeatedly broadcast recall notices, it will remove them. The remaining candidate information retained by the server includes: "Battery supplier Y's stock price plummeted 8% that day," "Competitor Z brand launched a 'lifetime battery warranty' promotional poster overnight," "Many users in car owner groups expressed anger that they couldn't get through to X brand's after-sales service hotline," and "An industry expert pointed out that this type of separator technology is ten years old." Subsequently, the server inputs this information into a large language model, which summarizes it into three sub-themes: supply chain and capital market turmoil, competitor countermeasures and leveraging, and public relations crisis and declining user trust. Based on these, the server generates three second content chapters. The server then sequentially fills these three second content chapters into the "In-Depth Impact Analysis" section of the report template. In this way, the server uses a low similarity threshold to filter other candidate information, removing homogeneous data that highly overlaps with the first content chapter. This helps reduce information redundancy among the generated content modules, minimizing information redundancy and improving the end-user experience.

[0069] In the embodiments of this application, based on the above technical solution, the server further obtains the client's information editing records for the generated content data. These information editing records contain modification operations performed on candidate information included in the content data. Subsequently, the server adds constraint information based on the modification operations in the information editing records to the prompt information template. This constraint information is used to instruct the large language model. In this embodiment, the server can modify the prompt information template based on the user's editing operations on the generated content data, thereby adjusting the prompt word content input into the large language model to make the model's output more accurate. Specifically, the server has already pushed the generated content data to the client's content draft interface for the user to adjust and confirm. When the user completes editing and clicks "confirm and publish," the server can record the differences between the finally confirmed content data and the original generated content data, or the modifications performed, such as deleting information from a specific source or modifying certain words. The server will correlate these modifications with candidate information to obtain labeled data with clear contextual information. After obtaining a large number of modification operations, the server can summarize constraint information with universal guiding significance from the discrete modification operations. The constraint information can contain multiple dimensions. For example, constraint information can include a vocabulary dimension. If the information editing record contains numerous modifications replacing common words (e.g., brakes) with technical terms (e.g., brake system), the server can automatically construct a mapping relationship to generate constraint information such as "When processing hardware appearance descriptions, force the replacement of common words with specific technical terms." Constraint information can also include a writing style dimension. If the information editing record contains numerous deletions of adjectives, exclamation marks, or conjunctions with strong subjective speculation, the server can generate writing style constraint rules, such as "Please maintain an objective and neutral tone; the use of exaggerated modifiers or subjective emotional evaluations is strictly prohibited." Constraint information can also include a logical structure dimension. For example, if the information editing record contains numerous adjustments to the subtopic hierarchy, such as moving paragraphs that were originally placed at the beginning to the end, the server can generate constraint information such as "Adjust the information weight order, placing specific discussions later." Subsequently, the server can append these constraint information to predefined prompt information templates. It can be understood that the aforementioned generated constraint information can be appended to both the predefined prompt information templates used to generate the first prompt word and the predefined prompt word templates used to generate the second prompt word. In this way, when the server receives a similar target topic again, the first or second prompt word it assembles will carry the previously corrected constraints, ensuring that the result generated by the large language model naturally meets the constraints without requiring additional modification. Through this method, the solution modifies the prompt message template based on editing operations on the generated content data, thus forming a closed-loop negative feedback mechanism based on user actions. This helps reduce the recurrence of similar errors and lowers the resource consumption of manually proofreading the generated content data.

[0070] In the embodiments of this application, based on the above technical solution, in the process of merging the first candidate information with the second candidate information contained in the target event node to obtain a candidate information set, the server first obtains the second candidate information corresponding to each target event node. Then, the server calculates the union between the first candidate information and the second candidate information according to the information identifiers of the first candidate information and the second candidate information, obtaining a union set. Next, the server uses a large language model to score the relevance between the description information of the target topic and the summary information of each candidate information in the union set, obtaining a relevance score corresponding to each candidate information. Finally, the server deletes candidate information whose corresponding relevance score is lower than a predetermined score threshold from the union set according to the relevance score corresponding to each candidate information, obtaining a candidate information set. In this embodiment, the server filters the candidate information during the merging of the first and second candidate information to avoid low-quality candidate information affecting the processing results of the large language model. Specifically, in the event knowledge graph, each event node carries multiple candidate information. Target event nodes are event nodes associated with the target topic, and they carry multiple second candidate information. The server calculates the union between the first and second candidate information. The server assigns a globally unique label to each individual candidate message. This label can be the Uniform Resource Locator (URL) of the original news article, a document ID in the underlying document library, or a unique data fingerprint generated by the system using a hash calculation of the core text of the information using an encryption algorithm. The server can build a global hash set in memory, traverse it, and extract the information labels of all candidate messages as primary keys for comparison. If the information label of a second candidate message already exists in the set, it means that the candidate message has already been recalled during the media retrieval stage. The server can discard the redundant candidate message or simply merge the labels. Only when a completely new information label is encountered is it formally inserted into the set. This can eliminate completely overlapping parts in the two recall results. After physical deduplication, the server uses a large model to filter out noisy data that consists only of keywords but differs significantly from the target topic in its core meaning. The server can obtain the summary information of each candidate message in the union set by using an extractive summarization algorithm or by extracting the summary field of the original data. Subsequently, the server constructs quantitative evaluation prompts, which require the large language model to score the relevance of the candidate message summary information based on the descriptive information of the target topic. After receiving the prompt words, the large language model performs semantic alignment and inference on the two, ultimately outputting a relevance score. After obtaining the quantified scores of all candidate information, the server loads a predetermined scoring threshold set in the system. This threshold can be a static absolute score, such as a fixed passing score of 60, or a dynamic score calculated based on the overall score distribution of the current union set, such as the median of all scores or the top-70% percentile.The server performs a fast traversal of the union set, identifying candidate information with a relevance score below a certain threshold as low-quality noise or off-topic information, and removing them from the union set. In this way, the solution extracts a summary of the information in the deduplicated set, inputs it into a large model for relevance scoring, and thus filters out low-relevance, off-topic noise, preventing noise from affecting subsequent screening and generation processes. This helps improve the accuracy of the output content data relative to the target topic.

[0071] In the embodiments of this application, based on the above technical solution, in the process of deleting candidate information with a corresponding relevance score lower than a predetermined score threshold from the union set according to the relevance score corresponding to each candidate information, and obtaining a candidate information set, the server can delete candidate information with a corresponding relevance score lower than a predetermined score threshold from the union set according to the relevance score corresponding to each candidate information, and obtain a filtering result set. Then, the server clusters the candidate information in the filtering result set according to the semantics of each candidate information, and obtains multiple information clusters. Next, for each information cluster, the server selects candidate information whose quality score meets a predetermined score condition according to the quality score of each candidate information in the information cluster, and obtains the information subset corresponding to each information cluster. Finally, the server merges the information subsets corresponding to each information cluster to obtain a candidate information set. In this embodiment, the server uses clustering to filter out candidate information with low relevance, thereby avoiding excessive information duplication and homogenization of topic content. Specifically, the server uses the relevance score to filter out candidate information with low relevance to the target topic in the manner described in the above embodiment, and obtains a filtering result set. Subsequently, the server uses a clustering algorithm to divide the semantically similar candidate information into different information clusters. Specifically, the server can use a pre-trained feature extraction model to map each candidate message (or its summary) in the filtered result set into a high-dimensional semantic vector. Then, it calculates the distance metric between the candidate message vectors in the vector space, such as cosine distance or Euclidean distance. The server can use a clustering algorithm to group vectors with similar distances into clusters. Each cluster is an information cluster, representing a specific sub-direction or sub-topic under the target theme. Within each information cluster, the server filters candidate messages based on a quality score. The score can be calculated based on parameters such as click-through rate, timeliness, content richness, source authority, and text standardization. Content richness can be assessed by calculating text length and entity density to evaluate the information entropy. Text standardization can be determined by using a language model to detect grammatical errors or garbled characters in the text. The server weights and sums the scores from the above dimensions to obtain the final quality score for the candidate message. The server then filters according to predetermined score conditions. These predetermined score conditions can be selecting the top K candidate messages within the cluster based on their quality scores, or selecting all candidate messages with quality scores higher than a certain absolute quality threshold, etc. The server merges the subsets of information generated from each information cluster to obtain a candidate information set. Please refer to [link / reference]. Figure 6 , Figure 6 This is a schematic diagram illustrating the clustering process used to filter candidate information in an embodiment of this application. For example... Figure 6As shown, the server calculates the similarity between candidate information in the union set 610 and the descriptive information of the target topic, and deletes candidate information with low similarity, forming a filtered result set 620. Subsequently, the server performs clustering based on the filtered result set 620, obtaining information clusters 630. Typically, the number or range of information clusters is predetermined. For each information cluster 630, the server sorts the candidate information according to its quantity score, selecting the top K, resulting in an information subset 640 for each information cluster 630. Finally, the server merges the information subsets 640 to obtain a candidate information set 650. In this way, the scheme divides candidate information into different sub-topics through clustering, and selects the best within each cluster based on quality scores. This ensures that the selected information has sufficient discriminative power, helps avoid the excessive influence of a single popular piece of information on the generated content data, and improves the data quality of the generated content data.

[0072] In the embodiments of this application, based on the above technical solution, the server can further extract the meta-information of each candidate information in the union set. The meta-information includes at least one of information source, quality score, and information status. Subsequently, the server deletes candidate information that does not meet predetermined filtering conditions from the union set according to the meta-information, wherein the predetermined filtering conditions include constraints set for the meta-information. In this embodiment, the server filters candidate information based on predetermined hard rules, thereby quickly cleaning up candidate content. Specifically, the information source is used to record which publishing entity initially generated the candidate information, generally divided into official messages and personal or self-media messages. The quality score refers to the prior static score given by the basic rule engine to the candidate information during the initial storage or indexing stage. This score can be calculated based on basic features, such as the total number of words in the text, the completeness of the paragraph structure, whether it contains a valid data table, and the proportion of garbled characters or special rare characters in the text. The information status is used to measure the timeliness and validity of the information. For example, its enumerated values ​​can include "published and valid", "expired", "retracted", and "debunked". Metadata can also include information tailored to different information formats. For example, images and videos can include clarity and resolution, while audio can include audio length and clarity. After extracting the metadata for each candidate piece of information, the server loads a series of pre-configured business constraints from the rules engine and performs rapid traversal and judgment on the union set. Pre-defined filtering conditions can be configured according to specific application scenarios. For example, constraints on information sources can use whitelists or blacklists. Constraints on quality scores can set basic numerical thresholds. Constraints on information status can require candidate information to be in a valid state. Constraints can be configured by the user on the client side. For example, a user can set the client side to only use information from official sources, which can then be restricted in the information source constraints. In this way, the solution can directly intercept low-quality data using pre-defined hard constraints, thereby avoiding low-quality data input into large language models that could cause model illusions, thus improving the compliance and credibility of candidate information.

[0073] In the embodiments of this application, based on the above technical solution, during the process of obtaining a content generation request for a target topic, the server can receive a generation instruction sent by the client. The generation instruction includes setting a topic name, topic description, and system parameters for the target topic. The system parameters are used to filter candidate information for the target topic according to predetermined filtering conditions. Then, the server concatenates the topic name and topic description to form the description information of the target topic. Subsequently, the server generates a content generation request for the target topic based on the description information and the system parameters. In this embodiment, the server sets the filtering rules for candidate information according to the system parameters input by the user. Specifically, the user completes form filling or configuration selection on the user client's interactive interface, thereby triggering the generation instruction. The generation instruction mainly includes a topic name, topic description, and system parameters. The system parameters may include, for example, a generation mode and a source configuration. The generation mode is used to distinguish whether in-depth analysis is required, while the source configuration is used to constrain whether candidate information is limited to official information. The server can predefine various concatenation templates, such as: "[Topic]: {Topic Name}; [Detailed Background]: {Topic Description}". The combined text is generated by filling the extracted topic name and description into placeholders in the template. Furthermore, during the concatenation process, the server can perform text preprocessing operations, such as removing redundant spaces, standardizing full-width and half-width characters, and filtering out non-compliant special symbols. Finally, the server generates a content generation request for the target topic based on the description information and system parameters. When this content generation request flows through the service bus, downstream processing nodes (such as the recall module, sorting module, and fusion module) can directly read the corresponding system parameters from the request object. For example, during the filtering and screening stages, the server can directly apply these predefined filtering conditions to filter the first candidate information in the initial recall and the second candidate information in the graph, quickly eliminating noisy data that does not meet the conditions. In this way, the solution concatenates the name and description on the server side, maximizing the preservation of the complete semantic features of the input, and independently encapsulates the system parameters as predefined filtering conditions, providing clear hard indicators for retrieval and screening. This helps to increase user participation in the content generation process and improve the matching between the generated results and user intent.

[0074] In the embodiments of this application, based on the above technical solution, after filtering and combining candidate information matching the target topic from the candidate information set according to the correlation between the target topic and the candidate information in the candidate information set to obtain the content data of the content generation request, the server can also display a content draft interface through the client. The content draft interface includes a first component, a second component, and a third component. The first component displays the names of each content chapter in the content data, the second component displays the information preview interface corresponding to the content chapter selected in the first component, and the third component displays the editing interface of the content chapter selected in the first component. Subsequently, the server updates the content data according to the information editing instruction sent by the client and sends it to the client to display the editing result in the third component, wherein the information editing instruction is triggered by the information editing operation in the third component. Finally, the server generates topic information containing the content data according to the confirmation instruction sent by the client and sends it to the client to display the information display page of the target topic in the client. In this embodiment, the content data generated by the server is a draft, and the server displays the draft to the user in the client, whereby the user can modify and confirm the draft. Specifically, the server sends the generated content data to the client, which then displays a content draft interface to the user. The content draft interface contains a first component, a second component, and a third component. The position and display format of each component on the page typically correspond to the predetermined content structure adopted by the content data. The first component displays the names of each content chapter in the content data. The content data usually contains at least one first chapter and multiple second chapters. The chapter names can be displayed in the first component as a list and links. Users can browse the specific content of each chapter on the page by clicking on the chapter names in the first component. The second component is a preview interface for the selected content chapter in the first component. Each content chapter can contain one or more candidate information items. The second component can directly display the content of these candidate information items, or it can only display the titles and access links of each candidate information item. The third component displays the editing interface for the selected content chapter in the first component. Each content chapter has corresponding chapter metadata, which is typically used to constrain the content and format of the content chapter displayed on the interface, such as chapter summaries, the resolution of images in the chapter, and chapter preview images. Users can edit this information using controls in the third component. After a user finishes editing on the client, their edits trigger an editing command and are sent to the server. The server then adjusts the content data according to the edits and sends the result back to the client, which then displays the edited result. It is understood that in some embodiments, the client can choose whether to send the edited content to the server.For example, text editing can be completed and displayed directly on the client side, while changes to information access links are sent to the server to update the content data. Information editing operations include various interactive commands issued by the user via input peripherals such as keyboard, mouse, and microphone, including but not limited to: adding, deleting, and rewriting characters; adjusting text formatting (such as bolding, highlighting, and adjusting line spacing); reordering paragraphs; pasting content from the external clipboard; and inserting local charts, images, videos, audio, or data tables. In some embodiments, the server or client can serialize and store the edited temporary results locally at a predetermined frequency (such as every 5 seconds or 2 seconds after the user stops input), triggering an automatic save request after debouncing, and asynchronously synchronizing the latest draft fragment to the server's draft table database. After editing, the user confirms the content draft on the content draft interface. Confirmation typically refers to the user triggering the "Submit," "Publish," "Generate," or "Export" control buttons on the interface. After receiving the confirmation command from the client, the server can update the content data from the draft state to the published state and further provide downstream notification services or cache preheating services. After receiving a "published successfully" confirmation response from the backend, the client can navigate to or refresh the page, close the current draft content interface, and render the information display page. This display page is typically a purely reading page with all editing controls removed and optimized layout, or a preview page in PDF / HTML format for users to download. Please refer to [link / reference]. Figure 7 , Figure 7 This is a schematic flowchart illustrating the prompt word correction process in an embodiment of this application. Figure 7As shown, prompt templates 701 and 702 are used to generate the first prompt word 703 and the second prompt word 704, respectively. The first prompt word 703, along with candidate information, is input into the large language model 705 for filtering and sorting to obtain a topic information sequence, thereby generating the first content chapter 705. The second prompt word 704, along with the remaining candidate information, is input into the large language model 705 for sub-topic classification to obtain multiple sub-topics and corresponding articles, thereby generating the second content chapter 707. The first content chapter 705 and the second content chapter 707 are combined to form content data 708. Content data 708 is displayed to the user, and when the user modifies content data 708, a user modification operation 709 is triggered. After modification, the user confirms the content data, thereby generating published content 710. The content modified by the user is uploaded to the server, forming a content modification record 711. Based on the content modification record 711, the server generates correction information for the prompt words, thereby correcting prompt templates 701 and 702. In this way, the server corrects the content data according to the user's editing quality through the interactive process, avoiding mismatch between the generated content and the actual needs. This helps to improve the reliability and flexibility of the content produced by the solution and enhance the overall user interaction experience.

[0075] In the embodiments of this application, based on the above technical solution, the description information includes the topic name and topic description information of the target topic. During the process of performing media retrieval based on the description information of the target topic to obtain first candidate information, the server can extract keywords from the topic name and the topic description information to obtain topic keywords. Then, based on the description information of the target topic, media retrieval is performed to obtain the first candidate information. Based on the topic keywords, first recall information associated with the topic keywords is retrieved. Furthermore, the server concatenates and embeds the topic name and the topic description information to obtain an embedding vector corresponding to the target topic. Next, based on the embedding vector, the server retrieves similar information vectors similar to the embedding vector in a vector database to obtain second recall information corresponding to the similar information vectors. The vector database contains multiple information vectors and corresponding event information. Finally, the server merges the first recall information and the second recall information as the first candidate information. In this embodiment, the server recalls the first candidate information through keyword retrieval and vector retrieval. Specifically, the server can use natural language processing algorithms to perform word segmentation, part-of-speech tagging, and stop word filtering on the topic name and topic description information of the target topic. Subsequently, the server extracts a core vocabulary set from the cleaned text using keyword extraction algorithms or named entity recognition methods based on pre-trained models. In practical applications, the server can also assign different initial search weights to these keywords based on part-of-speech or entity type to improve search accuracy. After obtaining the topic keywords, the server performs an inverted index query using the underlying search engine. The server constructs a Boolean query statement from the extracted topic keywords and matches it using a text similarity scoring algorithm to obtain information content containing the keywords. The server can calculate a relevance score and extract the top-ranked (e.g., the top 100) documents or multimedia data as the primary recall information. The server also concatenates the topic name and topic description information into a complete continuous text. Subsequently, the server inputs this complete text into a pre-trained deep language encoding model. This encoding model abstracts the contextual dependencies and semantic features in the text and outputs the text's embedding vector. After obtaining the embedding vector representing the target topic intent, the system initiates a query request to a backend professional vector database (such as Milvus, Faiss, or Pinecone). The vector database pre-stores all historical financial information, event slices, and research report paragraphs captured by the system offline, and all of this massive amount of data has been transformed into information vectors of the same dimension.During the retrieval process, the system utilizes Approximate Nearest Neighbor (ANN) algorithms (such as HNSW or IVF-PQ) to calculate the spatial distance between the target topic embedding vector and each information vector in the database (common metrics include cosine similarity, inner product, or Euclidean distance L2 distance). Vectors with closer distances or higher similarity scores indicate a stronger semantic fit between their corresponding original event information and the target topic. The system extracts the top M records with the highest similarity scores as the second recall information. After completing keyword-based literal recall and vector database-based semantic recall, the server merges the result sets generated by these two approaches. Since the result sets of the two approaches may overlap, the server performs deduplication, using the unique identifier of each piece of information to remove duplicates. In some embodiments, the server can also use algorithms such as inverse sorting fusion to perform preliminary scoring and reordering of the merged results, giving higher priority to information that ranks higher in both approaches. The data set after merging and deduplication is used as the first candidate information. In this way, the solution combines keyword recall and vector recall to ensure that the quantity and coverage of recalled information cover the core content of the target topic and improve the coverage rate of candidate information.

[0076] The complete solution of this application is described below with specific examples. Please refer to [link / reference]. Figure 8 , Figure 8 This is a schematic diagram illustrating the overall process of generating content data in an embodiment of this application. For example... Figure 8As shown, the user inputs the title and brief description of the topic 801. The server performs two-way retrieval based on the input: first, a media search is conducted from the resource database 802 based on keywords and semantics to obtain first candidate information 803; second, a search is conducted using a knowledge graph 804 to obtain second candidate information 805. The two candidate information sets are merged to obtain a union set 806. The server filters the union set 806 using hard rules, cluster optimization, and a large model 807 to obtain a candidate information set 808. If the knowledge graph 804 contains a predetermined information sequence, the first chapter content 811 can be directly obtained. If the knowledge graph 804 does not contain a predetermined information sequence, the server inputs the candidate information set 808 and the first prompt word 809 into the large language model 807 for filtering and sorting to obtain a topic information sequence 810, thereby generating the first chapter content 811. Unused information in the candidate information set 808 is considered as remaining candidate information 812. If necessary, the server further uses the large language model 807 to perform sub-topic reasoning based on the second prompt word 814 and the remaining candidate information 812, obtaining multiple word topics 814 and the remaining candidate articles corresponding to each sub-topic, thus combining them to obtain the second chapter content 815. The server merges the first chapter content 811 with the second chapter content 815 to obtain the output content data 816.

[0077] This example illustrates how a news platform can generate news features based on trending news. Suppose a company issues a recall notice for its X-brand electric vehicles, triggering numerous media reports. The news platform wants to publish a news feature on this event. Editors can input the news feature title and core event description on the client's interactive page, and set options such as in-depth research and excluding self-media from the information source, generating corresponding feature content. The client sends a content generation request to the backend server based on the user-defined form data. Upon receiving the content generation request, the backend server performs natural language segmentation on the user-input news feature title and core event description, deriving keywords such as "battery" and "recall." Based on these keywords, it retrieves information directly mentioning the event, serving as the first recall information. The server can also map the information combining the news feature title and core event description into embedded vectors and perform spatial distance similarity comparisons in a vector database, obtaining information related to vehicle recalls but not directly mentioning "X-brand recall," serving as the second recall information. The first and second recall information are combined as the first candidate information.

[0078] The server also searches for the first event node corresponding to the event in the automotive event knowledge graph, such as "Brand X vehicle issues a large-scale recall announcement due to power battery defects." Subsequently, the server traverses the downstream nodes of the first event node in the automotive event knowledge graph according to their hierarchical relationship, identifying them as second event nodes. Examples include nodes such as "significant intraday fluctuations in the stock price of the supporting power battery supplier," "a short-term surge in user inquiries and complaints about models of the same brand," "the regulatory department launches a special investigation into power battery safety," and "the automaker conducts an upgraded safety re-inspection of similar batteries." Both the first and second event nodes will carry relevant event information, such as related news and official statements. The server retrieves the information carried by the event nodes as second candidate information.

[0079] For the candidate information set formed by combining the first and second candidate information, the server can perform multi-level filtering. Specifically, the server can extract the unique information identifiers of the first and second candidate information and directly calculate their union in memory. If a news item is found to be both retrieved by media retrieval and referenced by graph nodes, the system eliminates redundancy, ensuring that each candidate information in the merged union set is objectively independent. The server can further parse the metadata of each candidate information in this union set (such as information source, basic quality score, and current status). Based on the server parameters initially set by the user, the server directly initiates the deletion of shill posts marked as "anonymous leaks from car enthusiast groups," false rumors from unofficial channels, and invalid data that is in a "retracted" state due to information errors. The server can further extract summaries of the remaining candidate information and input them along with the target topic into the large language model. The large model scores the relevance of the candidate information to the topic. Based on the preset passing score, the server deletes off-topic noise that only mentions "X brand cars" but discusses its "new car launch plan," "brand racing marketing," and "executive personnel changes." In addition, the server performs semantic vector clustering on the filtering results, automatically dividing them into several sub-topic clusters such as "Recall Scope and Batch Details," "Analysis of the Root Causes of Battery Defects," and "Automaker Remedial Measures and User Response Solutions." Within each cluster, the server selects the top few candidate articles with the highest quality scores. Finally, the server aggregates the candidate articles selected from each cluster to form a candidate information set.

[0080] Subsequently, the server first examines the first event node in the knowledge graph. This first event node can carry an event timeline configured for the event by news editors or the system, i.e., a predetermined information sequence. This sequence contains confirmed candidate information sequences that represent the current event's development. The server can directly combine the designed candidate information from these sequences into the first content chapter. Assuming the event is highly time-sensitive, the first event node does not carry a pre-configured event timeline. The server extracts summaries of the candidate information and combines them with a preset prompt word format to form the first prompt word, which is then input into the large language model. The large model outputs a topic information sequence according to the input prompt word requirements. Similar to the predetermined information sequence, the topic information sequence is composed of multiple candidate information IDs sorted together to represent the event's development timeline. The server then combines the corresponding candidate information into the first content chapter based on the topic information sequence.

[0081] Because the user selected deep analysis, the server detects the deep analysis identifier and then calculates the similarity between other unused candidate information from the candidate information set and the information in the first content section. The server selects candidate information with low similarity to the first content section as remaining candidate information, such as "case studies of similar past events," "the latest industry standards for power batteries," and "the impact of recall events on used cars." Using a pre-defined prompt template, the server assembles these divergent candidate information into second prompt words and inputs them into a large language model. The large model refines and summarizes these candidate information, outputting multiple sub-topics and candidate information under each sub-topic, such as "past cases," "industry standards," and "event impact." The server combines the candidate information according to classification to generate multiple second content sections, with one second content section corresponding to one sub-topic. Finally, the first content section and these second content sections are embedded into the overall webpage framework of the news topic.

[0082] After the background generation is complete, the server pushes the generated draft to the client, where it is displayed in the content draft interface. Users can preview the published presentation and make modifications in the content draft interface. User-modified content can be uploaded to the server as a modification record. The server can periodically modify the prompt template based on the modifications recorded in the modification record, thus making the server-generated draft more aligned with user expectations. For example, if users frequently delete information from specific sources or change emotional expressions and subjective evaluations in the draft, the server can add prompts such as "Objective statements are required" and "Use of information from XX source is prohibited" to the prompt template.

[0083] It should be noted that although the steps of the method in this application are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0084] The following describes the implementation of the apparatus of this application, which can be used to perform the data processing method in the above embodiments of this application. Figure 9 A schematic block diagram illustrating the composition of a data processing apparatus in an embodiment of this application is shown. The data processing apparatus is applied to a data processing system, such as... Figure 9 As shown, the data processing device 900 mainly includes: The content acquisition module 910 is configured to acquire a content generation request for a target topic, wherein the content generation request includes descriptive information of the target topic; The media retrieval module 920 is configured to perform media retrieval based on the description information of the target topic to obtain first candidate information; The graph query module 930 is configured to query target event nodes related to the target topic from a preset event knowledge graph. The event knowledge graph contains multiple event nodes and event association edges. The event nodes are used to represent events and contain second candidate information related to the events. The event association edges are used to represent the association relationship between events. The information merging module 940 is configured to merge the first candidate information with the second candidate information contained in the target event node to obtain a candidate information set; The information filtering module 950 is configured to filter and combine candidate information matching the target topic from the candidate information set based on the degree of correlation between the target topic and the candidate information in the candidate information set, so as to obtain the content data of the content generation request.

[0085] In the embodiments of this application, based on the above technical solution, the graph query module 930 is specifically configured to: locate a first event node corresponding to the target topic in the preset event knowledge graph according to the description information of the target topic; traverse the event nodes located downstream of the first event node in the event knowledge graph according to the event association edge with the first event node to obtain a second event node; and use the first event node and the second event node together as target event nodes related to the target topic.

[0086] In the embodiments of this application, based on the above technical solution, the information filtering module 950 is specifically configured to: obtain a first event node corresponding to the target topic in the event knowledge graph; if the first event node contains a predetermined information sequence set for the target topic, obtain candidate information corresponding to the predetermined information sequence from the candidate information set, and combine them to obtain a first content chapter, wherein the candidate information in the predetermined information sequence is selected according to the relevance between it and the target topic; fill the first content chapter into the corresponding position in the predetermined content structure to obtain the content data of the content generation request.

[0087] In the embodiments of this application, based on the above technical solution, the information filtering module 950 is further configured to: if the first event node does not contain the predetermined information sequence, generate a first prompt word based on the description information of the target topic, the summary information of the candidate information in the candidate information set, and the predetermined prompt information template, wherein the first prompt word is used to guide the large language model to generate a topic information sequence; input the first prompt word into the large language model so that the large language model can filter and sort the candidate information according to the relevance of the candidate information to the target topic to obtain the topic information sequence output by the language model; obtain the candidate information corresponding to the topic information sequence from the candidate information set and combine them to obtain a first content chapter.

[0088] In the embodiments of this application, based on the above technical solution, the information filtering module 950 is further configured to: if the content generation request contains a deep analysis identifier, then select the remaining candidate information whose similarity is lower than a predetermined threshold according to the similarity between other candidate information in the candidate information set and candidate information in the first content chapter, wherein the other candidate information is candidate information in the candidate information set that does not belong to the first content chapter; generate a second prompt word according to the remaining candidate information and a predetermined prompt information template, the second prompt word being used to guide the large language model to generate multiple sub-topics according to the remaining candidate information; input the second prompt word into the large language model to obtain multiple sub-topics output by the large language model and the remaining candidate information corresponding to each sub-topic; combine the remaining candidate information corresponding to each sub-topic to obtain multiple second content chapters, the second content chapters being used to fill the corresponding positions in the predetermined content structure.

[0089] In the embodiments of this application, based on the above technical solution, the information filtering module 950 is further configured to: obtain information editing records of the client for the generated content data, the information editing records containing modification operations on candidate information contained in the content data; and add constraint information based on the modification operations in the information editing records to the prompt information template, the constraint information being used to indicate the large language model.

[0090] In the embodiments of this application, based on the above technical solution, the information merging module 940 is specifically configured to: obtain second candidate information corresponding to each of the target event nodes; calculate the union between the first candidate information and the second candidate information according to the information identifiers of the first candidate information and the second candidate information to obtain a union set; score the relevance between the description information of the target topic and the summary information of each candidate information in the union set using a large language model to obtain a relevance score corresponding to each candidate information; and delete candidate information whose corresponding relevance score is lower than a predetermined score threshold from the union set according to the relevance score corresponding to each candidate information to obtain a candidate information set.

[0091] In the embodiments of this application, based on the above technical solution, the information merging module 940 is specifically configured to: delete candidate information whose corresponding relevance score is lower than a predetermined score threshold from the union set according to the relevance score corresponding to each candidate information, to obtain a filtering result set; cluster the candidate information in the filtering result set according to the semantics of each candidate information, to obtain multiple information clusters; for each information cluster, select candidate information whose quality score meets a predetermined score condition according to the quality score of each candidate information in the information cluster, to obtain an information subset corresponding to each information cluster; merge the information subsets corresponding to each information cluster to obtain a candidate information set.

[0092] In the embodiments of this application, based on the above technical solution, the information merging module 940 is further configured to: extract the meta-information of each candidate information in the union set, the meta-information including at least one of information source, quality score and information status; and delete candidate information that does not meet the predetermined filtering conditions from the union set according to the meta-information, wherein the predetermined filtering conditions include constraints set for the meta-information.

[0093] In the embodiments of this application, based on the above technical solution, the content acquisition module 910 is specifically configured to: receive a generation instruction sent by a client, the generation instruction including setting a topic name, topic description and system parameters for a target topic, the system parameters being used to filter candidate information of the target topic according to predetermined filtering conditions; concatenate the topic name and the topic description to form the description information of the target topic; and generate a content generation request for the target topic according to the description information and the system parameters.

[0094] In the embodiments of this application, based on the above technical solution, the information filtering module 950 is further configured to: display a content draft interface through a client, the content draft interface including a first component, a second component, and a third component, the first component displaying the names of each content chapter in the content data, the second component displaying an information preview interface corresponding to the content chapter selected in the first component, and the third component displaying an editing interface for the content chapter selected in the first component; updating the content data and sending it to the client according to the information editing instruction sent by the client, so as to display the editing result in the third component, wherein the information editing instruction is triggered by an information editing operation in the third component; generating topic information containing the content data and sending it to the client according to the confirmation instruction sent by the client, so as to display the information display page of the target topic in the client.

[0095] In the embodiments of this application, based on the above technical solution, the description information includes the topic name and topic description information of the target topic; the media retrieval module 920 is specifically configured to: extract keywords from the topic name and the topic description information to obtain topic keywords; retrieve first recall information associated with the topic keywords based on the topic keywords; concatenate and embed the topic name and the topic description information to obtain an embedding vector corresponding to the target topic; retrieve similar information vectors similar to the embedding vector in the vector database based on the embedding vector to obtain second recall information corresponding to the similar information vector, wherein the vector database contains multiple information vectors and corresponding event information; and merge the first recall information and the second recall information as first candidate information.

[0096] It should be noted that the apparatus provided in the above embodiments and the method provided in the above embodiments belong to the same concept, and the specific way in which each module performs the operation has been described in detail in the method embodiments, and will not be repeated here.

[0097] Figure 10 A schematic diagram of the structure of a computer system suitable for implementing the electronic device of the present application is shown.

[0098] It should be noted that, Figure 10 The computer system 1000 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0099] like Figure 10 As shown, the computer system 1000 includes a Central Processing Unit (CPU) 1001, which can perform various appropriate actions and processes based on programs stored in Read-Only Memory (ROM) 1002 or programs loaded from storage section 1008 into Random Access Memory (RAM) 1003. The RAM 1003 also stores various programs and data required for system operation. The CPU 1001, ROM 1002, and RAM 1003 are interconnected via a bus 1004. An Input / Output (I / O) interface 1005 is also connected to the bus 1004.

[0100] The following components are connected to I / O interface 1005: an input section 1006 including a keyboard, mouse, etc.; an output section 1007 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to I / O interface 1005 as needed. Removable media 1011, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 1010 as needed so that computer programs read from them can be installed into storage section 1008 as needed.

[0101] Specifically, according to embodiments of this application, the processes described in the various method flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1009, and / or installed from removable medium 1011. When the computer program is executed by central processing unit (CPU) 1001, it performs various functions defined in the system of this application.

[0102] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0103] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0104] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0105] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the methods according to the embodiments of this application.

[0106] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0107] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A data processing method, characterized in that, include: Obtain a content generation request for a target topic, wherein the content generation request contains descriptive information about the target topic; Media retrieval is performed based on the description information of the target topic to obtain the first candidate information; The target event nodes related to the target topic are queried from a preset event knowledge graph. The event knowledge graph contains multiple event nodes and event association edges. The event nodes are used to represent events and contain second candidate information related to the events. The event association edges are used to represent the association relationship between events. The first candidate information is merged with the second candidate information contained in the target event node to obtain a candidate information set; Based on the degree of correlation between the target topic and the candidate information in the candidate information set, candidate information matching the target topic is filtered and combined from the candidate information set to obtain the content data of the content generation request.

2. The method according to claim 1, characterized in that, The step of querying target event nodes related to the target topic from a preset event knowledge graph includes: Based on the description information of the target topic, locate the first event node corresponding to the target topic in the preset event knowledge graph; Based on the event association edges with the first event node, the event nodes located downstream of the first event node are traversed in the event knowledge graph to obtain the second event node; The first event node and the second event node are used together as target event nodes related to the target topic.

3. The method according to claim 1, characterized in that, The step of filtering and combining candidate information matching the target topic from the candidate information set based on the relevance between the target topic and the candidate information in the candidate information set to obtain the content data of the content generation request includes: Obtain the first event node corresponding to the target topic in the event knowledge graph; If the first event node contains a predetermined information sequence set for the target topic, then the candidate information corresponding to the predetermined information sequence is obtained from the candidate information set and combined to obtain the first content chapter. The candidate information in the predetermined information sequence is selected based on its relevance to the target topic. The first content chapter is filled into the corresponding position in the predetermined content structure to obtain the content data of the content generation request.

4. The method according to claim 3, characterized in that, Before filling the first content chapter into the corresponding position in the predetermined content structure to obtain the content data of the content generation request, the method further includes: If the first event node does not contain the predetermined information sequence, then a first prompt word is generated based on the description information of the target topic, the summary information of the candidate information in the candidate information set, and the predetermined prompt information template. The first prompt word is used to guide the large language model to generate the topic information sequence. The first prompt word is input into the large language model so that the large language model can filter and sort the candidate information according to the relevance of the candidate information to the target topic, and obtain the topic information sequence output by the language model. Candidate information corresponding to the topic information sequence is obtained from the candidate information set and combined to obtain the first content chapter.

5. The method according to claim 3, characterized in that, Before filling the first content chapter into the corresponding position in the predetermined content structure to obtain the content data of the content generation request, the method further includes: If the content generation request contains a deep analysis identifier, then based on the similarity between other candidate information in the candidate information set and candidate information in the first content chapter, the remaining candidate information corresponding to the similarity being lower than a predetermined threshold is selected, wherein the other candidate information is candidate information in the candidate information set that does not belong to the first content chapter; Based on the remaining candidate information and the predetermined prompt information template, a second prompt word is generated. The second prompt word is used to guide the large language model to generate multiple sub-topics based on the remaining candidate information. The second prompt word is input into the large language model to obtain multiple sub-topics and the remaining candidate information corresponding to each sub-topic. The remaining candidate information corresponding to each sub-topic is combined to obtain multiple second content chapters, which are used to fill the corresponding positions in the predetermined content structure.

6. The method according to claim 4 or 5, characterized in that, The method further includes: Obtain the client's information editing records for the generated content data, the information editing records containing modification operations performed on candidate information contained in the content data; Based on the modification operation in the information editing record, constraint information based on the modification operation is added to the prompt information template. The constraint information is used to indicate the large language model.

7. The method according to claim 1, characterized in that, The step of merging the first candidate information with the second candidate information contained in the target event node to obtain a candidate information set includes: Obtain the second candidate information corresponding to each of the target event nodes; Based on the information identifiers of the first candidate information and the second candidate information, calculate the union between the first candidate information and the second candidate information to obtain the union set; Using a large language model, the relevance between the descriptive information of the target topic and the summary information of each candidate information in the union set is scored to obtain the relevance score corresponding to each candidate information. Based on the relevance score corresponding to each candidate information, candidate information with a corresponding relevance score lower than a predetermined score threshold is deleted from the union set to obtain a candidate information set.

8. The method according to claim 7, characterized in that, The step of deleting candidate information with a relevance score lower than a predetermined score threshold from the union set based on the relevance score corresponding to each candidate information, to obtain a candidate information set, includes: Based on the relevance score corresponding to each candidate information, candidate information whose corresponding relevance score is lower than a predetermined score threshold is deleted from the union set to obtain a set of filtered results. Clustering is performed on the semantics of each candidate information in the filtering result set to obtain multiple information clusters; For each information cluster, based on the quality score of each candidate information in the information cluster, select the candidate information whose quality score meets the predetermined score condition to obtain the information subset corresponding to each information cluster; The subsets of information corresponding to each information cluster are merged to obtain the candidate information set.

9. The method according to claim 7, characterized in that, The method further includes: Extract the metadata of each candidate information in the union set, wherein the metadata includes at least one of information source, quality score and information status; Based on the metadata, candidate information that does not meet the predetermined filtering conditions is deleted from the union set, wherein the predetermined filtering conditions include constraints set for the metadata.

10. The method according to claim 1, characterized in that, The step of obtaining a content generation request for the target topic includes: The system receives a generation instruction sent by the client. The generation instruction includes setting a topic name, topic description, and system parameters for the target topic. The system parameters are used to filter candidate information of the target topic according to predetermined filtering conditions. The topic name and the topic description are concatenated to form the description information of the target topic; Based on the description information and the system parameters, a content generation request for the target topic is generated.

11. The method according to claim 1, characterized in that, After the step of filtering and combining candidate information matching the target topic from the candidate information set based on the relevance between the target topic and the candidate information in the candidate information set to obtain the content data of the content generation request, the method further includes: The client displays a content draft interface, which includes a first component, a second component, and a third component. The first component displays the name of each content chapter in the content data, the second component displays the information preview interface corresponding to the content chapter selected in the first component, and the third component displays the editing interface of the content chapter selected in the first component. According to the information editing instruction sent by the client, the content data is updated and sent to the client to display the editing result in the third component, wherein the information editing instruction is triggered by an information editing operation in the third component; Based on the confirmation command sent by the client, the topic information containing the content data is generated and sent to the client so as to display the information display page of the target topic in the client.

12. The method according to claim 1, characterized in that, The description information includes the topic name and topic description information of the target topic; the step of performing media retrieval based on the description information of the target topic to obtain the first candidate information includes: Keyword extraction is performed on the topic name and the topic description information to obtain topic keywords; Based on the topic keywords, retrieve the first recall information associated with the topic keywords; The topic name and the topic description information are concatenated and embedded to obtain the embedding vector corresponding to the target topic; Based on the embedding vector, similar information vectors that are similar to the embedding vector are retrieved from the vector database to obtain the second recall information corresponding to the similar information vectors. The vector database contains multiple information vectors and corresponding event information. The first recall information and the second recall information are combined as the first candidate information.

13. A data processing apparatus, characterized in that, include: The content acquisition module is configured to acquire a content generation request for a target topic, wherein the content generation request includes descriptive information of the target topic; The media retrieval module is configured to perform media retrieval based on the description information of the target topic to obtain first candidate information; The graph query module is configured to query target event nodes related to the target topic from a preset event knowledge graph. The event knowledge graph contains multiple event nodes and event association edges. The event nodes are used to represent events and contain second candidate information related to the events. The event association edges are used to represent the association relationships between events. The information merging module is configured to merge the first candidate information with the second candidate information contained in the target event node to obtain a candidate information set; The information filtering module is configured to filter and combine candidate information matching the target topic from the candidate information set based on the degree of relevance between the target topic and the candidate information in the candidate information set, so as to obtain the content data of the content generation request.

14. An electronic device, characterized in that, include: processor; Memory for storing the executable instructions of the processor; The processor is configured to execute the data processing method of any one of claims 1 to 12 by executing the executable instructions.

15. A computer-readable medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the data processing method as described in any one of claims 1 to 12.

16. A computer program product, characterized in that, The computer program product includes computer instructions stored in a computer-readable storage medium, a processor of a computer device reading the computer instructions from the computer-readable storage medium, and the processor executing the computer instructions to cause the computer device to perform the data processing method as described in any one of claims 1 to 12.

Images