Network guidance content automatic generation and distribution method based on large model

Abstract

The invention provides a method for automatically generating and distributing network guidance content based on a large model, which comprises the following steps of: 1, data monitoring: collecting related data of a monitoring target in the Internet; 2, analyzing a hot topic, processing the related data collected in the step 1, and obtaining the topic of the hot event and the corresponding data content; step 3, guide content generation: setting a caliber requirement of the guide content according to the theme of the hot event obtained in the step 2, and generating the guide content through a large language model; and 4, task distribution management: selecting a distribution method and a distribution strategy according to the guide content obtained in the step 3, and carrying out guide content distribution.

Description

TECHNICAL FIELD

[0001] The present application relates to a content automatic generation and distribution method, in particular to a large model-based network guide content automatic generation and distribution method. BACKGROUND

[0002] The information provided in this section is merely background information related to the present disclosure and can not necessarily be prior art.

[0003] With the rapid development of the Internet, network public opinion has become an important factor affecting the image and reputation of various institutions. However, the traditional public opinion information processing method has the problems of incomplete information collection and inaccurate analysis, which is difficult to meet the current needs of public opinion management.

[0004] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a large model-based network guide content automatic generation and distribution method to solve the problems of the prior art.

[0006] In order to solve the above technical problems, the present application discloses a large model-based network guide content automatic generation and distribution method, comprising the following steps:

[0007] Step 1, data monitoring, collecting relevant data of monitoring targets in the Internet;

[0008] Step 2, hot topic analysis, processing the relevant data collected in step 1 to obtain the theme of the hot event and the corresponding data content;

[0009] Step 3, guide content generation, setting the caliber requirements of the guide content according to the theme of the hot event obtained in step 2, and generating the guide content through a large language model;

[0010] Step 4, task distribution management, selecting a distribution method and strategy according to the guide content obtained in step 3, and distributing the guide content.

[0011] Further, the data monitoring of step 1 comprises:

[0012] Step 1-1, according to the set keywords, perform synonym evolution to obtain a dictionary composed of a domain keyword table and an exclusion word table of the target domain;

[0013] Step 1-2, according to the dictionary obtained in step 1-1, collect, store and preprocess the Internet data to obtain relevant data of the monitoring target; the relevant data is the data content corresponding to the set keywords.

[0014] Further, the hot topic analysis of step 2 comprises:

[0015] Step 2-1, based on the prompt word, using a large model to generate the topic and abstract text of each hot event for the relevant data collected in step 1;

[0016] Step 2-2: According to the topic and abstract text of the hot event, encode the abstract text based on the Q3 vector model, calculate the cosine similarity, and judge the similarity according to the cosine similarity, judge whether there is a similar hot event in the pre-set event graph library, if yes, execute step 2-3, otherwise execute step 2-4;

[0017] Step 2-3, updating the event graph of the similar hot event;

[0018] Step 2-4, constructing a new event graph of a hot event and adding it to the event graph library.

[0019] Further, the updating of the event graph of the similar hot event in step 2-3 comprises:

[0020] Step 2-3-1, using a large model to extract the relevant information of the hot event, the relevant information including: entity, relationship, attribute and event architecture, wherein the event architecture includes event subject, object and trigger word;

[0021] Step 2-3-2, based on the relevant information collected in step 2-3-1, updating the event graph of the similar hot event in the event graph library;

[0022] Step 2-3-3, calculating and updating the attributes of the similar hot event in the graph library, the attributes including: propagation breadth, propagation depth and emotional intensity.

[0023] Further, the construction of the new event graph of the hot event in step 2-4 comprises:

[0024] Step 2-4-1, using a large model to extract the relevant information of the hot event, the relevant information including: entity, relationship, attribute and event architecture, wherein the event architecture includes event subject, object and trigger word;

[0025] Step 2-4-2, based on the relevant information collected in step 2-4-1, constructing an event graph;

[0026] Step 2-4-3, calculate the attributes of the hot event in the event graph constructed in step 2-4-2, including: propagation breadth, propagation depth and sentiment intensity.

[0027] Further, the propagation breadth in step 2-3-3 and step 2-4-3, the calculation method is as follows:

[0028]

[0029] Where, B(e) is the propagation breadth, max represents the maximum value, ω is the amplification factor, U active (e) is the number of active users, the calculation method is as follows:

[0030]

[0031] Where, w i is the core user weight of the i-th user, U weight (e) is the space-time weighted user number, the calculation method is as follows:

[0032]

[0033] Where, α geo (d i ) is the geographical attenuation factor, α time (t i ) is the time attenuation factor, U(e) is the actual number of users, which is obtained according to different network platforms through preset rules.

[0034] Further, the propagation depth in step 2-3-3 and step 2-4-3, the calculation method is as follows:

[0035] Build the propagation chain topology, get the forwarding path of the hot event, use F k (e) to represent the forwarding volume of the hot event on the k-th layer forwarding path, then the calculation method of the propagation depth D(e) is as follows:

[0036]

[0037] Where, K is the number of forwarding path layers, I k,j is the user influence index.

[0038] Further, the sentiment intensity in step 2-3-3 and step 2-4-3, the calculation method is as follows:

[0039] Use natural language processing tools to perform sentiment analysis on the hot event, get the attributes of each sub-topic e m , including: positive value pos(e m ) and negative value neg(e m );

[0040] The emotional intensity S(e) is calculated as follows:

[0041] S(e) = |max(pos time (e))-max(neg time (e))| x (1+σ pos +σ neg )

[0042] Wherein, σ pos and σ neg are the standard deviations of positive and negative values, pos time (e) represents the positive emotional timeliness correction value, neg time (e) represents the negative emotional timeliness correction value, and the calculation method is as follows:

[0043] pos time (e) = τ(t m ) x pos(e m )

[0044] neg time (e) = τ(t m ) x neg(e m )

[0045] Wherein, τ(t m ) represents the time weight.

[0046] Further, the guided content generation of step 3 comprises:

[0047] Step 3-1, select a large language model according to the application scenario, and fine-tune according to the application field;

[0048] Step 3-2, according to the business requirements of guided content generation, build style knowledge base, dialect rule base and tone rule base;

[0049] Step 3-3, according to the theme of the hot event obtained in step 2, use vector retrieval and full-text retrieval in the style knowledge base, dialect rule base and tone rule base;

[0050] Step 3-4, fuse the retrieval results by reverse ranking fusion method to obtain the guided content.

[0051] Further, the task distribution management of step 4 comprises:

[0052] Step 4-1, set the guided content distribution mechanism, including: manual control distribution process and intelligent control distribution process;

[0053] Step 4-2, set the number of distributed tasks;

[0054] Step 4-3, select a task distribution mechanism for distribution, if a manual control distribution process is selected, step 4-4 is executed, if an intelligent control distribution process is selected, step 4-5 is executed;

[0055] Step 4-4, manually execute according to the distribution task requirements and operation steps, including: selecting the number of tasks, task execution mode, task distribution strategy and task effective time;

[0056] Step 4-5, automatically execute according to the distribution task quantity and the parameters set;

[0057] Step 4-6, after the distribution task is completed, the execution result feedback is received, and for the distribution task that has not been completed within the effective time, manual selection is required to resend or not to send.

[0058] Advantages:

[0059] 1. The present application realizes data collection of network public opinion, hot event analysis of topic model, large model intelligent guidance content generation and guided content task distribution management.

[0060] 2. Compared with the generation algorithm of traditional machine learning and deep learning, the large language model has more advantages in semantic understanding and logical generation.

[0061] 3. The task distribution management process of the present application is more intelligent and fast. DETAILED DESCRIPTION

[0062] The above and / or other aspects of the present application will become apparent from the following detailed description of the application taken in conjunction with the accompanying drawings.

[0063] Figure 1 The present application is a whole flow chart.

[0064] Figure 2 The data monitoring module flow chart.

[0065] Figure 3 The hot event analysis module flow chart of the event graph analysis model.

[0066] Figure 4 The intelligent guided content generation module flow chart of the optimized large language model.

[0067] Figure 5 The task distribution management module flow chart. DETAILED DESCRIPTION

[0068] The application provides a large model-based network guide content automatic generation and distribution method based on big data and intelligent analysis algorithm.

[0069] The hotspot event analysis based on the event graph is data analysis through big data processing and event graph generation and retrieval model, wherein the big data processing includes storage and framework algorithm, and the top hot events in a period of time are obtained through the text event graph analysis model.

[0070] The intelligent guide content generation based on the optimized large language model is to process the expected results through prompt word engineering optimization and RAG technology based on retrieval enhancement generation, take the hot topic content in the monitoring period as an external knowledge base, optimize the guide content generation effect of the large language model, and generate the guide content meeting the expectation.

[0071] The task distribution management intelligently distributes the generated guide content according to the task form and requirement. Generally, it includes manual distribution and automatic distribution, and a certain distribution strategy is set, including whether to randomly allocate and whether to repeatedly distribute. With the rise of internet video live broadcast and the rapid development of artificial intelligence, the execution subject is gradually changed from the original manual to robot execution, network big V and celebrity driving execution and other execution modes. The whole process tends to be intelligent and fast.

[0072] The application provides a large model-based network guide content automatic generation and distribution method, which includes the following four modules:

[0073] (1) Data monitoring module: collect the relevant data of the concerned subject;

[0074] (2) Hot event analysis module based on the event graph analysis model: obtain the hot topic and the data content corresponding to the hot topic.

[0075] (3) Intelligent guide content generation module: complete the intelligent generation of network guide content of specific public opinion hot events.

[0076] (4) Task distribution management module: in the guiding content distribution link, first face mechanism choice, one is manual control distribution process, user determines task quantity, distribution mode (covering manual, network robot, celebrity or big V, etc.), distribution strategy (such as random and whether corpus can be repeated) and task effective time according to actual task condition, two is intelligent control distribution process based on setting program and parameter, program automatically plans according to task quantity and preset parameter, like automatic selection distribution mode, strategy, prediction execution time. After task completion, each execution module receives feedback, and according to manual instruction, it is decided whether to resend for the task not completed in overtime.

[0077] The hot event analysis method based on the event graph analysis model disclosed by the application is data analysis through big data processing and event graph analysis model, processes massive internet data through the event graph model, analyzes the top-ranked topics and corresponding keywords in the time period, and retrieves the corresponding hot topic content in the stored data.

[0078] The intelligent guiding content generation method based on the optimized large language model disclosed by the application is to process the expected results of intelligent early warning through prompt word engineering optimization and RAG technology based on retrieval enhancement generation, optimize the large language model to improve the question and answer effect of whether the to-be-tested text is the predicted content, and screen out the expected intelligent early warning content. Compared with the early warning algorithm of traditional machine learning and deep learning, the large language model has more advantages in semantic understanding and judgment.

[0079] The task distribution management method based on the optimized large language model disclosed by the application mainly includes two kinds of control mechanisms. One is a manual control distribution mechanism, and the operator needs to set the specific number of tasks according to the actual demand of the task, select the execution subject of task distribution, and needs to rely on manual implementation, use network robots for batch operation, or rely on the influence of celebrities or big Vs. At the same time, the distribution strategy cannot be underestimated, and needs to be weighed in multiple modes such as random allocation and whether the corpus can be repeated, and the effective duration of the task is limited. The second is to realize intelligent control distribution relying on the set program and related parameters. In this mode, the program will determine the most suitable task distribution mode based on the task quantity and the pre-set parameters, plan the distribution strategy, and predict the execution time node, so that the whole distribution process can proceed in an orderly manner. After the execution of the distribution task is completed, each execution module enters a waiting state and receives execution result feedback information. If a task cannot be successfully completed within the specified effective time, the system will start the emergency handling mechanism according to the instructions given by manual operation, or require to resend the task to give a second chance to make up for the omissions, or directly judge not to send, to avoid resource waste. Whether manual or intelligent control can ensure the accuracy and efficiency of the guiding content distribution.

[0080] Embodiment:

[0081] The application proposes a large model-based network guide content automatic generation and distribution method. First, the relevant data of the concerned subject is collected through the data monitoring module; then the hot topic and corresponding data content are obtained through the hot event analysis module of the event graph analysis model; then these hot topics and data are used as auxiliary knowledge base knowledge, and the intelligent guide content generation module based on the optimized large language model is used to generate the guide content through the RAG technology based on the prompt word optimization and retrieval enhancement; finally, the generated guide content is quickly distributed to the execution operators according to different execution modes and distribution strategies through the task distribution management module, and the completion condition is judged by receiving the feedback information of the operators.

[0082] The overall process of the application is shown in Figure 1 , including data monitoring, hot event analysis, intelligent guide content generation and task distribution management process;

[0083] Figure 2 is the data monitoring module process diagram of the application using a crawler to obtain massive Internet information. As shown in Figure 2 , the purpose of the application is to realize multi-source accurate collection of information by combining massive Internet data with field screening. This method is realized through the following two steps:

[0084] Step 1-1: According to the keywords and synonym evolution of multiple fields, a dictionary yq_dict composed of field keyword table and exclusion word table is obtained, and the keywords involved include user attention and user-defined related content, etc.;

[0085] Step 1-2: Collect, store and process Internet data within each hour through big data, and the data source of the Internet data includes key website data crawled by a crawler and outsourcing cooperation data; wherein, the storage management is to use a column-oriented NoSQL database HBase to persistently store the Internet data; the framework algorithm includes distributed offline batch computing technology, distributed memory iteration computing technology and distributed streaming processing computing technology, and the framework algorithm uses MapReduce, Spark and Storm technology respectively;

[0086] Figure 3 is the hot event analysis module process diagram of the event graph analysis model of the application. As shown in Figure 3 , the purpose of the application is to reveal the core content, key topics and evolution trend of hot events by mining the theme structure in event text, so as to better understand the key points of public opinion events and public concerns. This method is realized through the following six steps:

[0087] Step 2-1: Generate topics and summaries based on public opinion news data; use prompt words and large models to achieve this. An example of prompt word content is as follows:

[0088] Please generate a concise and clear summary and event topic based on the news subject content, with a word limit of 150 words. The summary should cover the following points:

[0089] 1. Main event: The 2025 International Technology Exhibition was successfully held in [City Name].

[0090] 2. Key information: It showcased the latest achievements in cutting-edge technology fields such as artificial intelligence, 5G communication, and quantum computing.

[0091] 3. Impact and prospects: The exhibition promoted global technological cooperation and innovation, and is expected to drive technological progress and market development in related fields.

[0092] Step 2-2: Similarity judgment of public opinion events. Based on the event summary generated in step 2-1, generate text embedding representation through event summary combined with qwen3 embedding model, and calculate text similarity using cosine similarity. Cosine similarity measures similarity by calculating the cosine of the angle between two vectors, with the formula:

[0093]

[0094] According to the obtained similarity value Similarity and threshold value m, the similarity of events is judged. If Similarity > m, then the two events are similar, otherwise the two events are not similar.

[0095] Step 2-3: For similar events, use large models to extract event entities, relationships, attributes, and event schema extraction (including event subject, object, trigger word, etc.); update the event graph based on the extracted information; calculate the propagation breadth, propagation depth, and sentiment intensity attribute values in the updated graph.

[0096] Step 2-4: Numerical acquisition and calculation of propagation breadth (B(e)).

[0097] First, real-time acquisition of event-related data through multi-platform API interface:

[0098] (1) Social platforms (such as Weibo, Twitter): Call user interaction interface to collect independent user IDs participating in event discussion, and get (U(e)) (number of participating users) after deduplication.

[0099] (2) News platforms: Count the reading user IPs of articles, and calculate the number of independent access users combined with Cookie deduplication technology, supplemented to (U(e)).

[0100] (3) Video platform: Through the device ID in the playback record, the number of independent users of the video related to the viewing event is obtained.

[0101] Secondly, the traditional method only uses simple normalization of the number of users, and the innovation is reflected in:

[0102] (1) Spatiotemporal weighted user number: Introduce geographical decay factor a geo and time decay factor a time ; The formula is:

[0103]

[0104] Among them, (d i is the distance between the user's location and the event occurrence), (△t i is the difference between the user's participation time and the event occurrence time).(2) Active user correction: Weight the user interaction frequency (such as the number of posts and comments), and set the weight of core users (interaction ≥ 5 times) to 1.2 and the weight of ordinary users to 1.0. The calculation formula is:

[0105]

[0106] (w i = 1.2 if core user, otherwise 1.0)

[0107] The final spread width is:

[0108] (amplify the contribution of active users)

[0109] Step 2-5: Numerical acquisition and calculation of spread depth (D(e)).

[0110] (1) Construct the propagation chain topology: Through the forwarding relationship API of the social platform, obtain the forwarding path of each information, mark the user ID and timestamp of each layer of forwarding, and form F k (e) (the kth layer forwarding volume). Through depth-first search (DFS) traversal of the propagation chain, for example, the forwarding chain of a certain event reaches 7 layers, then n = 7.

[0111] (2) The traditional method only considers the weighted sum of the forwarding volume, and the innovation is reflected in:

[0112] a) Influence level weighting: Introduce user influence index I k (based on the number of fans and activity), the formula is:

[0113]

[0114] Among them, I k,i= log(1 + fans) * interaction rate Make the high-influence user's retweet in the deep level occupy higher weight.

[0115] b) Time decay retweet amount: Assign different weights to retweets at different times. The weight of recent retweets (within 24 hours) is 1.0, and the weight of long-term retweets (more than 7 days) is 0.3. Formula:

[0116]

[0117] (Δt k,i is the difference between the kth layer and the first retweet and the event start time).

[0118] Step 2-6: Numerical acquisition and calculation of emotional intensity (S(e)).

[0119] (1) Text sentiment data: Through natural language processing tools (such as HanLP, NLTK), sentiment analysis is performed on event-related text (comments, posts, news texts), and the positive pos(e i ) or negative neg(e i ) of each sub-topic is output.

[0120] (2) Traditional methods only calculate the difference between positive and negative extremes, and the innovation lies in:

[0121] ① Emotional fluctuation coefficient: Introduce the standard deviation σ pos and σ neg of sub-topic sentiment, formula:

[0122] S(e) = |max(pos(e i ))-max(neg(e i ))| × (1+σ pos +σ neg )

[0123] The greater the emotional fluctuation, the higher the σ value, indicating that the public opinion difference caused by the event is more intense, and the emotional intensity is amplified.

[0124] ② Emotional timeliness correction: Assign time weight τ(t i ) to sentiment values at different times (recent sentiment weight is higher), formula:

[0125] max(pos time (e)) = max(τ(t i ) × pos(e i ))

[0126] Similarly, calculate max(neg time (e)), finally:

[0127] S(e) = |max(postime (e))-max(neg time (e))

[0128] Figure 4 is the intelligent guiding content generation module flowchart of the optimized large language model of the present application. As shown in Figure 4 the present application aims to precisely control the guiding content output of the model by means of cutting-edge large language model generation technology and strategy, so that it can meet the needs of diversified scenarios and efficiently produce high-quality, logically coherent and targeted guiding content, thereby improving user interaction experience and public opinion processing work. This method is realized through the following five steps:

[0129] Step 3-1: Selection and deployment of large language model. For specific application scenarios, select a suitable pre-trained large language model. Considering the business demand for internal network deployment environment and the business being mainly in Chinese, select a domestic open source large model as the basic model for fine-tuning. The main selected models are GLM4-9B of Tsinghua Zhi Spectrum AI and Qwen2.5-72B of Alibaba Tongyi Qianwen to improve the pertinence and expressiveness of the model. Model quantization and VLLM acceleration methods can be used to improve the inference speed;

[0130] Step 3-2: According to the business demand of intelligent guiding content generation, build a style knowledge base like poet style. Build a triple feature library through NLP tools: (1) Vocabulary features: such as Li Bai's "bright moon" and "blue sky", Li Qingzhao's "sorrow" and "tears", count the Top50 high-frequency words of poets and calculate the emotional weight (such as "sorrow" in Li Qingzhao style with an emotional value of 0.8). (2) Rhetorical structure and corresponding poem: such as Du Fu's "well-structured couplets" (extracting "two orioles singing in the willow, a white egret flying in the blue sky" and other well-structured couplet templates), Li He's "metaphor" (such as "Kunshan jade broken, phoenix called, lotus shed tears, and fragrant orchid laughed"). (3) Emotional tendency: quantify the emotional distribution of poet's works (such as Tao Yuanming's "peaceful" in his pastoral poems, accounting for 70%, and "leisure" accounting for 25%). This part of text knowledge includes representative works of poets (such as Li Bai's "Drinking Alone under the Moon", Du Fu's "Spring Sigh"), creative style theory (such as "Li Bai's bold and elegant, using exaggeration and rhetoric"), and life prototype (such as Su Shi's "open-minded and broad-minded, good at using metaphors"). Store them in the structure of "poet ID-style label-high frequency vocabulary-rhetorical example".

[0131] Step 3-3: According to the intelligent guide content generation business requirements, build dialect and tone rule library. Dialect feature library: store according to "dialect type-rhythm rule-characteristic vocabulary-sentence structure", such as: Cantonese: retain "of" "how" "what" and other mood words, follow the nine-tone six-tone of Cantonese (such as generating poetry, the last word is preferred to be selected as the yangping tone). Sichuan dialect: include "Ba suitable" "know" "no" and other words, and end the sentence with "th" "oh" to enhance the oral feeling. Configure tone adaptation model: design conversion rules for each tone, such as: humorous tone: add irony rhetoric on the basis of the poet's original style (such as imitating Li Bai's style to evaluate hot pot: "spicy sky three thousand feet, suspect is the Milky Way falls into the stomach——is spicy straight jump feet th"). Romantic tone: strengthen the emotional characteristics of the poet, such as Li Qingzhao's style of film review: "silver screen light and thin, missing around the heart, a parting, how to resist the wind of the late coming".

[0132] Step 3-4: According to the intelligent guide content generation business requirements, introduce multi-condition weighted retrieval and context fusion. Multi-condition weighted retrieval includes three aspects: ①Style retrieval: according to the selected poet, recall Top20 related features from the style feature library (such as selecting "Su Shi + humorous tone", recalling "metaphor rhetoric" "open-minded vocabulary" "interrogative sentence structure"). ②Dialect retrieval: match the dialect rule library, such as selecting "Sichuan dialect + Du Fu style", recalling "rhythm rule (press into sound rhyme) + characteristic word ('Ba suitable' comfortable)". ③Tone retrieval: retrieve the corresponding conversion template, such as "humorous tone" needs to add "exaggeration + colloquial suffix" (such as "yo" "ma"). The context fusioner mainly integrates the retrieval results into the generation prompt (Prompt), for example: evaluate a comedy film in the style of Su Shi, using Northeastern dialect, colloquial tone, and containing:

[0133] ①High-frequency words: "laugh" "drunk" "human world"

[0134] ②Rhetoric: metaphor (such as comparing the laugh point to "spring rain")

[0135] ③Northeastern dialect features: "lala" "gaga" "what"

[0136] ④Sentence structure: interrogative sentence (such as "life in the world, what's the point of laughing?")

[0137] Step 3-5: According to the intelligent guide content generation business requirements, configure the generation quantity (such as 50), generate multiple sets of text through batch processing, and then filter and calibrate through the following rules: (1)Style consistency: use cosine similarity to calculate the matching degree of the generated text and the poet's feature library (threshold ≥0.75); if the matching degree is less than the threshold, re-perform style consistency generation. (2)Dialect accuracy: call the dialect dictionary to check the usage rate of characteristic words (such as Cantonese net evaluation with a proportion of mood words ≥15%). If the usage rate of characteristic words is less than the threshold, replace the corresponding dialect mood words.

[0138] Figure 5 is the task distribution management module flowchart of the present application. As shown in Figure 5 the present application aims to solve the problems of poor timeliness in task distribution, chaotic execution results, and tedious manual operation. This method is realized through the following five steps:

[0139] Step 4-1: Select the guided content distribution mechanism, mainly including manual control distribution process and intelligent control distribution process based on setting program and related parameters, and the parameters can be modified.

[0140] Step 4-2: Set the number of tasks. The number of tasks refers to the number of operators executing the tasks. Generally, it is set according to different task requirements.

[0141] Step 4-3: Select the task distribution method. If manual control distribution process is selected, manually execute related options according to actual task requirements and operation steps, including selecting task execution method (manual execution, network robot execution, celebrity or big V execution), task distribution strategy (random allocation, corpus repeatable distribution, random allocation, corpus non-repeatable distribution, full allocation), and task effective time.

[0142] For network robots, batch creation or import of robot accounts on the platform is required, and random but seemingly reasonable avatars, nicknames, personal profiles, etc. are generated for each account to simulate real users and reduce the risk of being identified by the platform. For the target platform, study the page structure, interface rules, interaction logic, etc. of the target platform to determine the content publishing entry, required parameters (such as article ID, comment content field, etc.), and platform restrictions that may be set (such as content length, frequency restrictions). According to the characteristics of the platform, write adaptive code to ensure that the program can interact smoothly with the platform.

[0143] Task distribution strategies mainly include three types: (1) Random allocation, corpus repeatable distribution. For example: 10 corpora, 10 people: the distribution result is 1 for each person. 10 corpora, 3 people: distribute 3, 3, and 4. 10 corpora, 12 people: under the premise of distributing 10 corpora for one round, the last 2 people each have 1. (2) Random allocation, corpus non-repeatable distribution. For example: 10 corpora, 10 people: the distribution result is 1 corpus for each person. 10 corpora, 3 people: distribute 3, 3, and 4. 10 corpora, 12 people: the first 10 people each have 1, and the last 2 people have no task execution. (3) Full allocation (each person gets all the corpora): for example: 10 corpora, 10 people: each person has 10.

[0144] Setting task effective time refers to that each performer of the distributed task must complete the execution within the effective time, and if the execution is performed beyond the time, it is invalid. When manually setting the task effective time, various factors such as network environment and operator reaction time need to be considered.

[0145] Step 4-4: If the intelligent control distribution process based on the setting program and related parameters is selected, the program will automatically execute according to the task quantity and the parameters that have been set. The automatic execution includes two strategies:

[0146] Strategy one: the automatic execution process is as follows:

[0147] (1) The task execution mode is evenly distributed according to the task quantity, and if it cannot be evenly distributed, the remaining quantity is uniformly executed by the network robot.

[0148] (2) The task distribution strategy adopts random allocation and corpus repeated distribution mode.

[0149] Strategy two: the task demand, task execution mode (manual execution, network robot execution, celebrity or big V execution), and task distribution strategy (random allocation, corpus repeatable distribution, random allocation, corpus non-repeatable distribution, and full allocation) are input as prompt words into the task distribution intelligent agent based on a large model. The intelligent agent outputs the optimal task distribution mode and strategy; including task execution mode (manual execution, network robot execution, celebrity or big V execution), task distribution strategy (random allocation, corpus repeatable distribution, random allocation, corpus non-repeatable distribution, and full allocation), and simultaneously intelligently predicts the task completion time.

[0150] Step 4-5: Task completion monitoring and feedback adjustment. After the task is completed, each execution module will wait to receive the execution result feedback, including whether it is successfully submitted, whether there is an error message; for the tasks that are not completed within the effective time, the system will select to require resending or not sending according to the manual operation.

[0151] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium can store a computer program, and the computer program can run the invention content of a network-guided content automatic generation and distribution method based on a large model and part or all steps in each embodiment when executed by the data processing unit. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0152] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present application can be realized by means of a computer program and a corresponding general hardware platform. Based on such understanding, the technical solutions in the embodiments of the present application can be embodied in the form of a computer program, i.e. a software product, which can be stored in a storage medium, including a plurality of instructions for causing a device (which can be a personal computer, a server, a single-chip microcomputer, an MCU, or a network device, etc.) comprising a data processing unit to execute the method described in various embodiments or some parts of the embodiments of the present application.

[0153] The present application provides a network guide content automatic generation and distribution method based on a large model. There are many methods and approaches to realize the technical solutions. The above description is only the preferred embodiment of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, some improvements and refinements can be made. These improvements and refinements should also be considered as the protection scope of the present application. The components not explicitly described in the embodiments can be realized by using existing technology.

Claims

1. A large model-based network boot content automatic generation and distribution method, characterized in that, The method comprises the following steps: Step 1, data monitoring, collecting relevant data of the monitoring target in the Internet; Step 2, hot topic analysis, processing the relevant data collected in step 1 to obtain the theme of the hot event and the corresponding data content; Step 3, guided content generation, setting the caliber requirements of the guided content according to the theme of the hot event obtained in step 2, and generating the guided content through a large language model; Step 4, task distribution management, selecting a distribution method and strategy according to the guided content obtained in step 3, and distributing the guided content.

2. The method of claim 1, wherein, Step 1, data monitoring, includes: Step 1-1, according to the set of keywords, perform synonym evolution to obtain a dictionary composed of a domain keyword table and an exclusion word table of the target domain; Step 1-2, according to the dictionary obtained in step 1-1, collect, store and preprocess the Internet data to obtain the relevant data of the monitoring target; the relevant data is the data content corresponding to the set of keywords.

3. The method of claim 2, wherein, Step 2, hot topic analysis, includes: Step 2-1, based on the prompt word, generate the theme and abstract text of each hot event using a large model on the relevant data collected in step 1; Step 2-2: encode the abstract text based on the Qianwen3 vector model and calculate the cosine similarity, and perform similarity judgment according to the cosine similarity, to determine whether there is a similar hot event in the pre-set event graph library, if yes, execute step 2-3, otherwise execute step 2-4; Step 2-3, update the event graph of the similar hot event; Step 2-4, construct a new event graph of the hot event and add it to the event graph library.

4. The large model-based network boot content automatic generation and distribution method according to claim 3, characterized in that, The step 2-3 of updating the event graph of the similar hot event, includes: Step 2-3-1, extract the relevant information of the hot event using a large model, the relevant information includes: entity, relationship, attribute and event architecture, wherein the event architecture includes event subject, object and trigger word; Step 2-3-2, update the event graph of the similar hot event in the event graph library based on the relevant information collected in step 2-3-1; Step 2-3-3, calculate and update the attributes of the similar hot event in the graph library, the attributes include: propagation breadth, propagation depth and emotional intensity.

5. The large model-based network boot content automatic generation and distribution method according to claim 4, characterized in that, The step 2-4 of constructing a new event graph of a hot event, includes: Step 2-4-1, extract the relevant information of the hot event using a large model, the relevant information includes: entity, relationship, attribute and event architecture, wherein the event architecture includes event subject, object and trigger word; Step 2-4-2, construct an event graph based on the relevant information collected in step 2-4-1; Step 2-4-3, calculate the attributes of the hot event in the event graph constructed in step 2-4-2, the attributes include: propagation breadth, propagation depth and emotional intensity.

6. The large model-based network boot content automatic generation and distribution method according to claim 5, characterized in that, The propagation breadth in step 2-3-3 and step 2-4-3 is calculated as follows: Where B(e) is the propagation breadth, max represents the maximum value, ω is the amplification factor, U active (e) is the number of active users, and the calculation method is as follows: where w i is the core user weight of the i-th user, U weight (e) is the spatio-temporal weighted user number, calculated as follows: wherein, α geo (d i ) is a geographic attenuation factor, α time (t i ) is a time attenuation factor, and U(e) is the actual number of users, which is obtained according to different network platforms through a preset rule.

7. The large model-based network boot content automatic generation and distribution method according to claim 6, characterized in that, The propagation depth in step 2-3-3 and step 2-4-3 is calculated as follows: Construct the propagation chain topology, obtain the forwarding path of the hot event, and use F k (e) represents the forwarding volume of the hot event on the kth layer forwarding path. The calculation method of the propagation depth D(e) is as follows: Wherein, K is the number of forwarding path layers, I k,j is the user influence index.

8. The large model-based network boot content automatic generation and distribution method according to claim 7, characterized in that, The emotional intensity in step 2-3-3 and step 2-4-3 is calculated as follows: Using natural language processing tools to perform sentiment analysis on the hot events, obtaining the attributes of each sub-topic e m , including: positive value pos(e m ) and negative value neg(e m ); Calculate the sentiment intensity S(e) as follows: S(e) = |max(pos time (e)) - max(neg time (e)) | x (1 + σ pos + σ neg ) where σ pos and σ neg are the standard deviations of positive and negative values, respectively, pos time (e) represents a positive sentiment time-dependent correction value, and neg time (e) represents a negative sentiment time-dependent correction value, and are calculated as follows: pos time (e) = τ(t m ) x pos(e m ) neg time (e) = τ(t m ) x neg(e m ) where τ(t m ) represents a time weight.

9. The large model-based network boot content automatic generation and distribution method of claim 8, wherein, The guided content generation of step 3 comprises: Step 3-1, select a large language model according to the application scenario, and fine-tune according to the application field; Step 3-2, according to the business requirements of guided content generation, construct style knowledge base, dialect rule base and tone rule base; Step 3-3, according to the theme of the hot event obtained in step 2, use vector retrieval and full-text retrieval in the style knowledge base, dialect rule base and tone rule base; Step 3-4, fuse the retrieval results by reverse ranking fusion method to obtain the guided content.

10. The large model-based network boot content automatic generation and distribution method of claim 9, wherein, The task distribution management of step 4 comprises: Step 4-1, set the guided content distribution mechanism, including: manual control distribution process and intelligent control distribution process; Step 4-2, set the number of distributed tasks; Step 4-3, select the task distribution mechanism for distribution, if the manual control distribution process is selected, step 4-4 is executed, if the intelligent control distribution process is selected, step 4-5 is executed; Step 4-4, manually execute according to the distribution task requirements and operation steps, including: selecting the number of tasks, task execution mode, task distribution strategy and task effective time; Step 4-5, automatically execute according to the number of distribution tasks and the parameters set; Step 4-6, after the distribution task is completed, receive the execution result feedback, for the distribution task which is not completed within the effective time, manually select to require to resend or not to send.

Images