Content generation method, device, electronic device and storage medium

Abstract

The present invention provides a content generation method, device, electronic device and storage medium, the method comprising: obtaining hot events according to a preset time period; performing multi-level risk factor identification on the hot events to determine hot events containing risk factors; for each hot event predicted to have risk factors, calculating the public attention score of the hot event based on the sentiment vector and topic vector of the hot event; when it is determined that the public attention score of the hot event is greater than a preset score threshold, determining the product information of the hot event based on the risk information corresponding to the hot event; constructing corresponding copy content based on the title, document content, product information, and materials of the hot event. Then, the corresponding copy content is generated. Thus, the corresponding marketing content is accurately and quickly generated.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a content generation method, device, electronic device and storage medium. Background Art

[0002] In current insurance marketing practices, salespeople often leverage breaking social news to promote content on social media platforms, aiming to increase user attention and increase insurance conversion rates. This marketing strategy, which leverages breaking news, social media, and user risk perception, is known as "hotspot marketing."

[0003] Currently, marketing content is often manually created by retrieving relevant materials from pre-built, high-quality resource libraries. Because hotspot marketing requires real-time capture of breaking news and rapid identification of risk factors, this approach is incapable of accurately and quickly generating relevant marketing content.

[0004] Therefore, there is an urgent need for an automated content generation technology solution specifically for hot marketing scenarios in the insurance industry. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a content generation method, device, electronic device, and storage medium to solve the problem in the prior art that corresponding marketing content cannot be accurately and quickly generated.

[0006] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0007] A first aspect of an embodiment of the present invention provides a content generation method, the method comprising:

[0008] Acquire hot events according to a preset time period, where the number of hot events is multiple;

[0009] Conduct multi-level risk factor identification on the hot events to determine the hot events containing risk factors;

[0010] For each hot event predicted to have risk factors, calculate the public attention score of the hot event based on the sentiment vector and topic vector of the hot event;

[0011] When it is determined that the public attention score of the hot event is greater than a preset score threshold, determining product information of the hot event based on the risk information corresponding to the hot event;

[0012] The corresponding copywriting content is constructed based on the title of the hot event, the document content, the product information, and the materials, and the materials are collected based on the title of the hot event and the document content.

[0013] Optionally, multi-level risk factor identification is performed on the hot events to determine hot events containing risk factors, including:

[0014] Splitting the title and text content of the hot event into text windows;

[0015] Call the preset classification model to perform classification prediction on each text window and obtain the prediction result;

[0016] Determine whether the prediction result of any text window is a risk factor;

[0017] If so, it is determined that the hot event contains risk factors.

[0018] Optionally, for each hot event predicted to have a risk factor, a public attention score of the hot event is calculated based on the sentiment vector and topic vector of the hot event, including:

[0019] For each hot event predicted to have risk factors, sentiment analysis and topic dimension analysis are performed on the hot event to determine the corresponding sentiment vector and topic vector;

[0020] Perform feature concatenation on the emotion vectors of different dimensions and the topic vector to obtain an input vector;

[0021] The input vector is input into a pre-built value assessment model for processing to obtain a public attention score for the hot event.

[0022] Optionally, determining product information of the hot event based on risk information corresponding to the hot event includes:

[0023] Generate risk factor analysis results based on the risk information corresponding to the hot event;

[0024] Generate a hypothetical product description document based on the risk factor analysis results corresponding to the hot event;

[0025] Calculating a semantic matching score for the hypothetical product description document;

[0026] The product information of the hot event is determined based on the risk factor analysis results and the semantic matching score.

[0027] Optionally, generating risk factor analysis results based on the risk information corresponding to the hot event includes:

[0028] For each hot event predicted to have risk factors, extract risk information from the hot event;

[0029] Processing the risk information to obtain structured information of the implicit risk;

[0030] The risk information and the structured information of the implicit risk are combined to obtain the risk factor analysis results.

[0031] Optionally, determining product information of the hot event based on the risk factor analysis result and the semantic matching score includes:

[0032] Extracting query keywords from the risk factor analysis results corresponding to the hot events;

[0033] Determining a keyword matching score for each product from a product information database based on the query keyword;

[0034] Calculating based on the semantic matching score and the keyword matching score to obtain a comprehensive matching score;

[0035] Each product in the product information database is sorted in descending order according to the comprehensive matching score, and the product information that ranks first is selected.

[0036] Optionally, corresponding copywriting content is constructed based on the title of the hot event, document content, product information, and materials, including:

[0037] Collect corresponding materials based on the title and document content of the hot event, wherein each hot event corresponds to a piece of material;

[0038] Determining a preset number of marketing content documents based on a preset dynamic creation example library and semantic similarity of titles of the hot events;

[0039] Performing reasoning path deconstruction on the preset number of marketing content documents to obtain corresponding creation prompt words;

[0040] Based on the creation prompt words, the title and text content of the hot event, materials, and product information, corresponding copywriting content is generated;

[0041] Select corresponding pictures from the materials as illustrations for the copy content.

[0042] A second aspect of an embodiment of the present invention provides a content generation device, the device comprising:

[0043] An acquisition unit, configured to acquire a plurality of hot events according to a preset time period;

[0044] an identification unit, configured to identify multi-level risk factors of the hot events and determine the hot events containing risk factors;

[0045] a processing unit configured to calculate, for each hot event predicted to have a risk factor, a public attention score of the hot event based on the sentiment vector and the topic vector of the hot event;

[0046] a determining unit, configured to determine product information of the hot event based on risk information corresponding to the hot event when it is determined that the public attention score of the hot event is greater than a preset score threshold;

[0047] The generation unit is used to construct corresponding copy content based on the title of the hot event, document content, product information, and materials, where the materials are collected based on the title of the hot event and the document content.

[0048] The third aspect of an embodiment of the present invention shows an electronic device, which includes a processor and a memory, wherein the memory is used to store program code and data for content generation, and the processor is used to call the program instructions in the memory to execute the content generation method shown in the first aspect of the embodiment of the present invention.

[0049] A fourth aspect of an embodiment of the present invention shows a storage medium, which includes a stored program. When the program runs, the device where the storage medium is located is controlled to execute the content generation method shown in the first aspect of the embodiment of the present invention.

[0050] Based on the above-mentioned embodiment of the present invention, a content generation method, device, electronic device and storage medium are provided. The method includes: obtaining hot events according to a preset time period, and the number of the hot events is multiple; performing multi-level risk factor identification on the hot events to determine the hot events containing risk factors; for each hot event predicted to have risk factors, calculating the public attention score of the hot event based on the sentiment vector and topic vector of the hot event; when it is determined that the public attention score of the hot event is greater than a preset score threshold, determining the product information of the hot event based on the risk information corresponding to the hot event; constructing corresponding copy content based on the title, document content, product information, and materials of the hot event, and the materials are collected based on the title and document content of the hot event. In the embodiment of the present invention, risk factors are identified for hot events, and the product information of the hot event is determined based on the public attention score calculated based on the sentiment vector and topic vector of the hot event, and then the corresponding copy content is generated. Therefore, the corresponding marketing content is generated accurately and quickly. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0052] Figure 1 A flow chart of a content generation method according to an embodiment of the present invention;

[0053] Figure 2 A schematic flow chart of another content generation method according to an embodiment of the present invention;

[0054] Figure 3 The figure is a schematic structural diagram of a content generation device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0056] The terms "first," "second," "third," "fourth," and the like (if any) in the specification and claims of this application and in the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or sequential sequence. It should be understood that the terms used in this manner are interchangeable where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusions, e.g., a process, method, system, product, or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0057] It should be noted that the descriptions of "first", "second", etc. in the present invention are for descriptive purposes only and should not be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" or "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but this must be based on the fact that they can be implemented by ordinary technicians in this field. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0058] In this application, the terms "comprises," "comprising," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not preclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0059] See also Figure 1 , is a flow chart of a content generation method according to an embodiment of the present invention, the method comprising:

[0060] Step S1: acquiring hot events according to a preset time period, wherein the number of hot events is multiple;

[0061] Among them, the capture of hot events involved in this application has been obtained with full consent and authorization, and the collection, use and processing of relevant information comply with the relevant laws, regulations and standards of relevant countries and regions.

[0062] In the specific implementation of step S1, an open source API interface for aggregating hot data is called according to a preset time period to crawl an event list from a preset news or social media application to capture the top 50 hot events.

[0063] It should be noted that the preset time period is set in advance by the computer personnel based on multiple experiments, and can generally be set to 30 minutes.

[0064] Step S2: Identify the multi-level risk factors of the hot events to determine the hot events containing risk factors;

[0065] In the process of implementing step S2, hot events are acquired according to a preset time period and stored in a hot event library. Then, multi-level risk factor identification is performed on the newly added hot events in the hot event library to determine whether the newly added hot events are hot events containing risk factors. If so, step S3 is executed; if not, no processing is performed temporarily.

[0066] Step S3: For each hot event predicted to have risk factors, a public attention score of the hot event is calculated based on the sentiment vector and topic vector of the hot event.

[0067] In the specific implementation of step S3, the emotion vectors and topic vectors of different dimensions are first calculated based on the title and content of the hot event; then the emotion vectors and topic vectors are processed to determine the public attention score of the hot event.

[0068] Step S4: Determine whether the public attention score of the hot event is greater than a preset score threshold. If it is greater than or equal to the threshold, execute step S5; if it is less than the threshold, discard it.

[0069] In the specific implementation of step S4, the hot event will enter the subsequent insurance product matching process only when the public attention score, that is, the marketing value score is greater than or equal to the preset score threshold; therefore, the public attention score of each hot event is compared to see whether it is greater than the preset score threshold. If it is greater than or equal to, step S5 is executed; if it is less than, it is discarded.

[0070] It should be noted that the preset score threshold is set based on multiple experiments or experience, and can generally be set to 6 points.

[0071] Step S5: Determine product information of the hot event based on the risk information corresponding to the hot event.

[0072] In the specific implementation of step S5, risk information is first extracted from the hot event, and structured information of implicit risks is determined by processing the risk information; then it is merged with the risk information to obtain a risk factor analysis result; then, product information matching the hot event is selected from the product information library based on the semantic matching score and keyword matching score calculated based on the risk factor analysis result.

[0073] It should be noted that the product information database is pre-built and stores product numbers, product names, product descriptions, and insurance types of different products.

[0074] It should be noted that the products here refer to insurance products.

[0075] Step S6: Construct corresponding copy content based on the title, document content, product information, and materials of the hot event.

[0076] In the specific implementation of step S6, first, collect the hot-spot related web pages and illustrations corresponding to the hot event and use them as the corresponding materials; then, retrieve the marketing content documents related to the hot event, that is, the real marketing copy; then generate the marketing content with both pictures and texts based on the marketing content documents, the title of the hot event, the document content, the product information, and the materials, and polish and proofread them line by line.

[0077] The material is collected based on the title of the hot event and the document content.

[0078] In an embodiment of the present invention, a real-time monitoring mechanism for hot events based on multi-source social media data is established; hot events are periodically captured through a timed scheduling strategy; risk factors are identified for hot events, and the product information of the hot events is determined based on the public attention score calculated based on the sentiment vector and topic vector of the hot events, thereby improving the timeliness and user attention relevance of insurance marketing content, and providing a new technical path for hot marketing in the insurance industry. Corresponding copywriting content is then generated. The traditional hot marketing work that takes several hours is compressed to the minute level, and the entire process from hot spot discovery to content publishing is automated, effectively solving the core pain points of insurance hot marketing, such as "slow capture of hot spots, difficulty in finding materials, and tiring copywriting", thereby accurately and quickly generating corresponding marketing content.

[0079] Optionally, based on the content generation method shown in the embodiment of the present invention, the specific implementation of step S2 of identifying multi-level risk factors for the hot events and determining the hot events containing risk factors includes the following steps:

[0080] Step S2.1: Split the title and text content of the hot event into text windows;

[0081] In the specific implementation of step S2.1, a rolling window method is used to process long texts. The titles and text contents of the newly added hot events are divided according to a window size of 128 tokens. Each text window overlaps 64 tokens to obtain multiple text windows.

[0082] Optionally, before executing step S2.1, the process further includes:

[0083] A heartbeat detection mechanism is used to monitor each hot event in the hot event library at preset time intervals to see if there is a field marked as "risk identification". If not, it means that the hot event is a newly added hot event; then the "risk identification" field is added to the hot event record and false is written; and step S2.1 is executed. Otherwise, the execution is returned to perform detection, and the hot event library is continued to be traversed at the next time interval.

[0084] It should be noted that the preset time interval is set in advance based on multiple experiments or experience, and can generally be set to 5 minutes.

[0085] Step S2.2: Call the preset classification model to perform classification prediction on each text window and obtain the prediction result;

[0086] It should be noted that the preset classification model is a binary classification model, which is pre-trained based on text windows with risk factors and text windows without risk factors, and is used to distinguish whether a text window has risk factors or not.

[0087] It should be noted that the forecast results include risk factors and risk-free factors.

[0088] In the specific implementation of step S2.2, the preset classification model is called to perform a binary classification prediction of risk factors (with risk factors or without risk factors) for each text window, and a prediction result containing each text window is obtained.

[0089] Step S2.3: Determine whether any text window has a prediction result that indicates a risk factor. If so, proceed to step S2.4. If not, do not process it for the time being.

[0090] In the specific implementation of step S2.3, if the prediction result of any window is determined to be "risk factor", the hot event is determined to contain risk factors, that is, step S2.3 is executed. Otherwise, no processing is performed temporarily.

[0091] Step S2.4: Determine whether the hotspot event contains risk factors.

[0092] In an embodiment of the present invention, each text window of a hot event is classified and predicted using a preset classification model. When the prediction result of any text window is that there is a risk factor, it is determined that the hot event contains a risk factor.

[0093] Optionally, based on the content generation method shown in the above-mentioned embodiment of the present invention, the specific implementation step S3 includes the following steps in the process of calculating the public attention score of each hot event whose prediction result is a risk factor based on the emotion vector and topic vector of the hot event.

[0094] Step S3.1: For each hot event predicted to have risk factors, sentiment analysis and topic dimension analysis are performed on the hot event to determine the corresponding sentiment vector and topic vector.

[0095] It should be noted that the specific implementation of step S3.1 includes the following steps:

[0096] Step S3.1.1: For each hot event predicted to have risk factors, sentiment analysis is performed on the title and text content of the hot event to obtain sentiment vectors of different dimensions.

[0097] It should be noted that the machine learning algorithm is pre-trained based on the titles and text contents of different historical hot events and their corresponding emotion categories, so as to construct a classification model that can analyze the titles and text contents of historical hot events and obtain the emotion probabilities of different emotion categories.

[0098] It should be noted that the emotion category includes emotions such as positive, angry, sad, fearful, surprised, and no emotion.

[0099] In the specific implementation of step 3.1.1, for each hot event predicted to have risk factors, that is, each hot event in the hot database, the pre-set emotion classification model is called to perform emotion analysis on the title and text content of the hot event to obtain the emotion probabilities of different emotion categories, that is, emotion vectors of different dimensions, and store them under the "emotion vector" field.

[0100] Step S3.1.2: For each hot event predicted to have risk factors, a topic dimension analysis is performed on the hot event based on the title and text content of the hot event to determine a topic vector.

[0101] In the specific implementation of step S3.1.2, for each hot event predicted to have risk factors, a pre-set unsupervised learning algorithm (Latent Dirichlet Allocation, LDA) topic model based on a probability generation model is called to perform topic dimension analysis on the title and text content of the hot event to determine the topic probability distribution vector of the preset dimension, that is, the topic vector, and then store this feature under the "topic vector" field.

[0102] It should be noted that the LDA topic model is based on the large-scale, high-quality news dataset open sourced by the Large Model Corpus Data Alliance, namely the Open News Archive dataset (OpenNewsArchive). The LDA topic model contains 100 predefined topic categories. By performing topic modeling on the text of hot events, a 100-dimensional topic probability distribution vector is extracted as the topic feature and stored in the "topic vector" field.

[0103] It should be noted that the preset dimension refers to 100 dimensions.

[0104] Step S3.2: Feature concatenation is performed on the emotion vectors of different dimensions and the topic vector to obtain an input vector.

[0105] In the specific implementation of step S3.2, the 6-dimensional emotional probability distribution vector of the hot event is and 100-dimensional topic probability distribution vector Perform feature concatenation to form a unified input representation of 106 dimensions , which is the input vector.

[0106] Step S3.3: Input the input vector into a pre-built value assessment model for processing to obtain a public attention score for the hot event.

[0107] In the specific implementation process of step S3.3, the pre-built value assessment model performs deep feature fusion on the input vector through a multi-layer fully connected neural network, and outputs an attention prediction score of 1-10 points, that is, the public attention score of the hot event.

[0108] Specifically, the pre-built value assessment model adopts a two-layer fully connected neural network architecture, and first performs a first-layer feature fusion on the input vector.

[0109] Input vector It is a 106-dimensional feature vector, which contains the combined representation of the emotion vector and the topic vector.

[0110] The input vector is transformed through the first layer to obtain a high-dimensional feature h, as shown in formula (1).

[0111] Formula (1):

[0112] .

[0113] in, represents the first layer weight matrix, represents a set of real numbers, d is the number of hidden layer neurons, and 106 is the input feature dimension, i.e. is a The real matrix of represents the first layer bias vector, express dimensional real vector space, that is is included ReLU is a rectified linear unit activation function, which is used to introduce nonlinearity and maintain gradient flow.

[0114] Then, the high-dimensional feature (Right now dimensional real vector) is transformed nonlinearly to capture the complex interaction pattern between emotions and topics. Then, the second layer of feature fusion is performed to compress the high-dimensional features is the scalar z, as shown in formula (2).

[0115] Formula (2):

[0116] .

[0117] in, represents the second layer weight matrix, express dimensional real matrix space, used to transform dimensional features are mapped to 1-dimensional output, represents the second layer bias term, represents the set of real numbers, i.e. is a single real number used to adjust the output baseline.

[0118] Finally, the scalar z is normalized by the Sigmoid function and linear mapping to obtain the public attention score of the hot event, which is then substituted into formula (3) to calculate the corresponding public attention score.

[0119] Formula (3):

[0120] .

[0121] in, For the Sigmoid function, the scalar (real number) is mapped to the (0,1) interval for normalization, and then transformed by linear transformation " ” It is scaled to a score range of 1-10 and outputs the predicted score of public attention of hot events.

[0122] It should be noted that the training process of the pre-built value assessment model includes:

[0123] The regularized mean square error loss function is adopted, as shown in formula (4), combined with the adaptive learning rate mechanism of the Adam optimizer and the back propagation algorithm for parameter optimization, and then as shown in formula (5).

[0124] Formula (4):

[0125] .

[0126] Formula (5):

[0127] .

[0128] End-to-end model training is achieved through formula (5).

[0129] Specifically, we used the sentiment vectors and topic vectors of different historical hot events as a dataset, and then divided the entire dataset into training set, validation set, and test set in a ratio of 7:2:1;

[0130] The training set is used to learn model parameters and update gradients through formulas (4) and (5), the validation set is used for hyperparameter tuning and early stopping mechanism monitoring, and the test set is used for final performance evaluation after model training is completed. Indicators such as are used to objectively measure the generalization ability and prediction accuracy of the model on unseen data, thereby establishing a reliable mapping relationship from multimodal features to public attention scores, namely the value assessment model.

[0131] Furthermore, it should be noted that the public attention score of the hot event, that is, the marketing value score is stored in the "public attention" score field.

[0132] In an embodiment of the present invention, the corresponding sentiment vector and topic vector are first calculated for each hot event predicted to have risk factors; then, the public attention score of the hot event is calculated based on the sentiment vector and topic vector of the hot event to determine whether each hot event has sufficient public attention.

[0133] Optionally, based on the content generation method shown in the above embodiment of the present invention, the specific implementation of step S5 in the process of determining the product information of the hot event based on the risk information corresponding to the hot event includes the following steps.

[0134] Step S5.1: Generate risk factor analysis results based on the risk information corresponding to the hot event.

[0135] It should be noted that the specific implementation of step S5.1 includes the following steps.

[0136] Step S5.1.1: For each hotspot event predicted to have risk factors, extract risk information from the hotspot event.

[0137] In the specific implementation of step S5.1.1, for each hot event predicted to have risk factors, the first layer of screening is used to select hot events containing risk factors. That is, the large language model constructed by the technology of efficient fine-tuning of large language model parameters through low-rank matrix decomposition (Low-Rank Adaptation, LoRA) performs formatted extraction of risk factors from the hot events to obtain the risk information screened out by the first layer.

[0138] It should be noted that the large language model built through LoRA is fine-tuned based on 10,000+ insurance risk annotation data. It can output risk information from hot event texts in a structured manner in Json format.

[0139] For example, the risk information of hot event A obtained through step S5.1.1 is {"involved population":["population category 1","population category 2"], "risk type":["risk type 1","risk type 2"]}.

[0140] Step S5.1.2: Process the risk information to obtain structured information of the implicit risk.

[0141] In the specific implementation of step S5.1.2, the prompt engineering in the large language model guides the large model to directly perform risk reasoning based on the identified risk information, and generates reasoning text consisting of possible implicit risks, associated risks and potential consequences. Then, the large language model constructed by LoRA is called again to perform structured extraction on the reasoning text to obtain the structured information of the implicit risks screened out in the second layer.

[0142] Step S5.1.3: Merge the risk information and the structured information of the implicit risk to obtain the risk factor analysis results.

[0143] In the specific implementation of step S5.1.3, the risk information obtained in step S5.1.1 and the structured information of the implicit risk obtained in step S5.1.2 are superimposed and spliced ​​to obtain the risk factor analysis result.

[0144] In the specific implementation of step S5.1.3, the risk information and the structured information of implicit risks, that is, the list values ​​of the same fields in the two Json structures are superimposed and spliced ​​to form a complete risk factor analysis result, and written into the "Risk Factor Analysis Result" field.

[0145] It should be noted that the risk factor analysis results can be {"involved population": [list of direct risk population + list of implicit risk population], "risk type": [list of direct risk type + list of implicit risk type]}.

[0146] For example, for the hot event "XXX suffered a sudden myocardial infarction, and the family members did not know the first aid measures and delayed treatment", the first-level rapid screening determined that it contained risk factors. The second-level structured extraction output {"involved population": ["XXX", "XXX family members"], "risk type": ["cardiovascular disease risk", "lack of first aid knowledge risk"]}. The third-level reasoning generated and extracted {"involved population": ["middle-aged and elderly groups", "family caregivers"], "risk type": ["family care ability risk", "medical expense burden risk"]}. The final merged result, that is, the risk factor analysis result is {"involved population": ["elderly people", "elderly family members", "middle-aged and elderly groups", "family caregivers"], "risk type": ["cardiovascular disease risk", "lack of first aid knowledge risk", "family care ability risk", "medical expense burden risk"]}.

[0147] Optionally, after completing the above analysis, the "Risk Identification" field of the data record is updated to true.

[0148] Step S5.2: Generate a hypothetical product description document based on the risk factor analysis results corresponding to the hot event.

[0149] The specific process of building a product description model includes: collecting risk factor analysis results of different hot events, and training and learning a variety of hypothetical description documents containing insurance products to build a product description model.

[0150] In the specific implementation of step S5.2, the pre-built product description model is called to analyze the risk factor analysis results to obtain a hypothetical product description document.

[0151] It should be noted that this hypothetical product description document can cover the insurance product characteristics that identify risks.

[0152] Step S5.3: Calculate the semantic matching score of the hypothetical product description document.

[0153] It should be noted that the first path matching strategy is a Hypothetical Document Embeddings (HyDE) semantic matching strategy.

[0154] In the specific implementation of step S5.3, the first path matching strategy is executed to convert the hypothetical document into a 1024-dimensional vector, i.e., a text vector. Then, the cosine similarity between the product description vector in the product information database and the text vector is calculated, as shown in formula (6), to obtain the corresponding HyDE semantic matching score .

[0155] Formula (6):

[0156] .

[0157] Among them, A is a 1024-dimensional vector, that is, a text vector; B is the description vector of each insurance product in the product information database.

[0158] Step S5.4: Determine product information of the hot event based on the risk factor analysis results and the semantic matching score.

[0159] It should be noted that the specific process of implementing step S5.4 includes the following steps.

[0160] Step S5.4.1: Extract query keywords from the risk factor analysis results corresponding to the hot event.

[0161] In the specific implementation of step S5.4.1, the involved population and risk type in the risk factor analysis results are extracted as query keywords.

[0162] Step S5.4.2: Determine a keyword matching score for each product from the product information database based on the query keyword.

[0163] It should be noted that the product information database is used to store the product information of all currently collected marketing products.

[0164] In the specific implementation of step S5.4.2, the ranking function algorithm is used to search the "target population label" and "product coverage" fields of the product information database respectively, and the Bool query clause "Should" of the ES database is used to implement multi-field search. That is, the "Should" clause follows the "or" logic. In other words, the BM25 relevance score corresponding to the keyword of each product is determined from the product information database based on the query keyword, and the score is substituted into formula (7) for calculation to obtain the keyword matching score of each product. .

[0165] Among them, the document can match any condition but not necessarily. The more matching conditions, the higher the score.

[0166] Formula (7):

[0167] .

[0168] in, Indicates the A Should clause condition, that is, a matching condition for a query field; Indicates the total number of Should clauses, that is, the total number of fields involved in the retrieval; Indicates a weighted summation of the scores of all Should clauses.

[0169] It is The BM25 relevance score corresponding to each Should clause. In this scenario, each Should clause corresponds to a match of a field (such as "target population label" or "product coverage"). It only represents the matching score of a single field, which is calculated based on word frequency, inverse document frequency, etc., and does not directly include cross-field combination logic.

[0170] It is The weight factors of the Should clauses are used to adjust the importance of different fields in the final score.

[0171] Coordination Factor It is pre-set and is used to reward documents that match multiple conditions at the same time, thereby achieving accurate product relevance ranking.

[0172] It should be further explained that for each Should clause, the corresponding BM25 relevance score is implemented as follows:

[0173] If one of the fields matched by the Should clause is "target population label" and "product coverage", first, the population label matching score is calculated based on the "target population label" matched by the Should clause. , as shown in formula (8).

[0174] Formula (8):

[0175] .

[0176] Next, the coverage matching score is calculated based on the "product coverage" matched by the Should clause. , as shown in formula (9).

[0177] Formula (9):

[0178] .

[0179] Then, and Perform weighted calculation to obtain the BM25 correlation score, where and Fields are given equal weight.

[0180] Step S5.4.3: Calculate based on the semantic matching score and the keyword matching score to obtain a comprehensive matching score.

[0181] In the specific implementation of step S5.4.3, the semantic matching score of the product description document and the keyword matching score are substituted into formula (10) for calculation to obtain a comprehensive matching score.

[0182] Formula (10):

[0183] Comprehensive matching score = a × semantic matching score + b × keyword matching score.

[0184] Among them, a is the weight of the semantic matching score, and b is the weight of the keyword matching score.

[0185] It should be noted that a and b are set by technical personnel based on actual conditions.

[0186] Step S5.4.4: Sort each product in the product information database in descending order according to the comprehensive matching score, and select the product information with the highest ranking.

[0187] It should be noted that product information includes product number, product name, product description, insurance type, and matching score.

[0188] In the specific implementation of step S5.4.4, all products are sorted in descending order according to the comprehensive scores, and the product information with the highest score is selected and written into the "Related Products" field.

[0189] In this embodiment of the present invention, the hot event is first processed to obtain risk information and structured information about the underlying risks. This information is then combined to generate a risk factor analysis result. Finally, the risk factor analysis result and the semantic matching score are used to determine product information related to the hot event. The risk factor analysis result and the semantic matching score of the hot event are then used to determine relevant insurance product information, allowing for the subsequent generation of new marketing content solutions based on the product information and the hot event.

[0190] Optionally, based on the content generation method shown in the embodiment of the present invention, the specific implementation of step S6 is to construct corresponding copy content based on the title of the hot event, document content, product information, and materials, including the following steps:

[0191] Step S6.1: Collect corresponding materials based on the title of the hot event and the document content.

[0192] Among them, each hot event corresponds to a piece of material.

[0193] It should be noted that the specific implementation of step S6.1 includes the following steps.

[0194] Step S6.1.1: Within a preset time period, search based on the title of the hot event to determine relevant web page fields.

[0195] Optionally, a multi-source directional search engine aggregation API is deployed within the search engine framework. This API can call multiple mainstream search engine interfaces at the same time and provide more comprehensive search coverage through deduplication algorithms and result fusion strategies.

[0196] It should be noted that the preset time period refers to the period from when a hot event in the hot event database is detected to contain a risk factor field to a certain period of time.

[0197] The certain period of time is set in advance based on experience, and can generally be set to 2 days.

[0198] In the specific implementation of step S6.1.1, first, determine whether the "Risk Identification" field in the hotspot database monitored at the first time interval is true, whether the "Risk Factor Analysis Result" field of the hotspot event is not null, and whether the "Material Collection Completed" field does not exist;

[0199] It should be noted that the first time interval is set in advance by technicians based on multiple experiments, and can generally be set to 3 minutes.

[0200] If the "Risk Identification" field is true, the "Risk Factor Analysis Result" field of the hot event is not null, and the "Material Collection Completed" field does not exist, add the "Material Collection Completed" field and write false. This means that when it is detected that a hot event in the hot event database contains a risk factor field, the targeted collection and multi-dimensional screening of related materials are triggered to provide high-quality background information for subsequent content generation.

[0201] Next, we process the records of hot events that contain risk factors and for which no materials have been collected. We use the titles of hot events as search terms and perform parallel searches through the search engine API to determine the top 20 search results related to the search terms. After deduplication, we retain the titles, links, and summary information of all web pages, i.e., the initial web pages.

[0202] Then, a two-level intelligent screening algorithm is used to filter out relevant web page fields from the initial web pages.

[0203] Specifically, the first layer encodes the vectors of the webpage title and the hot event title of each initial webpage based on the preset embedding model, and calculates their cosine similarity;

[0204] It should be noted that the preset embedding model here is used to store the cosine similarity algorithm for calculating the cosine similarity between the webpage title of the initial webpage field and the vector of the hot event title.

[0205] For each initial web page in the second layer, the preset large model is called to process the summary information of the initial web page to obtain the corresponding keyword density and information completeness; then the text length in the initial web page is counted; finally, the preset large model is called again to perform quality scoring based on the text length, keyword density, and information completeness in the initial web page.

[0206] It should be noted that the preset large model is pre-trained in the first layer based on the summary information of different initial web pages, as well as the corresponding keyword density and information completeness; after the first layer of training is completed, training is carried out based on information such as the internal text length, keyword density, information completeness, and corresponding quality scores of different initial web pages to construct a preset large model that can perform quality scores on initial web pages. In other words, the preset large model is provided with information such as the internal text length, keyword density, information completeness, and the quality scores of the summary information of different historical initial web pages.

[0207] After two levels of screening, for each initial web page, the product of the cosine similarity and the first weight, as well as the product of the quality score and the second weight are calculated and added together to obtain a comprehensive score; the initial web pages are sorted from large to small according to the comprehensive score, and the web page titles and links of a preset number of initial web pages are selected and stored in the relevant web page field to obtain the relevant web page field.

[0208] It should be noted that the first weight and the second weight are both set in advance by technicians based on multiple experiments or experience. The first weight can be set to 0.6, and the second weight can be set to 0.4.

[0209] The preset number is set in advance by technicians based on multiple experiments and is generally set to 10.

[0210] Step S6.1.2: For the hot event records existing in the relevant web page fields, start parallel web page content crawling to determine the materials of the hot event.

[0211] This step uses an asynchronous request framework to concurrently request all web links in the relevant web pages, and sets the request timeout to 10 seconds.

[0212] For the response content of each web page, the web page body content is extracted through the tag selector, filtering out non-content tags such as script, language specification style, function nav, etc.

[0213] Extract high-definition image links from web pages, filter thumbnails smaller than 300x300 pixels, and store them;

[0214] A BERT-based text summarization model is used to intelligently summarize the extracted long texts, compressing articles over 1,000 words into a core summary of 200-300 words.

[0215] The parsing results are stored in the standardized JSON format: {"texts":[{"url":"webpage link","title":"webpage title","content":"text content","extract_time":"extraction time"}],"images":[{"url":["image link"]}]}, and stored in the "Material" field. After all webpage content is crawled and parsed, this step updates the "Material Collection Completed" field to true, indicating that the material collection process for this hot event record has been completed.

[0216] Step S6.2: Determine a preset number of marketing content documents based on a preset dynamic creation example library and the semantic similarity of the titles of the hot events.

[0217] In the specific implementation of step S6.2, for each marketing content document in the dynamic creation example library, the semantic similarity between the hot event title and the example title in the marketing content document is calculated; a weighted sum is performed based on the semantic similarity, the number of readings and likes of the hot event and other interactive data to obtain the corresponding score, and the scores are sorted, and a preset number of marketing copywriting with the highest scores are selected, that is, the marketing content document is used as a creation reference template, and is stored in the "Creation Reference Example" field of the hot event record.

[0218] It should be noted that the weights corresponding to interactive data such as semantic similarity, number of readings and likes of hot events are all set by technical personnel based on actual conditions.

[0219] The weight of semantic similarity is greater than the weight of interactive data such as the number of readings and likes of hot events.

[0220] It should be noted that the preset number is set in advance by technical personnel based on actual conditions and can generally be set to 5.

[0221] The preset dynamic creation example library continuously captures insurance marketing content in various media applications through automated extraction, and extracts corresponding basic information; then, the insurance marketing content is stored based on the basic information.

[0222] Basic information includes example number, title, content, insurance type, number of reads, number of likes, number of reposts, number of favorites, and number of comments, etc.

[0223] It is further worth noting that the sample library adopts an incremental update mechanism, automatically crawling the latest marketing content and performing quality screening every day.

[0224] Optionally, before executing step S6.2, the following steps may be further included:

[0225] A heartbeat detection mechanism is used to monitor event records of completed material collection and yet to generate content at a second time interval.

[0226] Specifically, determine whether the "Material Collection Completed" field in the hot event database is true and whether the "Marketing Copy" field does not exist. If it is detected that the "Material Collection Completed" field is true and the "Marketing Copy" field does not exist, create a "Marketing Copy Generated" field and write false to it, then execute step S6.2.

[0227] It should be noted that the second time interval is set in advance by technicians based on multiple experiments, and can generally be set to 2 minutes.

[0228] Step S6.3: Deconstruct the reasoning path of the preset number of marketing content documents to obtain corresponding creation prompt words;

[0229] In the specific implementation of step S6.3, LoRA fine-tuning training is performed based on a pre-built large language model, that is, the "creation reference examples", that is, the potential creative thinking chains of a preset number of marketing content documents are deconstructed through the reverse thinking chain distiller. The distiller can automatically analyze the structural pattern of each marketing content document (such as: opening introduction, pain point mining, product placement, and call to action), rhetorical techniques (such as: metaphor, parallelism, and rhetorical questions, etc.) and emotional mobilization strategies, reversely extract reusable marketing copywriting creative thinking chains, and convert them into structured creative prompts.

[0230] Step S6.4: Processing the creative prompt words, the title and text content of the hot event, the materials, and the product information to generate corresponding copy content;

[0231] In the specific implementation of step S6.4, the pre-built copy content generation model is called, and the "Title" and "Document Content" fields of the hot database are used as the basic information of the hot event, the high-quality materials in the "Material" field are used as supplementary information, and the "Related Products" field is used as product information. The distilled thought chain, that is, the creation prompt words, is used as the creation guidance framework, and personalized marketing copy, that is, copy content, is generated through preset templates.

[0232] It should be noted that the pre-built copy content generation model is pre-based on the creation prompt words of different historical hot data, the titles and text content of the hot events, materials, product information, and preset templates as training data to train the copy content generation model, so that the trained copy content generation model can be processed based on the creation prompt words, the titles and text content of the hot events, materials, and product information to generate corresponding copy content.

[0233] Step S6.5: Select a corresponding picture from the material as an accompanying picture for the copy content.

[0234] In the specific implementation of step S6.5, a multimodal large language model is used to analyze the content theme, emotional tone and product features of the complete copy, and the four most matching pictures are intelligently selected from all the picture materials in the "Material" field as accompanying pictures.

[0235] It should be noted that the copy and accompanying pictures are stored in JSON format, specifically {"text_content":"Marketing copy text","selected_images":[{"image_url":"Select image link","match_reason":"Match reason"}],"compliance_check":"Compliance check result","quality_score":"Quality score"} are stored in the "Marketing copy" field, and the "Marketing copy generated" field is updated to true, indicating that the content generation process of the event has been completed.

[0236] This embodiment of the present invention provides a new technical approach for hot marketing in the insurance industry. First, materials related to hot events containing risk factor fields are collected. Next, a preset number of marketing content documents with the highest semantic similarity to the titles of the hot events are determined from a preset dynamic creation example library. These documents are then deconstructed using reasoning paths to obtain corresponding creative prompts, which are then used to generate corresponding copywriting content. This allows for accurate and rapid generation of relevant marketing content.

[0237] Optionally, based on the content generation method shown in the above embodiment of the present invention, after executing step S6.5, the method further includes:

[0238] Step S6.6: Determine whether there is a description in the copy content that is consistent with the product information of the hot event. If not, execute step S6.7; if so, execute step S6.8.

[0239] In the specific implementation of step S6.6, first check the product consistency of the insurance products in the copy content, split the copy according to the paragraph granularity and analyze it, and verify whether the product description in each paragraph is consistent with the product information description in the "Related Products" field. If not, execute step S6.7; if so, execute step S6.8.

[0240] Optionally, step S6.6 can be implemented using a trained large language model.

[0241] Step S6.7: Modify the product description in the copy based on the product information of the hot event.

[0242] In the specific implementation of step S6.7, the product description in the copy is automatically updated using the product information of the hot event to obtain a new copy.

[0243] Step S6.8: Correct the content in the document that poses compliance risks.

[0244] In the specific implementation of step S6.8, a large language model is used to identify compliance risks in each paragraph of the copy. If non-compliant content is found, the corresponding content description is revised based on regulations such as the Advertising Law and the Insurance Law.

[0245] It should be noted that the large language model here is constructed based on laws and regulations such as the Advertising Law and the Insurance Law, various paragraphs of text, and content marked as having compliance risk identification.

[0246] Step S6.9: Optimize the text.

[0247] In the specific implementation of step S6.9, a pre-trained large language model is used to analyze the language fluency, logical coherence, and emotional appeal of each paragraph, and intelligently polish and reconstruct paragraphs that do not meet the quality standards to optimize the copy.

[0248] In this application, inaccurate content is automatically rewritten through product description consistency verification to ensure that the marketing copy is both timely and industry-specific, achieving a dual improvement in content quality and generation efficiency.

[0249] Optionally, based on the method shown in the above embodiment of the present invention, the embodiment of the present invention also shows a flow chart of another content generation method, combined with Figure 1 , see Figure 2 In step S1, after obtaining hot events according to a preset time period, where the number of hot events is multiple, the method further includes:

[0250] Step S7: filtering hot events based on the hash value of the title text corresponding to each hot event.

[0251] In the process of implementing step S7, for each hot event, the corresponding hash value is calculated based on the title text of the hot event; the hot events are filtered based on the hash value of the title text corresponding to each hot event; the fields corresponding to the remaining hot events are written into the hot event library; and step S2 is executed for each hot event in the hot event library.

[0252] It should be noted that, for each hot event, the process of calculating the corresponding hash value based on the title text of the hot event includes:

[0253] For each hot event, the title text of the hot event is first identified; then, a cryptographic hash function is called to process the title text to generate a hash value of fixed bytes, that is, a hash value of fixed length.

[0254] Optionally, the cryptographic hash function may be the sha256 algorithm.

[0255] It should be noted that the process of filtering hot events based on the hash value of the title text corresponding to each hot event includes:

[0256] Compare the hash values ​​of the title texts corresponding to each of the hot events. If the same hash values ​​exist, only retain the hot event corresponding to one hash value and remove the other hot events, that is, exclude the hot events with the same title and only retain one, thereby ensuring that there are no identical hot event records in the database.

[0257] It should be noted that the process of writing the remaining fields corresponding to the hot events into the hot event library includes:

[0258] Extract and filter the remaining hot events' titles, links, content texts and other fields, and write them into the hot events database.

[0259] After step S7 is executed, steps S2 to S6 are continued to be executed.

[0260] In an embodiment of the present invention, a real-time monitoring mechanism for hot events based on multi-source social media data is implemented. Hot events are periodically captured using a timed scheduling strategy. Duplicate hot events are removed using their hash values. Risk factors are then identified for the remaining hot events. Product information for these hot events is then determined based on the public attention scores calculated based on their sentiment and topic vectors. This improves the timeliness and relevance of insurance marketing content, providing a new technical approach for hot marketing in the insurance industry. Corresponding copywriting is then generated, resulting in accurate and rapid generation of relevant marketing content.

[0261] Based on the content generation method shown in the above embodiment of the present invention, the embodiment of the present invention also shows a content generation device, such as Figure 3 As shown, the device includes:

[0262] An acquisition unit 1 is configured to acquire a plurality of hot events according to a preset time period;

[0263] Identification unit 2, used to identify multi-level risk factors of the hot events and determine the hot events containing risk factors;

[0264] Processing unit 3, configured to calculate, for each hot event predicted to have a risk factor, a public attention score of the hot event based on the sentiment vector and topic vector of the hot event;

[0265] Determining unit 4, configured to determine product information of the hot event based on risk information corresponding to the hot event when determining that the public attention score of the hot event is greater than a preset score threshold;

[0266] The generating unit 5 is used to construct corresponding copy content based on the title of the hot event, document content, product information, and materials, where the materials are collected based on the title of the hot event and document content.

[0267] The specific principles and execution processes of each unit in the content generation device disclosed in the above embodiment of the present invention are the same as the corresponding contents in the content generation method provided in the above embodiment of the present invention. Please refer to the corresponding parts of the content generation method disclosed in the above embodiment of the present invention, and no further details will be given here.

[0268] In this embodiment of the present invention, a real-time monitoring mechanism for hot events based on multi-source social media data is implemented. Hot events are periodically captured through a timed scheduling strategy. Risk factors for these hot events are identified, and product information for these hot events is determined based on a public attention score calculated based on the sentiment and topic vectors of the hot events. This improves the timeliness and relevance of insurance marketing content to user attention, providing a new technical approach for hot event marketing in the insurance industry. This in turn generates corresponding copywriting content, resulting in accurate and rapid generation of relevant marketing content.

[0269] Optionally, based on the content generation device shown in the above embodiment of the present invention, the identification unit 2 is specifically configured to:

[0270] Splitting the title and text content of the hot event into text windows;

[0271] Call the preset classification model to perform classification prediction on each text window and obtain the prediction result;

[0272] Determine whether the prediction result of any text window is a risk factor;

[0273] If so, it is determined that the hot event contains risk factors.

[0274] Optionally, based on the content generation device shown in the above embodiment of the present invention, the processing unit 3 is specifically configured to:

[0275] For each hot event predicted to have risk factors, sentiment analysis and topic dimension analysis are performed on the hot event to determine the corresponding sentiment vector and topic vector;

[0276] Perform feature concatenation on the emotion vectors of different dimensions and the topic vector to obtain an input vector;

[0277] The input vector is input into a pre-built value assessment model for processing to obtain a public attention score for the hot event.

[0278] Optionally, based on the content generation device shown in the above embodiment of the present invention, the determination unit 4 is specifically configured to:

[0279] Generate risk factor analysis results based on the risk information corresponding to the hot event;

[0280] Generate a hypothetical product description document based on the risk factor analysis results corresponding to the hot event;

[0281] Calculating a semantic matching score for the hypothetical product description document;

[0282] The product information of the hot event is determined based on the risk factor analysis results and the semantic matching score.

[0283] The risk factor analysis results are generated based on the risk information corresponding to the hot event, including:

[0284] For each hot event predicted to have risk factors, extract risk information from the hot event;

[0285] Processing the risk information to obtain structured information of the implicit risk;

[0286] The risk information and the structured information of the implicit risk are combined to obtain the risk factor analysis results.

[0287] The process of determining the product information of the hot event based on the risk factor analysis result and the semantic matching score includes:

[0288] Extracting query keywords from the risk factor analysis results corresponding to the hot events;

[0289] Determining a keyword matching score for each product from a product information database based on the query keyword;

[0290] Calculating based on the semantic matching score and the keyword matching score to obtain a comprehensive matching score;

[0291] Each product in the product information database is sorted in descending order according to the comprehensive matching score, and the product information that ranks first is selected.

[0292] Optionally, based on the content generation device shown in the above embodiment of the present invention, the generation unit 5 is specifically configured to:

[0293] Collect corresponding materials based on the title and document content of the hot event, wherein each hot event corresponds to a piece of material;

[0294] Determining a preset number of marketing content documents based on a preset dynamic creation example library and semantic similarity of titles of the hot events;

[0295] Performing reasoning path deconstruction on the preset number of marketing content documents to obtain corresponding creation prompt words;

[0296] Based on the creation prompt words, the title and text content of the hot event, materials, and product information, corresponding copywriting content is generated;

[0297] Select corresponding pictures from the materials as illustrations for the copy content.

[0298] An embodiment of the present application provides an electronic device, which includes a processor and a memory, wherein the memory is used to store content generation control program code and data, and the processor is used to call program instructions in the memory to execute the steps shown in the content generation method in the above embodiment.

[0299] An embodiment of the present invention provides a storage medium, which includes the electronic device provided by the above-mentioned embodiment of the present application, and the electronic device is used to execute the content generation method disclosed in the embodiment of the present application.

[0300] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple. For relevant parts, refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0301] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0302] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A content generation method, characterized in that: The method comprises: Acquire hot events according to a preset time period, where the number of hot events is multiple; Conduct multi-level risk factor identification on the hot events to determine the hot events containing risk factors; For each hot event predicted to have risk factors, calculate the public attention score of the hot event based on the sentiment vector and topic vector of the hot event; When it is determined that the public attention score of the hot event is greater than a preset score threshold, determining product information of the hot event based on the risk information corresponding to the hot event; Constructing corresponding copywriting content based on the title of the hot event, document content, product information, and materials, where the materials are collected based on the title of the hot event and document content; Wherein, determining the product information of the hot event based on the risk information corresponding to the hot event includes: Generate risk factor analysis results based on the risk information corresponding to the hot event; Generate a hypothetical product description document based on the risk factor analysis results corresponding to the hot event; Calculating a semantic matching score for the hypothetical product description document; The product information of the hot event is determined based on the risk factor analysis result and the semantic matching score, wherein the product refers to an insurance product.

2. The method according to claim 1, characterized in that Conduct multi-level risk factor identification for the hot events and determine the hot events containing risk factors, including: Splitting the title and text content of the hot event into text windows; Call the preset classification model to perform classification prediction on each text window and obtain the prediction result; Determine whether the prediction result of any text window is a risk factor; If so, it is determined that the hot event contains risk factors.

3. The method according to claim 1, characterized in that For each hot event predicted to have a risk factor, a public attention score of the hot event is calculated based on the sentiment vector and topic vector of the hot event, including: For each hot event predicted to have risk factors, sentiment analysis and topic dimension analysis are performed on the hot event to determine the corresponding sentiment vector and topic vector; Perform feature concatenation on the emotion vectors of different dimensions and the topic vector to obtain an input vector; The input vector is input into a pre-built value assessment model for processing to obtain a public attention score for the hot event.

4. The method according to claim 1, wherein Generate risk factor analysis results based on the risk information corresponding to the hot event, including: For each hot event predicted to have risk factors, extract risk information from the hot event; Processing the risk information to obtain structured information of the implicit risk; The risk information and the structured information of the implicit risk are combined to obtain the risk factor analysis results.

5. The method according to claim 1, wherein Determining product information of the hot event based on the risk factor analysis results and the semantic matching score includes: Extracting query keywords from the risk factor analysis results corresponding to the hot events; Determining a keyword matching score for each product from a product information database based on the query keyword; Calculating based on the semantic matching score and the keyword matching score to obtain a comprehensive matching score; Each product in the product information database is sorted in descending order according to the comprehensive matching score, and the product information that ranks first is selected.

6. The method according to claim 1, characterized in that Construct corresponding copywriting content based on the title of the hot event, document content, product information, and materials, including: Collect corresponding materials based on the title and document content of the hot event, wherein each hot event corresponds to a piece of material; Determining a preset number of marketing content documents based on a preset dynamic creation example library and semantic similarity of titles of the hot events; Performing reasoning path deconstruction on the preset number of marketing content documents to obtain corresponding creation prompt words; Based on the creation prompt words, the title and text content of the hot event, materials, and product information, corresponding copywriting content is generated; Select corresponding pictures from the materials as illustrations for the copy content.

7. A content generating device, characterized in that: The device comprises: An acquisition unit, configured to acquire a plurality of hot events according to a preset time period; an identification unit, configured to identify multi-level risk factors of the hot events and determine the hot events containing risk factors; a processing unit configured to calculate, for each hot event predicted to have a risk factor, a public attention score of the hot event based on the sentiment vector and the topic vector of the hot event; a determining unit, configured to determine product information of the hot event based on risk information corresponding to the hot event when it is determined that the public attention score of the hot event is greater than a preset score threshold; A generating unit, configured to construct corresponding copy content based on the title of the hot event, document content, product information, and materials, wherein the materials are collected based on the title of the hot event and the document content; The unit for determining the product information of the hot event based on the risk information corresponding to the hot event is specifically configured to: Generate risk factor analysis results based on the risk information corresponding to the hot event; Generate a hypothetical product description document based on the risk factor analysis results corresponding to the hot event; Calculating a semantic matching score for the hypothetical product description document; The product information of the hot event is determined based on the risk factor analysis result and the semantic matching score, wherein the product refers to an insurance product.

8. An electronic device, characterized in that: The electronic device includes a processor and a memory, the memory is used to store program code and data for content generation, and the processor is used to call program instructions in the memory to execute the content generation method according to any one of claims 1 to 6.

9. A storage medium, characterized in that: The storage medium includes a stored program, wherein when the program is executed, the device where the storage medium is located is controlled to execute the content generation method according to any one of claims 1 to 6.

Images