Event element atlas-based propagation strategy intelligent analysis system

Abstract

The application belongs to the technical field of event information retrieval, and discloses a spreading strategy intelligent analysis system based on an event element graph. A source index management module is used to complete the qualification grading and credibility scoring quantization labeling of multi-field sources, to perform normalized indexing, bidirectional mapping indexing and keyword inverted indexing construction on full-channel and multi-format materials, to generate a unique traceability code for each material, and to realize the standardized management and rapid tracing of the materials. The semantic analysis capability of an element extraction coding module is used to divide information units into three core elements, namely, event background, objective facts and multi-party views, to establish the mapping association between the elements and the sources through semantic hash coding and cross-source same-reference alignment, and to construct an event element graph with a three-layer core structure and four types of composite index mechanisms, so as to realize multi-dimensional retrieval and rapid calling of the elements and to improve the information processing regularity and calling efficiency.

Description

Technical Field

[0001] This invention belongs to the field of event information retrieval technology, specifically an intelligent analysis system for propagation strategies based on event element maps. Background Technology

[0002] In the current multimedia communication landscape, the efficiency of event-based content production and the accuracy of its dissemination have become core components of media organizations' competitiveness. Content production tools are gradually incorporating natural language processing technology to assist in manuscript writing and material organization. However, existing technical solutions still have the following technical problems:

[0003] Existing solutions are unable to perform hierarchical element decomposition and classification of original event materials, making it difficult to distinguish information units with different attributes such as event background, objective facts, and multiple viewpoints. They are also unable to establish logical connections between various elements, resulting in fragmented event cognition and difficulty in providing unified and structured data support.

[0004] Current fact-checking methods rely heavily on manual cross-verification of multiple sources. They cannot conduct multi-dimensional cross-verification based on an authoritative and credible source system, making it difficult to identify and filter false information at the source. They also cannot manage the verification results with traceable marking, and the reliability and transparency of content production cannot be reliably guaranteed.

[0005] Existing intelligent content generation technologies rely on splicing materials and rewriting sentences as their core approach. They cannot simulate professional investigative thinking to conduct multiple rounds of questioning and logical deduction of events, making it difficult to uncover the deep connections and essential motivations behind events. The resulting content remains at the level of superficial information integration and cannot meet the production requirements of in-depth investigative reports.

[0006] Existing intelligent content creation solutions mostly output standard manuscripts in general formats, failing to tailor customized content arrangement based on media positioning, target audience characteristics, and dissemination platform attributes, and unable to match corresponding dissemination strategies based on event element characteristics; the content production and dissemination delivery stages are disconnected, making it difficult to achieve precise reach for high-quality content, and failing to achieve a balance between large-scale production and maximum dissemination effect. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent analysis system for propagation strategies based on event element graphs, in order to solve one or more problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an intelligent analysis system for propagation strategies based on event element graphs, comprising the following modules:

[0009] Furthermore, the information source index management module performs qualification classification and credibility score quantification labeling on information sources in different fields, includes multiple types of information source subjects, and covers multiple content fields; performs information source verification and index update, periodically calculates credibility scores based on dimensions such as information correction rate, cross-source expression consistency, and information release stability of information sources, synchronously updates the weight labels in the information source index library, and removes information source subjects with credibility scores lower than the preset threshold;

[0010] Perform targeted retrieval and normalized indexing of original event materials, unify the format and index metadata of materials of different formats and distribution channels, generate a unique source traceability code for each material, record the source link, collection time and credibility weight of the material, and construct a two-way mapping index between the source and the material and a keyword inverted index structure.

[0011] Furthermore, the element extraction and encoding module performs hierarchical element extraction and standardized retrieval encoding on the material data that has been normalized and indexed, based on the semantic understanding capability of the semantic extraction large language model.

[0012] Based on the fine-tuned semantic extraction big language model, the attribute features of each information unit in the material are identified, and the information units in the material are automatically split into event background elements, objective fact elements, and multi-party viewpoint elements. The event background elements include the preconditions, industry environment, historical related information, etc. The objective fact elements include verifiable factual information such as the time, place, participating entities, event process, and final result of the event. The multi-party viewpoint elements include subjective content such as the public statements, analysis and judgment, and evaluation statements of different entities.

[0013] The extracted elements are labeled with attributes and semantic hashed. A unique element retrieval code is generated according to the element's level, information type, and associated subject. Automatic alignment of cross-source elements is performed, and elements from different sources that express different things but point to the same fact are encoded and associated. An element-source mapping index is established, and standardized searchable element units are output.

[0014] Furthermore, the element graph construction module uses element units as graph nodes, establishes connection edges between nodes based on multiple dimensions such as semantic association, logical causality, time series, and subject association, and labels each connection edge with association strength weight and association type label to form a network structure event element graph.

[0015] The event element map is hierarchically divided into three core layers: basic facts layer, background association layer, and opinion game layer. The objective fact nodes in the basic facts layer are synchronously attached with attribute information such as verification status, credibility score, and evidence chain.

[0016] A composite indexing mechanism for event element maps is established, which includes four types of index structures: inverted index of element type, time axis clustering index, association strength weight index, and verification status filtering index. It supports retrieval and information retrieval by element type, association strength, time dimension, and credibility.

[0017] Furthermore, the fact verification and update module extracts objective fact nodes from the event element map and automatically matches them with corresponding materials in the trusted source index for cross-verification through element retrieval codes;

[0018] By leveraging the semantic comparison capabilities of the semantic extraction big language model, the consistency of the statements of the same fact in different information sources is verified. Fact nodes with differing statements are marked with risk and assigned credibility values. The differing content and corresponding source tracing codes are simultaneously annotated to form a searchable node evidence chain index.

[0019] The verification results are synchronously updated to the attribute fields and corresponding index structure of the corresponding objective fact nodes, and the verification status, credibility score and credible source support basis of each fact node are updated in real time. For information that cannot be verified through the included sources, a targeted search query is automatically generated, a list of to-be-verified indexes is generated and an incremental search supplement to the credible source database is triggered.

[0020] Furthermore, the deep reasoning retrieval module is equipped with a reasoning big language model, which uses multi-round iterative retrieval as the core logic to perform deep retrieval and reasoning on events;

[0021] Based on the existing index information in the event element map, multi-round structured search queries are automatically generated, including multiple dimensions such as event cause tracing, correlation impact inference, stakeholder analysis, and development trend prediction. According to each round of search query, the trusted information source index library is called to perform targeted information retrieval. After each round of retrieval is completed, the supplementary elements obtained by the retrieval are aligned and updated in the event element map, and all corresponding index structures are updated synchronously.

[0022] Once the entire deduction process is completed, the system automatically generates analytical clues and logical chains, while also retaining the deduction retrieval path index for the entire process.

[0023] Furthermore, the retrieval aggregation output module has multiple built-in information organization architecture templates to cover different retrieval scenarios, and each template corresponds to clear information section division and logical arrangement rules;

[0024] Based on the scenario type and information granularity requirements of the search needs, the corresponding information organization structure and content detail level are matched, and the corresponding level of element content is automatically retrieved from the event element map through composite index, and the content is integrated and structured according to the logical order of the information organization structure.

[0025] It supports search output adaptation for multiple distribution channels. It automatically adjusts the length, expression and presentation of search results according to the content format requirements of different information distribution platforms and users' information acquisition habits, and generates structured information content adapted to multi-channel distribution.

[0026] Furthermore, the strategy optimization and adaptation module combines event attributes, search content characteristics, target user information needs profiles, and the search characteristics of the distribution platform to generate search adaptation strategy solutions from multiple dimensions such as search push timing, distribution channel combination, information granularity adjustment, and user reach methods.

[0027] Establish a data collection link for search results, collect and statistically analyze the results data after information distribution, and extract core search results indicators.

[0028] Based on performance data, we review the actual search conversion effects of different strategy combinations, iteratively optimize the matching rules and adaptation models of search adaptation strategies, and simultaneously update the inverted index for channel and element matching.

[0029] The beneficial effects of this invention are as follows:

[0030] 1. This invention completes the qualification classification and credibility scoring quantitative labeling of information sources in multiple fields through the information source index management module. It performs normalized indexing, bidirectional mapping indexing, and keyword inverted index construction on materials from all channels and in multiple formats, generating a unique traceability code for each material, realizing standardized management and rapid traceability of materials. Through the semantic parsing capability of the element extraction and coding module, information units are divided into three core elements: event background, objective facts, and multiple viewpoints. Through semantic hash coding and cross-source alignment, a mapping relationship between elements and information sources is established. A three-layer core structure event element map is constructed, with four types of composite indexing mechanisms to realize multi-dimensional retrieval and rapid access of elements, improving the regularity of information processing and access efficiency.

[0031] 2. The fact verification and update module of this invention automatically extracts objective fact nodes from the event element graph, matches them with materials in the trusted source index for multi-source cross-verification, verifies the consistency of cross-source descriptions of the same fact through semantic comparison capabilities, automatically marks and assigns credibility values ​​to nodes with description differences, and synchronously retains the difference content and corresponding source tracing codes to form a searchable node evidence chain index. All verification results are updated in real time to the attribute fields and corresponding index structures of the corresponding objective fact nodes, ensuring the dynamic searchability of the node credibility status. For information that cannot be verified temporarily, a list to be verified is automatically generated and incremental retrieval is triggered to supplement it, improving verification efficiency and content reliability.

[0032] 3. The deep reasoning and retrieval module of this invention is equipped with a large reasoning language model. It simulates professional investigative thinking through multi-round iterative retrieval, conducting in-depth analysis from multiple dimensions, including tracing the causes of events, deducing related impacts, analyzing stakeholders, and predicting development trends. Supplementary elements obtained in each round of retrieval are synchronously aligned and updated to the event element map and all corresponding indexes. Ultimately, a complete set of analytical clues, logical chains, and a full-process reasoning path index are generated, meeting the production requirements of in-depth investigative content. The retrieval aggregation and output module has built-in multi-scenario information organization templates, which can match different needs to generate structured content and adapt to multi-channel distribution formats. The strategy optimization and adaptation module generates dissemination strategies based on event, content, user, and platform characteristics. By collecting distribution effect data, iteratively optimizing matching rules, it connects content production and dissemination, improving dissemination accuracy and actual conversion effects. Attached Figure Description

[0033] Figure 1 This is the main flowchart of the system of the present invention;

[0034] Figure 2 Update the sub-flowchart for fact verification of this invention. Detailed Implementation

[0035] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings, using a sudden public opinion event involving a brand of pre-prepared food products failing a random inspection as a specific application scenario. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0036] like Figures 1 to 2 As shown, this embodiment describes in detail the operating logic of the intelligent analysis system for dissemination strategies based on event element graphs, covering the entire process of material collection, element analysis, fact verification, in-depth analysis, content generation, and dissemination strategy optimization for food safety public opinion.

[0037] In this embodiment, the information source index management module performs qualification classification and credibility scoring quantitative labeling of information sources in the field of food safety, and includes various information source entities such as official channels of market supervision departments, national food safety testing databases, mainstream central and local official media, and leading food industry vertical media, covering multiple content dimensions such as regulatory reports, industry analysis, and public opinion dissemination;

[0038] Perform source verification and index updates. Calculate a credibility score monthly based on three dimensions: information correction rate, cross-source representation consistency, and information release stability. Update the weight labels in the source index database synchronously and remove source entities with a credibility score below 0.3.

[0039] Before calculation, the statistical period is a calendar month. All data released by the information source subject within the period are collected, and quantitative values ​​of three dimensions are calculated respectively: information correction rate is the ratio of the number of correction statements released within the statistical period to the total number of content released in the same period; cross-source expression consistency is the average semantic overlap between factual content on the same topic and the official announcements of the State Administration for Market Regulation; information release stability is the coefficient of variation of the release time interval (the ratio of standard deviation to mean). The lower the coefficient of variation, the higher the stability.

[0040] The quantitative values ​​of information correction rate and cross-source representation consistency are positively normalized and mapped to the 0-1 range, while the quantitative values ​​of information release stability are negatively normalized and mapped to the 0-1 range. The higher the coefficient of variation, the lower the value after mapping. Then, the values ​​are weighted and summed according to the dimensional weights of 0.3:0.5:0.2, and finally mapped to the credibility score range of 0-1, which corresponds to the credibility score labeling results of the information source subject.

[0041] Source credibility scoring formula:

[0042]

[0043] This represents the credibility score of the information source, with a value range of [0, 1]. The higher the score, the higher the overall credibility of the information source.

[0044] The positively normalized quantitative value of the information correction rate dimension is represented, with a value range of [0, 1]. It is obtained by positively normalizing the ratio of the number of information source correction statements to the total number of content published in the same period within the statistical period. The lower the value, the less information source error correction and the higher the accuracy of the content.

[0045] This represents the positively normalized quantitative value of the cross-source expression consistency dimension, with a value range of [0, 1]. It is the average semantic overlap between factual content on the same topic and authoritative benchmark information sources. The higher the value, the higher the degree of fit between the source expression and authoritative information.

[0046] The value represents the inverse normalization quantification value of the information release stability dimension, with a value range of [0, 1]. It is obtained by inverse normalization of the dispersion coefficient (the ratio of the standard deviation to the mean) of the information source release time interval. The higher the value, the more stable the information source release rhythm.

[0047] In response to the public opinion surrounding the substandard pre-prepared food inspection, we conducted targeted retrieval and normalized indexing of original event materials. We standardized the format and added metadata indexes to materials from different formats and distribution channels, including official announcements, media press releases, self-media articles, short video scripts, and industry report PDFs. We generated a unique source tracing code for each piece of material, recorded the source link, collection time, and credibility weight of the material, and constructed a two-way mapping index between sources and materials, as well as a keyword inverted index structure.

[0048] In the format unification stage, the materials from different carriers are uniformly converted into UTF-8 encoded plain text format, and four basic fields are extracted simultaneously: title, official release time, release subject, and original access address. In the metadata indexing and tagging stage, three types of tags are added to the materials: food safety field, genre type such as regulatory notices / news reports / netizen comments, and batch of the first day of public opinion occurrence. All fields are named according to the same rules and stored in JSON format to form a standardized material metadata data package.

[0049] The bidirectional mapping index comprises two parts: a forward index and a reverse index. The forward index uses the unique code of the source entity as the key, for example, using the code "SCJG001" for the "Provincial Market Supervision Bureau" as the key. The corresponding value is a list of source codes for all public opinion-related materials under that source. Each source code is accompanied by a metadata summary of three types: material collection time, credibility weight, and content format type. The reverse index uses the material's source code as the key, and the corresponding value is the source entity code to which the material belongs, along with a metadata summary of the material, enabling rapid source mapping from material to source. A full-text inverted index is simultaneously constructed at the keyword and element tag dimensions, supporting the location of related materials and sources through feature keywords.

[0050] The source tracing code adopts a three-segment fixed-length combination structure, which includes a 7-digit source main code segment, an 8-digit acquisition time code segment, and a 4-digit material sequence number segment. The source main code segment corresponds to the unique identification code of the source to which the material belongs, the acquisition time code segment is generated by converting the year, month, and day of the material acquisition time, and the material sequence number segment is the sequential number of the materials acquired from the same source and in the same batch. The combination of the three code segments forms a globally unique material tracing identifier.

[0051] For example, the source code of the third inspection report collected by the Provincial Market Supervision Bureau on October 15, 2025 is “SCJG001202510150003”.

[0052] In this embodiment, the element extraction and coding module performs hierarchical element extraction and standardized retrieval coding on the public opinion material data that has been normalized and indexed, based on the semantic understanding capability of the semantic extraction large language model.

[0053] The input to the semantic extraction large language model consists of the plain text of the inspection report after normalization and indexing, and the attached metadata fields. Before input, the material text exceeding the threshold of the number of tokens in the model context window is segmented according to the window length. Before segmentation, the text is first tokenized. The token quantity of a single segment reserves space for auxiliary identifiers and the overlapping area of ​​the context. The corresponding source tracing code and material collection timestamp are embedded at the beginning of each segment as auxiliary identifiers. Adjacent segments have a 15% text overlap area. After extraction, semantic merging and deduplication verification are performed on the same elements across segments.

[0054] The model output is a set of structured element triples. For example, for a random inspection report, the output is a triple of "total bacterial count exceeds the national standard by 2 times", "objective fact", and "brand involved". The element category label is limited to three enumeration values: event background, objective fact, and multiple viewpoints.

[0055] The model output is directly fed into the subsequent attribute labeling stage. The feature category label is used to match the semantic hash encoding rules of the corresponding level, and the feature association subject field is used for initial screening and matching of cross-source same-point alignment.

[0056] Based on a finely tuned semantic extraction big language model in the field of food safety, the system identifies the attribute features of each information unit in public opinion materials and automatically splits the information units in the materials into three core elements:

[0057] Background elements include the sampling standards for the pre-prepared food industry, the brand's past sampling records, and historical cases of quality problems with similar products.

[0058] Objective factual elements include verifiable factual information such as the sampling time, the city where the sampling was conducted, the brand and product model involved, the specific value of the total bacterial count exceeding the standard, and the penalties imposed by the regulatory authorities;

[0059] Multiple perspectives include responses from the companies involved, analyses and comments from food industry experts, consumer complaints and feedback, and subjective content such as interpretations and evaluations from social media.

[0060] The extracted elements are labeled with attributes and semantic hashed, and a unique element retrieval code is generated according to the element's level, information type, and associated subject.

[0061] Semantic hashing encoding generates a 16-bit hexadecimal feature code based on the core semantic vector of the original text of the element. Elements with the same semantic reference have a hash code hexadecimal character overlap of more than 85% matching threshold, which is used for initial fast matching of elements with the same reference.

[0062] The element retrieval code adopts a six-segment fixed-length combination structure, which includes a 2-digit element level identifier segment, a 2-digit information type identifier segment, an 8-digit associated subject code segment, a 4-digit time feature code segment, a 4-digit location feature code segment, and a 16-digit semantic hash feature segment. The whole code adopts a hexadecimal character format, which corresponds to the element's level, information sub-type, unique associated subject code, event time feature, event location feature, and element semantic hash value, respectively, serving as the globally unique retrieval identifier for the element.

[0063] Automatic alignment of elements across information sources is performed. Elements with different expressions such as "total bacterial count exceeds the standard" and "microbial indicators are unqualified" in different media but point to the same sampling result are encoded and associated, an element-source mapping index is established, and standardized searchable element units are output.

[0064] The alignment process first performs an initial screening based on the associated subject coding segment in the element retrieval code to select a set of candidate elements that match the brands involved. For objective fact elements, in addition to the initial screening, the time feature coding segment and location feature coding segment in the retrieval code are added for precise matching and verification. For event background elements and multi-party viewpoint elements, the matching and verification of the event identifier of this sampling is added.

[0065] After verification, the semantic hash code of the elements in the candidate set is used to calculate the Hamming distance in hexadecimal character bits. Elements with a distance less than 3 are determined to be homonymous elements. All homonymous elements are assigned a unified homonymous cluster identifier, such as "TZ-20251015-001". A bidirectional mapping association is established between the retrieval code of all homonymous elements and the homonymous cluster identifier.

[0066] In this embodiment, the element graph construction module uses standardized element units as graph nodes, establishes association edges between nodes based on four dimensions: semantic association, logical causality, time series, and subject association, and labels each association edge with association strength weight and association type label to form a network structure event element graph;

[0067] The association strength weights are uniformly mapped to the range of 0 to 1, and the quantification rules for each dimension are as follows: The semantic association dimension directly determines the weight value based on the overlap of semantic hash codes between nodes; the logical causal dimension is assigned values ​​of 0.8 and 0.3 respectively based on whether the elements are connected by causal conjunctions such as "cause" and "trigger" in the original text; the time series dimension starts with a weight of 1.0 when the time interval is 0 days, and decreases in the opposite direction as the number of days increases, with the weight decreasing by 0.05 for every additional day, down to a minimum of 0.1; the subject association dimension determines the weight value based on the ratio of the number of shared associated subjects between two nodes to the union of the total number of associated subjects between the two nodes, with the ratio being positively correlated with the weight value, and the value range being 0 to 1.

[0068] The final association strength weight is the weighted average of the multi-dimensional calculated values. The corresponding weight is matched according to the dominant type of the association edge: if the association edge is dominated by logical causality, the weights are allocated according to logical causality, semantic association, time series, and subject association at a ratio of 0.6:0.2:0.1:0.1; if the association edge is dominated by subject association, the weights are allocated according to semantic association, logical causality, time series, and subject association at a ratio of 0.1:0.1:0.2:0.6. The weights are calculated and updated synchronously with the graph node updates.

[0069] Formula for the comprehensive strength weight of logically causally dominant related edges:

[0070]

[0071] The weight represents the overall association strength of logically causally dominant association edges, with a value range of [0, 1]. The higher the weight, the stronger the association between the two graph nodes.

[0072] This represents the strength of the logical causal relationship, with a value range of [0, 1]. If the corresponding elements of the two nodes are connected by explicit causal conjunctions such as "cause" or "trigger" in the original text, the value is 0.8; otherwise, the value is 0.3.

[0073] This represents the semantic association strength value, ranging from [0, 1], and is directly determined by the overlap of the semantic hash codes of the two nodes;

[0074] This represents the correlation strength value of the time series dimension, with a value range of [0.1, 1]. The initial value is 1.0 when the time interval is 0 days. The weight decreases by 0.05 for each additional day of the interval, with a minimum retention of 0.1.

[0075] This represents the association strength value of the subject association dimension, with a value range of [0, 1]. It is equal to the ratio of the number of shared associated subjects between the two nodes to the number of the union of all associated subjects between the two nodes.

[0076] The event element map is hierarchically divided into three core layers: the basic fact layer, the background association layer, and the opinion game layer. The basic fact layer contains objective fact nodes such as sampling time, exceeding the standard value, and punishment measures. Each objective fact node is simultaneously attached with attribute information such as verification status, credibility score, and evidence chain. The background association layer contains background nodes such as industry standards and brand historical sampling records. The opinion game layer contains opinion nodes such as corporate responses, expert comments, and netizens' feedback.

[0077] The relationships between levels follow fixed constraints. Nodes in the basic fact layer are associated with nodes in the background association layer and the opinion game layer, respectively. Nodes in the background association layer and the opinion game layer can establish direct association edges based on semantic association and subject association. The conditions for establishing an association are that there is a shared associated subject between the nodes, the semantic matching degree is higher than 0.7, or there is a clear temporal causal connection relationship. The association edge is simultaneously labeled with its respective level type.

[0078] A composite indexing mechanism for event element maps is established, which includes four types of index structures: element type inverted index, time axis clustering index, association strength weight index, and verification status filtering index. It supports retrieval and information retrieval by element type, association strength, time dimension, and credibility. For example, all verified fact nodes can be extracted with one click through the verification status filtering index.

[0079] The inverted index by element type uses element category labels as keys. For example, using "objective facts" as the key, the corresponding value is a list of search codes for all sampling times, exceeding values, and involved entities. The time-axis clustering index supports switching between hourly, daily, and monthly granularities. Using the time interval "2025-10-15" as the key, the corresponding value is a list of search codes for all public opinion element nodes published on that day.

[0080] The association strength weight index uses the retrieval code of the "Total Colony Count Exceeds Standard" node as the key, and the corresponding value is a list of associated nodes whose associated edges are sorted in descending order of weight. Each record is also accompanied by labels such as causal association and subject association, along with corresponding weight values. The verification status filtering index uses the "Verification Passed" status identifier as the key, and the corresponding value is a list of retrieval codes of all fact nodes that have undergone multi-source verification.

[0081] In this embodiment, the fact verification and update module extracts objective fact nodes from the event element map and automatically matches the corresponding materials in the trusted information source index library for cross-verification through element retrieval coding;

[0082] By leveraging the semantic comparison capabilities of a large language model through semantic extraction, the consistency of statements in the same sampled results across different sources, such as official reports and mainstream media, is verified. Fact nodes with discrepancies in expression are flagged for risk and assigned credibility values. The discrepancies are simultaneously annotated with corresponding source tracing codes, forming a searchable index of node evidence chains. Verification status is categorized into three levels: verified, pending verification, and awaiting verification, determined based on both expression consistency scores and the number of supporting sources.

[0083] If the expression consistency score is higher than 0.8 and there are no fewer than 2 credible sources supporting it, it is considered to have passed the verification; if the expression consistency score is higher than 0.8 but there is only 1 credible source supporting it, or if the score is in the range of 0.4-0.8 and there is no fewer than 1 credible source supporting it, it is considered to be questionable and pending verification; if there are 0 credible sources supporting it, or if the score is lower than 0.4, it is considered to be pending verification. Different levels correspond to different status indicators and weight coefficients, which are synchronously mapped to the node credibility assignment results.

[0084] When performing semantic comparison, the model input consists of two material fragments: a report from the provincial market supervision bureau and a local media report, corresponding to the fact that "total bacterial count exceeds the standard." Each fragment is accompanied by the credibility weight of the corresponding source (0.95 and 0.82, respectively). The model outputs two items: first, a consistency score of 0.92, which quantifies the degree of consistency between the two sources' descriptions of the fact that the count exceeds the standard; and second, the result of locating the difference fragments, which marks the numerical discrepancy range and the type of quantitative difference between "3 times the standard" in the local media report and "2 times the standard" in the official report.

[0085] The output consistency score is weighted by the average credibility weight of all supporting sources at a ratio of 0.6:0.4, and the resulting value in the 0-1 range is used as the credibility value of the fact node. The difference fragment location result is mapped one-to-one with the corresponding source tracing code and stored in the node evidence chain sub-index.

[0086] Formula for assigning credibility values ​​to objective fact nodes:

[0087]

[0088] The final score represents the credibility of the objective fact node, with a value range of [0, 1]. The higher the score, the higher the credibility of the fact node.

[0089] This represents the consistency score of multiple sources, with a value range of [0, 1]. It is output by the semantic comparison model and quantifies the degree of fit between different sources in their statements of the same fact.

[0090] This represents the average credibility score of all credible sources supporting the fact, with a value range of [0, 1], reflecting the overall credibility level of the sources supporting the fact.

[0091] The node evidence chain sub-index uses the retrieval code of the fact node "total colony count exceeds the standard" as the key. The corresponding value contains five related information items: difference content "exceeds the standard by 3 times", original text start and end position offset "characters 128-133", corresponding source tracing code "MT002202510150007", source credibility weight 0.82, and expression consistency score 0.92.

[0092] Source tracing codes are directly linked to the full-text storage address of the corresponding material. The consistency score is retained to two decimal places. All records are sorted in descending order of source credibility weight, and pagination retrieval is supported. The entire data is included in the scope of the associated data of the verification status filtering index.

[0093] The verification results are synchronously updated to the attribute fields and corresponding index structure of the corresponding objective fact nodes, and the verification status, credibility score and credible source support basis of each fact node are updated in real time. For information such as "all products involved have been removed from the shelves" that cannot be verified through the included sources, a targeted search query of "a certain brand of pre-prepared dishes has been removed from the shelves and recalled" is automatically generated, an index of the list to be verified is generated and incremental search and supplementation of the credible source expansion library of the food supervision system and the brand's official channels is triggered.

[0094] Incremental retrieval is limited to pre-defined trusted sources such as information databases under market supervision departments and official brand accounts. The retrieval query is generated based on the core semantics of the node to be verified and the brand involved. The enterprise removal announcements obtained by the retrieval are first verified for the corresponding source's qualifications and credibility rating. Sources that meet the access standards are included in the source index library. Then, material normalization indexing and element extraction processing are performed. Finally, the nodes to be verified are aligned with the same index. The verified matching elements are synchronously updated to the event element map and the corresponding composite index structure. At the same time, the verification status of the corresponding node is refreshed to "verification passed".

[0095] In this embodiment, the deep inference retrieval module is equipped with a reasoning big language model and uses multi-round iterative retrieval as the core logic to perform deep retrieval and inference on this public opinion event.

[0096] Taking the event cause tracing dimension as an example, the model input consists of verified basic facts of "excessive total bacterial count," a summary of background elements of "hygiene standards for pre-prepared food production," and deduction instructions for "tracing the cause of the exceedance." Before input, the core content and association strength weights of the corresponding nodes are extracted through a composite index and organized into a structured prompt sequence. The model outputs standardized retrieval queries for "reasons in the production process for excessive total bacterial count in pre-prepared food" and "limited food production supervision areas," with the retrieval query directly calling the trusted source index library to perform targeted retrieval. After each round of retrieval, supplementary elements such as production process loopholes and non-standard personnel operations are updated to the graph through same-index alignment and then reorganized into a new input sequence for the next round of deduction.

[0097] Based on the existing index information in the event element graph, multi-dimensional and multi-round structured search queries are automatically generated: query queries such as "production process loopholes and raw material contamination" are generated from the event cause tracing dimension; query queries such as "brand sales impacting industry regulatory tightening" are generated from the correlation impact inference dimension; query queries such as "upstream and downstream suppliers of the involved company" are generated from the stakeholder analysis dimension; and query queries such as "rectification measures and industry standard updates" are generated from the development trend prediction dimension. Based on each round of query, a trusted source index library is called to perform targeted retrieval. After each round of retrieval is completed, the supplementary elements are aligned and updated in the event element graph, and all corresponding index structures are updated synchronously.

[0098] After each round of retrieval, the termination conditions for the corresponding dimension are checked. The current dimension's inference is terminated if any of the following conditions are met: First, the number of newly added elements supplemented in two consecutive rounds of retrieval is less than 3, indicating that there is no more effective information to be mined; second, the current dimension's inference rounds reach the upper limit of 5 rounds to avoid unlimited retrieval; third, all effective queries that can be generated based on existing elements under the current dimension have been executed, and no new effective queries can be generated. After the single dimension terminates, it automatically switches to the next inference dimension. When all dimensions trigger the termination conditions, the overall inference process ends.

[0099] After the entire deduction process is completed, the system automatically generates analytical clues and logical chains, while retaining a deduction retrieval path index for the entire process. The path index is stored in clusters according to four deduction dimensions: event cause, related impact, stakeholders, and development trend. Each path records five types of information: the corresponding deduction round number, the original text of the retrieval query, the retrieval trigger node code, the list of newly added elements in this round, and the source tracing code. A two-way mapping relationship is established between the path index entries and the corresponding analytical clues and logical chain nodes. The complete cause retrieval deduction link can be traced back through the "total colony count exceeds the standard" node code.

[0100] The path index is stored in clusters according to four inference dimensions. For example, the path record of the second round under the cause dimension is as follows: round number 2, retrieval query "hygiene standards of pre-prepared vegetable production workshop", trigger node code "0201BB000001A3F2...", list of newly added element codes "[0103CC000002...]", and associated source code "SCJG001202510150012". The path index is bidirectionally mapped to the analysis clues and logical chain nodes, and the complete inference link can be traced back through any node.

[0101] When tracing the logical chain, the core fact of "pre-prepared food failing inspection" is used as the starting point. Nodes are traversed from high to low correlation strength, forming the main logical chain in a causal sequence of "raw material contamination → substandard production disinfection → excessive total bacterial count → failing inspection → regulatory penalties." Branch nodes such as "other products from the same brand passing inspection" serve as supplementary chains. Analytical clues are extracted according to four dimensions. For example, the causal dimension extracts the clue "non-standard operation by production workshop personnel as the core cause," simultaneously attaching corresponding element codes and deductive path identifiers.

[0102] In this embodiment, the retrieval aggregation output module has multiple built-in information organization structure templates, covering different retrieval scenarios such as event summaries, fact-checking, in-depth analysis, and opinion summaries. For example, the event summary template is set up with four main sections: "Event Overview → Core Facts → Responses from All Parties → Impact Analysis," and the fact-checking template is set up with four main sections: "Check Items → Comparison of Multi-Source Evidence → Verification Conclusion → Risk Warning." Each template corresponds to clear information section division and logical arrangement rules.

[0103] In response to the official notification requirements for this public opinion event, the analysis scenario is identified as "event summary" and the granularity parameter is "standard". The corresponding event summary template is matched, and the element retrieval rules corresponding to the standard granularity are used to select the top 70% of the elements with the strongest correlation to the core event nodes. The elements of the corresponding level are automatically retrieved from the event element map through a composite index. The content is integrated and structured according to the template's logical order, and a formal element retrieval list is generated and sent to the retrieval execution module.

[0104] After retrieving the elements, deduplication is first performed based on the same-index cluster identifier. For example, within the same-index cluster of "total bacterial count exceeds the standard," the elements reported by the Provincial Market Supervision Bureau are retained, while statements from less credible self-media are removed. After deduplication, the elements are sorted according to the template section rules: the core facts section is sorted in chronological order of sampling, notification, and punishment; the responses from all parties are sorted in descending order of the credibility of the sources (enterprises, experts, and netizens); and the background information section is sorted in descending order of its relevance to the event exceeding the standard. After sorting, the template sections are filled in sequentially. Logical connecting statements are automatically generated between sections based on the semantic content of the preceding and following sections to ensure content coherence.

[0105] Supports search output adaptation for multiple distribution channels: For official WeChat accounts, outputs a complete review article of about 1,500 words, using paragraph formatting and bolding of core data; for short video platforms, extracts core facts and risk warnings within 300 words, using short sentence line breaks and adapting to subtitle copywriting scenarios; for industry reference channels, outputs a detailed analysis report of more than 3,000 words, supplementing the complete evidence chain and reasoning logic, using a chapter-based structured format.

[0106] In this embodiment, the strategy optimization and adaptation module combines the attributes of food safety public opinion, the characteristics of debunking content, consumer demand profiles, and the dissemination characteristics of various platforms to generate a complete public opinion response and dissemination strategy: the timing of the push is selected during the golden response period of 12 hours when public opinion ferments; the channel combination adopts a matrix mode of "first release on official WeChat account + reprint by mainstream media + dissemination on short video platforms"; the information granularity adopts a layered configuration of "simplified version of short video + standard version of WeChat account + detailed version of official website"; and the user reach adopts a combination of "targeted push to followed users + top exposure of topic".

[0107] Establish a data collection link for search results, collect and statistically analyze data such as the reach rate of debunking content after distribution, the decline in rumor search volume, the proportion of positive comments, and the number of official information forwards, and extract core dissemination effect indicators.

[0108] The model takes four types of feature data as input: food safety incident tags, fact-checking content features, consumer risk preference tags, and short video platform characteristic parameters. The output is the recommendation weight value for each dissemination strategy dimension. The model uses the debunking reach rate and the rumor decay speed as supervision tags for incremental iteration.

[0109] For example, before the iteration, the matching weight of factual content on short video channels was 0.6. After the effect data was verified, this combination had the highest reach rate in debunking rumors. After the iteration, the weight was updated to 0.85, and it was updated in the inverted index of channel and element matching to achieve continuous optimization of strategy matching accuracy in food safety public opinion scenarios.

[0110] The channel-element matching index contains two sets of structures: forward index and inverted index. The forward index uses the "short video platform" identifier as the key, and the corresponding values ​​are a list of suitable element types [objective facts (0.85), multiple viewpoints (0.78)] and their corresponding recommendation weights. The inverted index uses the "event background" element tag as the key, and the corresponding values ​​are a list of suitable platforms [official public accounts (0.92), industry references (0.88)] and their corresponding matching scores. When the index is updated, the weight values ​​and matching scores are adjusted based on the distribution effect data, and the items in the list are arranged in descending order of scores.

[0111] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0112] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent analysis system for propagation strategies based on event element graphs, characterized in that: The system performs qualification classification and credibility quantification of information source entities, performs verification and index update of information source entities, and completes event material collection, normalized indexing and bidirectional inverted index structure construction. Hierarchical semantic parsing is performed on the event materials after normalization indexing to identify the attributes of information units and divide them into three core elements. Attribute marking, semantic hashing, and cross-source subject alignment of the elements are performed to establish a mapping index between elements and source subjects and output standardized element units. Using standardized element units as nodes in the event element graph, establishing association edges based on association logic and labeling association strength and type, hierarchically dividing the event element graph, and constructing a composite index structure; Extract objective fact nodes from the event element map, perform cross-validation by matching credible event materials in the information source index, verify the consistency of factual statements through semantic comparison, perform risk marking and credibility assignment, synchronously update the verification results to the event element map and composite index structure, generate a list to be verified and trigger incremental retrieval; Perform iterative retrieval and deduction, generate structured queries based on the index information of the event element map, and update the event element map and composite index structure after each round of retrieval by aligning the supplemented elements with the same index, and sort out and generate analysis clues, logical chains and deduction retrieval path indexes; Based on the built-in information organization architecture template, the system matches the information organization architecture template according to the search requirements, retrieves elements from the event element map, performs integration and layout, and generates structured content. By combining event, content, user, and platform characteristics, a search adaptation strategy is generated. Data on distribution effectiveness is collected and analyzed, and the strategy matching rules and composite index structure are iteratively optimized.

2. The intelligent analysis system for propagation strategies based on event element graphs according to claim 1, characterized in that, It includes multiple types of information source entities covering multiple content areas, and uses credibility scoring to classify and quantitatively label the qualifications of each type of information source entity. When performing the verification and index update of the information source subject, a credibility score is calculated periodically based on multiple dimensions, including information correction rate, cross-source expression consistency, and information release stability. The weight labels in the information source index are updated synchronously, and information source subjects with credibility scores lower than a preset threshold are removed.

3. The intelligent analysis system for propagation strategies based on event element graphs according to claim 2, characterized in that, The original event materials are collected using a targeted retrieval method. The event materials of different formats and distribution channels are standardized in format and marked with metadata index. A unique source traceability code is generated for each event material. The source link, collection time and credibility weight of the event material are recorded. A bidirectional inverted index structure of the source subject and the event material is constructed.

4. The intelligent analysis system for propagation strategies based on event element graphs according to claim 3, characterized in that, Based on the semantic extraction big language model, hierarchical element extraction is performed on the event material. After identifying the attribute features of the information units, they are divided into the three core elements, namely event background elements, objective fact elements, and multiple viewpoint elements. The event background elements include the event's preconditions, industry environment, and historical information; the objective fact elements include verifiable time, location, participating entities, event process, and outcome information; and the multi-party viewpoint elements include public statements, analyses, judgments, and evaluations from different entities.

5. The intelligent analysis system for propagation strategies based on event element graphs according to claim 4, characterized in that, The extracted elements are tagged with attributes and semantic hashed. A unique element retrieval code is generated by combining the element's level, information type, and associated subject. Automatic alignment of elements pointing to the same object across information source subjects is performed. Elements from different information source subjects that express different ideas but point to the same object are encoded and associated. The objective fact elements point to the same fact, and the event background elements and the multi-viewpoint elements point to the same subject or the same associated event. A mapping index between elements and information source subjects is established, and the searchable standardized element units are output.

6. The intelligent analysis system for propagation strategies based on event element graphs according to claim 5, characterized in that, The event element graph nodes are established from multiple dimensions, including semantic association, logical causality, time series, and subject association. Each association edge is labeled with association strength weight and association type label to form a network structure of the event element graph. The event element map is divided into four levels: basic fact layer, background association layer, opinion game layer, and verification status layer, which correspond to the three types of core elements and node verification status information, respectively. Construct the composite index structure that includes feature type inverted index, time axis clustering index, association strength weight index, and verification status filtering index.

7. The intelligent analysis system for propagation strategies based on event element graphs according to claim 6, characterized in that, The element retrieval code is used to automatically match the corresponding event material in the information source index database, and cross-validation is performed on the objective fact node. Based on the semantic comparison capability of the semantic extraction big language model, the consistency of the expression of the same fact in different source subjects is verified. For the objective fact nodes with expression differences, risk marking and credibility assignment are performed, and the difference content and the corresponding source tracing code are simultaneously marked to form a searchable node evidence chain sub-index, which is included in the scope of the verification status screening index. The verification results are synchronously updated to the event element map and the composite index structure. For information that cannot be verified by the source subject in the source index library, the list to be verified is generated and incremental retrieval is triggered.

8. The intelligent analysis system for propagation strategies based on event element graphs according to claim 7, characterized in that, The system uses a large reasoning language model to perform multiple rounds of iterative retrieval and deduction. Based on the existing index information of the event element map, it generates a structured query covering multiple dimensions, including event cause tracing, correlation and impact deduction, stakeholder analysis, and development trend prediction. After each round of retrieval is completed, the supplemented elements are aligned with the same index and updated to the event element map and the composite index structure; after all deductions are completed, the analysis clues and logical chains are sorted out and generated, and the deduction retrieval path index of the entire process is retained.

9. The intelligent analysis system for propagation strategies based on event element graphs according to claim 8, characterized in that, The system includes multiple built-in information organization architecture templates covering different retrieval scenarios, with each template corresponding to specific information module division and logical arrangement rules. According to the scenario type and information granularity requirements of the search needs, the corresponding information organization architecture template is matched, and the corresponding level of element content is retrieved from the event element map through the composite index structure, and the content is integrated and structured in logical order. It supports output adaptation to multiple distribution channels, adjusting the length, expression, and presentation of the output content according to the format requirements of the distribution platform and the user's information acquisition habits.

10. The intelligent analysis system for propagation strategies based on event element graphs according to claim 9, characterized in that, By combining event attributes, search content characteristics, target user demand profiles and distribution platform search characteristics, the search adaptation strategy scheme is generated from multiple dimensions, including search push timing, distribution channel combination, information granularity adjustment and user reach method. Establish a data collection link for search results, collect and statistically analyze the results data after distribution, and extract core performance indicators; Based on the aforementioned performance data, the strategy matching rules and adaptation model are iteratively optimized, and the matching inverted index of the distribution channels and elements is updated synchronously.

Images