Intelligent programming method and system for broadcasting programs

Abstract

The application discloses an intelligent arrangement method and system for broadcast programs, and relates to the technical field of intelligent broadcast program arrangement.The method comprises the following steps: collecting broadcast materials and interactive data, and constructing a structured material pool; generating an initial program list based on program template rules, and injecting random entropy to generate bridging content at the connection of block items; performing logical verification and dynamic adjustment on the bridging content to obtain an optimized program list; generating serial languages matched with the optimized program list based on a diffusion model, and optimizing the serial languages; scoring the optimized program list and the serial languages, automatically filtering out high-risk content, and forming a safe broadcast program list for broadcast. The method realizes second-level compliance verification and safe filtering, greatly reduces the cost of manual review, and improves the broadcast efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent broadcast program scheduling technology, and in particular to an intelligent scheduling method and system for broadcast programs. Background Technology

[0002] Currently, radio program scheduling mainly relies on predefined program templates and manual configuration processes. Users maintain calendar-style program templates through a web interface, specifying sections such as time announcements, scripts, news, music, commercials, and trailers for each time slot, and selecting material sources and broadcast methods for each section. This method can quickly generate daily program schedules, but the transitions between sections still largely rely on fixed connecting phrases or preset segments, lacking the ability to dynamically adjust content transitions.

[0003] In actual operation, program schedules often face uncertainties such as a surge in listener interaction, breaking news, or traffic information interruptions. These situations can disrupt the original segment order, making it difficult for fixed transitions to balance continuity and diversity. At the same time, the existing programming lacks real-time logical verification and optimization methods for transition content, and it cannot proactively adjust the connection strategy based on user behavior data or content entropy indicators, resulting in a less than smooth program experience. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an intelligent scheduling method for broadcast programs to address the problems of existing broadcast program scheduling systems in terms of dynamic and diversified connection, online causal verification and optimization, and automated compliance risk management. These problems include the inability to proactively generate diversified transition content, the lack of a real-time causal effect evaluation mechanism, and the difficulty in achieving second-level semantic compliance and automatic screening of copyright risks.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an intelligent arrangement method for broadcast programs, which includes: collecting broadcast materials and interactive data, and constructing a structured material pool;

[0008] An initial program list is generated based on program template rules, and random entropy is injected at the junction of section items to generate bridging content.

[0009] Perform logical verification and dynamic adjustment on the bridging content to obtain an optimized program schedule;

[0010] Generate and optimize program schedule matching based on a diffusion model, and optimize the program schedule matching.

[0011] The optimized program schedule and connecting phrases are scored, high-risk content is automatically filtered out, and a safe broadcast program schedule is generated for radio broadcast.

[0012] As a preferred embodiment of the intelligent programming method for broadcast programs described in this invention, the broadcast materials include music, news, weather, traffic conditions, advertisements, stories, dialogue, trailers, long programs, short programs, opening remarks, and closing remarks; the interactive data includes song requests, messages, comments, votes, and audio uploads submitted by users through mini-programs, applications, or websites.

[0013] The construction of the structured content pool includes tagging the broadcast content, archiving the interactive data, and aggregating all tagged broadcast content and archived interactive data into a database to form a structured content pool that supports multi-condition retrieval, automatic arrangement, and AI content generation.

[0014] As a preferred embodiment of the intelligent program scheduling method for broadcast programs described in this invention, the generation of the initial program list includes: extracting corresponding candidate material sets from the structured material pool according to the type, duration, and playback order constraints of each section and section item in the program template; calculating the interaction diversity entropy according to the proportion of various user interactions in the total interactions, and statistically analyzing the proportion of each subtype of material in each candidate material set to obtain a content diversity index; and linearly weighting the content diversity index and the interaction diversity entropy according to preset weights to obtain a bridging entropy intensity value.

[0015] When the bridging entropy intensity value is not lower than a preset threshold, a local perturbation is performed on each pair of adjacent segment items; the perturbation includes mapping the bridging entropy intensity value to a perturbation probability according to a preset rule; generating a random number for each pair of adjacent segment items and comparing it with the perturbation probability; if the random number is not higher than the perturbation probability, the order is swapped; and ensuring that after the swap, the number of consecutive occurrences of the same type of segment items does not exceed the upper limit, and the program list still conforms to the template order constraint.

[0016] Before performing the perturbation, if the bridging entropy strength value is not lower than the insertion threshold or the content difference between adjacent segment items is not lower than a preset threshold, then bridging content is inserted between adjacent segment items; the bridging content includes at least one of impromptu connecting words, interactive summaries, hot news flashes, or sound effect clips, and the length and tone style of the bridging content are determined by interpolation of the bridging entropy strength value within their respective preset intervals; the segment item sequence after completing the perturbation and inserting the bridging content is used as the initial program list.

[0017] As a preferred embodiment of the intelligent scheduling method for broadcast programs described in this invention, the method for obtaining an optimized program list includes: extracting the bridging entropy intensity value of each bridging position from the initial program list; inputting the bridging entropy intensity value and the corresponding interaction data of the bridging position into a causal inference model, and obtaining causal weights for each bridging position from the causal inference model; combining the bridging entropy intensity value and the causal weights proportionally to generate a joint control factor for each bridging position; sequentially traversing all bridging positions and optimizing each bridging position according to the numerical range of the joint control factor, wherein the optimization adjustment includes version reselection, sequence perturbation, and bridging content enhancement; and generating and outputting the optimized program list after traversing and completing the dynamic adjustment of all bridging positions.

[0018] As a preferred embodiment of the intelligent scheduling method for broadcast programs described in this invention, the step of generating and optimizing the program schedule matching concatenation includes: extracting the joint control factor of each bridging position in the optimized program schedule; modifying the basic noise coefficient based on the value of the joint control factor before noise injection in the forward diffusion process; and in each reverse denoising iteration, using the joint control factor as a gradient correction weight to adjust the gradient output of the noise prediction network, and superimposing the corrected gradient into the denoising result.

[0019] During the sampling phase, based on the default number of denoising iterations and the default sampling step size of the diffusion model, the joint control factor is compared with a preset threshold. When the joint control factor is higher than the preset threshold, the default number of denoising iterations is increased to the integer part of the product of the default number of iterations plus the difference between the joint control factor and the threshold and the predetermined iteration increment. The sampling step size is shortened by calculating the difference between the joint control factor and the preset threshold, multiplying the difference by a predetermined distance reduction coefficient to obtain the step size reduction magnitude. The result of subtracting the step size magnitude from the default sampling step size is used as the new sampling step size. When the joint control factor is lower than the threshold, the default number of iterations and sampling step size are maintained.

[0020] As a preferred embodiment of the intelligent scheduling method for broadcast programs described in this invention, the optimization of the connecting phrases includes: calculating a diversity score, a causal guidance score, and a generation depth score for each candidate connecting phrase based on the recorded noise injection ratio, gradient correction weights, and sampling parameters; the diversity score is obtained by calculating the information entropy of the word frequency distribution of the candidate connecting phrase and normalizing it according to the total text length; the causal guidance score is obtained by taking the arithmetic mean of the gradient correction weights for the candidate connecting phrase in all denoising iterations; the generation depth score is obtained by subtracting a preset minimum number of iterations from the actual number of denoising iterations of the candidate connecting phrase, dividing the difference by the difference between the maximum number of iterations and the minimum number of iterations, and using the result as the normalized generation depth score; the diversity score, causal guidance score, and generation depth score are linearly combined according to preset weights to obtain a comprehensive score, and candidate connecting phrases with a comprehensive score lower than the minimum passing threshold are eliminated;

[0021] The audience retention rate is defined as the ratio of the number of remaining listeners at the end of a program segment to the number of listeners at the beginning of the segment. Based on historical broadcast data, the average retention rate is statistically obtained under different bridging positions, similar contexts, and text features, and a retention rate prediction model is constructed. Each candidate linking phrase, its corresponding joint control factor, and the context features are input into the retention rate prediction model to calculate the expected retention rate of the candidate linking phrase, and candidate linking phrases with expected retention rates lower than the historical average retention rate are eliminated. The remaining candidate linking phrases are compared with the preceding and following text segments for semantic similarity, and candidate linking phrases with semantic similarity lower than a preset consistency threshold are eliminated. The final remaining linking phrases that meet the contextual coherence and their corresponding joint control factors are used as the final broadcast script and submitted together with the optimized program schedule.

[0022] As a preferred embodiment of the intelligent scheduling method for broadcast programs described in this invention, the process of forming a safe broadcast program list includes: reviewing the content of each section item text and connecting phrase based on a violation word library, a pre-trained deep classification model, and a copyright database interface, and generating a risk score; classifying the risk scores according to preset grading rules, and automatically replacing high-risk content while submitting medium-risk content for manual review; during automatic replacement, selecting matching content from a safe alternative library for replacement based on the tags, type, and duration of the original material; and after the manual review is completed, updating the review results to the program list to form a safe broadcast program list.

[0023] Secondly, this invention provides an intelligent program scheduling system for broadcast programs, comprising: a data module for collecting broadcast materials and interactive data, and constructing a structured material pool; a bridging module for generating an initial program list based on program template rules, and injecting random entropy at the junctions of section items to generate bridging content; an optimization module for performing logical verification and dynamic adjustment on the bridging content to obtain an optimized program list; a connecting phrase module for generating connecting phrases that match the optimized program list based on a diffusion model, and optimizing the connecting phrases; and a scoring module for scoring the optimized program list and connecting phrases, automatically filtering out high-risk content, and forming a safe broadcast program list for broadcast.

[0024] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the intelligent scheduling method for broadcast programs as described in the first aspect of the present invention.

[0025] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent scheduling method for broadcast programs as described in the first aspect of the present invention.

[0026] The beneficial effects of this invention are as follows: By injecting random entropy at the junctions of program segments, dynamic diversification of transitional content is achieved, enhancing the audience's sense of novelty compared to traditional fixed linking phrases; relying on online causal reasoning to evaluate the impact of transition operations on retention rates in real time, and automatically adjusting bridging strategies through joint control factors, ensuring program coherence while achieving optimal user experience; the improved diffusion model adaptively applies control factors in each noise injection and denoising iteration, enabling the AI ​​linking phrase style and length to accurately match context and audience preferences; finally, combining a deep classification model and a copyright database interface to perform risk scoring and automatic replacement of program schedules and linking phrases, completing second-level compliance verification and security screening, significantly reducing manual review costs and improving broadcast efficiency. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart of an intelligent scheduling method for broadcast programs. Detailed Implementation

[0029] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0030] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0031] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0032] Example 1, referring to Figure 1 As one embodiment of the present invention, this embodiment provides an intelligent scheduling method for broadcast programs, comprising the following steps:

[0033] S1: Collect broadcast materials and interactive data, and build a structured material pool.

[0034] The broadcast materials include music, news, weather, traffic conditions, advertisements, stories, dialogue, trailers, long programs, short programs, opening remarks, and closing remarks; the interactive data includes song requests, messages, comments, votes, and audio uploads submitted by users through mini-programs, applications, or websites.

[0035] Building a structured content pool includes: tagging broadcast content, adding the following tags to each piece of content and storing them as field attributes: category (news, music, talk show, advertisement, etc.); mood / style (relaxed, serious, enthusiastic, etc.); playback time suitability (morning rush hour, nighttime, etc.); content length range (≤30 seconds, 30 seconds to 2 minutes, etc.); AI broadcast readability (human only, AI-enabled, both); recommendation priority (normal, priority, strong recommendation); and archiving interactive data, structurally recording song requests, messages, comments, votes, and uploaded audio, and adding tags such as interaction type, timestamp, and user preferences.

[0036] All tagged broadcast materials and archived interactive data are aggregated into a database to form a structured material pool that supports multi-condition retrieval, automatic arrangement, and AI content generation, which can be used in subsequent program schedule arrangement, entropy injection, and diffusion-style content generation.

[0037] The system unifies the collection and preprocessing of broadcast materials and interactive data. By tagging each piece of material with multi-dimensional tags such as program category, emotional style, playback time suitability, content length range, AI broadcast readability, and recommendation priority, the localized resource library gains a rich semantic hierarchy for the first time. Compared to traditional material pools that are only categorized by type or simple duration, this solution can quickly locate fine-grained content that meets program requirements during the retrieval stage, significantly reducing manual screening time. For example, during the morning rush hour, simply searching for tags such as "music," "high-energy," "≤2 minutes," "AI-enabled," and "priority" can quickly yield 120 candidate materials that meet the criteria.

[0038] All interactive actions—song requests, messages, comments, polls, and audio uploads—are structured and stored in the same database according to interaction type, timestamp, user ID, and preference tags. By comparing the proportion of each interaction in the total, audience engagement metrics can be calculated in real time, providing dynamic parameters for injecting entropy into subsequent programming. Unlike the previous method of manually reviewing logs or analyzing reports, this interaction archiving method creates a closed loop between trending behaviors and program materials within a time window, ensuring that content inserted during peak interaction periods meets user needs and improves the relevance between the program and the audience.

[0039] Furthermore, the entire resource pool is supported by a dual-engine architecture of Elasticsearch and a relational database, enabling multi-condition composite queries and parallel retrieval. Combined with scheduled tasks and triggers, it achieves seamless integration of the "automatic orchestration—entropy injection—diffusion generation" process. Facing large-scale, multi-format, and high-concurrency scheduling scenarios, it effectively solves the bottlenecks of previous systems in terms of real-time performance, accuracy, and scalability.

[0040] S2: Generate an initial program list based on the program template rules, and inject random entropy at the junction of the segment items to generate bridging content. Perform logical verification and dynamic adjustment on the bridging content to obtain an optimized program list.

[0041] Program template :Include Each sector Each section There are several sections below. It also specifies the content type and playback order constraints for each item.

[0042] Structured Material Pool Prepare multiple versions for each type of material, including music, news, advertising, and stories, and attach a set of tags to each piece of material.

[0043]

[0044] User behavior set : Counting different interaction types within a specified time window, such as the number of song requests. Number of comments Number of comments wait.

[0045] Contextual features Such as holiday signs, types of emergencies, etc.

[0046] Real-time interaction intensity : Sector Item The number of related interactions in the last five minutes.

[0047] in, This represents a collection of program templates; This indicates the total number of sections contained in the template; Indicates the first Each section (such as music section, advertising section, story section); Indicates the first The first in the section Each section item (i.e., playback unit); This represents the sector index, with values ​​ranging from 1 to... ; Indicates the segment item in the segment Sequential index in; Indicates the plate The collection of all sub-items; Represents a structured resource pool; This indicates the set of tags attached to each piece of content; type indicates the content type (e.g., music, advertisement, news, etc.); duration indicates the duration of the content; style indicates the style of the content (e.g., cheerful, calm, formal, etc.); copyright indicates the copyright status tag of the content; Represents a set of user interaction behaviors; Indicates the number of times a song request was made; Indicates the number of times a message was left; Indicates the number of times a comment was made; This indicates the current external environment context characteristics.

[0048] Candidate material set extraction for each section item in the template According to its type label, from The candidate set is obtained through screening:

[0049]

[0050] in, Indicates the section item The corresponding set of candidate materials; Represents a single asset in the asset pool; type Express material Content type tag; type Indicates the section item Required content type tags; Indicates the template number The first section Item content.

[0051] Calculate the total number of interactions:

[0052]

[0053] And obtain the interaction diversity entropy:

[0054]

[0055] in, Indicates the total number of interactions; Indicates the first The number of times a type of interactive behavior occurs; This represents the entropy of user interaction diversity; Indicates the first The proportion of this type of interactive behavior in the total interactions; This indicates the category number of the user interaction behavior.

[0056] For candidate set By subtype Group the elements, and denote the number of elements in each group as . The total is:

[0057]

[0058] Calculate content entropy:

[0059]

[0060] in, Indicates candidate set Total number of materials in the file; Indicates that the candidate set belongs to a subtype The quantity of materials; Indexes representing subtypes (such as style, language); Indicates the section item Content entropy.

[0061] Constructing the bridging vector:

[0062]

[0063] Linearly map it to the bridging strength factor:

[0064]

[0065] in, Indicates the section item The bridging vector; Indicates the section item The intensity of real-time interaction; Indicates the characteristics of the current environment context; Indicates the section item The bridging entropy intensity factor; The mapping weights represent the entropy of user interaction; The mapping weights represent the content entropy; Mapping weights representing the intensity of real-time interaction; Represents the mapping weights of the environment context.

[0066] For candidate set In Version 1: If If the probability is equal, then random selection is performed; otherwise, the first number is selected according to the historical preferred sequence. Version number 1.

[0067] For adjacent Item set ;like Then, for each pair of adjacent terms, the probability is:

[0068]

[0069] generate ,like Then swap the two positions; if the swap results in the number of consecutive occurrences of the same type exceeding the limit, then... If so, then skip the swap.

[0070] in, This indicates the probability of executing a local perturbation; This represents a function that takes the minimum value. Represents a random variable; It represents a uniform distribution of 0 to 1.

[0071] In each pair of adjacent terms If Or the degree of thematic difference Then, insert bridging content, which can be impromptu dialogue, interactive summary, hot news, or sound effects trailer.

[0072] The length of the bridging section is:

[0073]

[0074] The style varies between the two preset templates. Interpolation. If Then it enters high-entropy mode. Expand the range of version candidates and the perturbation window. Increase to Prioritize inserting interactive summaries or trending news.

[0075] in, Indicates the actual generated length of the current bridged content; Indicates the preset minimum bridging length; Indicates the preset maximum bridging length; This represents the function that takes the maximum value. This represents the intensity threshold that triggers the high-entropy mode; Indicates the threshold for topic dissimilarity; Indicates the version selection threshold; Indicates the order perturbation threshold; This indicates the bridging insertion threshold.

[0076] The final output program schedule is as follows:

[0077]

[0078] in, This indicates the optimized and complete program schedule; Indicates the first The optimized program content; This indicates the total number of items in the optimized program schedule.

[0079] Read the initial program list, where each item Includes: material tags, bridging vectors Entropy intensity factor .

[0080] Obtain real-time audience interaction data ( (For interactive category index) and current environment context labels .

[0081] Maintain a simplified causal model, with nodes including each bridging decision. (Regarding position) Operations and performance indicators (Audience retention rate). Collect user retention or interaction rate data in real time within a sliding window for each location. Calculate incremental causal gain:

[0082]

[0083] in, Indicates the bridging location causal gain value; This refers to metrics such as user retention rate or engagement rate. Indicates position A binary variable indicating whether to perform the bridging operation; To express mathematical expectation; do This indicates the intervention condition that forces the operation to be performed at the bridging location; do This indicates the condition under which bridging operations are not performed. An online Bayesian filter is then used to... Smooth updates, suppressing noise.

[0084] Will Mapped to causal feedback factors:

[0085]

[0086] Combination With feedback factors Calculate the joint control factor:

[0087]

[0088] in, Indicates the bridging location The causal feedback factor has a value range between [0, 1]. The hyperparameter representing the slope of the causal feedback curve; Indicates the bridging location The joint control factor; The fusion weight represents the entropy factor, with a value range of [0, 1]. Indicates the bridging location The entropy intensity factor.

[0089] In each location The utility function is defined as follows:

[0090]

[0091] in, Indicates control factor In position Multi-objective utility value; The weight hyperparameters representing the causal gain term; The penalty weight for the coherence offset term; Indicates the reward weight of diversity enhancement items; Indicates position Deviation in thematic coherence between adjacent content; Indicates position Diversity increment factor at the location.

[0092] For each Solve using an approximation of gradient ascent:

[0093]

[0094] The number of iterations should be controlled within 1 to 3 times to ensure that the single-point decision latency is less than 5ms.

[0095] For each position Based on the optimal joint factor Compare with a preset threshold and execute:

[0096] when When reselecting a version, start from the position. Remove the current version from the candidate material list and select causal gain. Replace the highest-ranked version.

[0097] when At that time, perform sequential perturbation with probability. With the next item Swap; if it leads to consecutive overruns of the same type. Then skip it.

[0098] when At that time, enhance the bridging content, in the location Insert bridging segment; bridging length according to Confirmed. The preferred type is interactive summary or trending news.

[0099] when At this time, a conservative strategy is chosen, preserving the version and order in S2 without inserting additional bridging. In this embodiment: , , .

[0100] in, Indicates position The optimal control factor is obtained by solving the problem. This indicates the control factor threshold for version reselection; The control factor threshold represents the sequential perturbation. This indicates the control factor threshold for bridging content enhancement; Indicates the minimum insertion length of the bridged content; Indicates the maximum insertion length of the bridged content; This indicates the maximum allowable limit for consecutive occurrences of items of the same type.

[0101] Traverse positions in order to The process iterates through all bridging locations, applying the aforementioned dynamic adjustment rules one by one and updating the program schedule in real time; each update is written to the decision log and recorded. Selected version identifier, whether disturbance is present, bridging segment type and length; after completing all position adjustments, output the final optimized program schedule. :

[0102]

[0103] in, This indicates that after traversal and adjustment, the position at the th position is... The final optimized program list entries for each location. Includes updated material identifiers, sequence status, bridging content, and , , , .

[0104] During the program generation phase, referencing the predefined segment types and item order in the template, the system automatically extracts a set of candidate materials matching each segment item and injects diverse transitional content at the junctions of adjacent segment items according to random entropy intensity. The injected bridging elements include impromptu connecting phrases, interactive summaries, trending news, or sound effects trailers. The timing and length of each type of content insertion are dynamically determined by real-time calculated entropy intensity and content difference to ensure the transitions are neither abrupt nor redundant, and to appropriately shuffle the material order to create a sense of novelty. This process is fully automated, adhering to playback order constraints while providing flexible content insertion during high-interaction periods or thematic nodes, forming the first initial program list.

[0105] Compared to traditional programming methods that rely solely on static tag matching or manual fine-tuning, this solution uses random entropy injection to achieve controllable perturbation of content order within the same segment and dynamic generation of multiple version candidates. This significantly improves program coherence while also catering to diverse user experiences. Before each perturbation, the system detects the consecutive occurrences of similar content to avoid repetition—a detail reinforcement that is easily overlooked in manual operations. When there are significant differences in content, hot topic summaries are prioritized, avoiding the slow response to sudden events that occurs in manual programming. This allows for more accurate content recommendations and transitions during peak audience attention periods.

[0106] After initial bridging, a lightweight causal reasoning model is used to analyze the entropy intensity and interaction data at all bridging points, generating joint control factors for further fine-tuning. This fine-tuning includes version reselection, sequence fine-tuning, and bridging enhancement. Each step is carried out while maintaining the original template's logical sequence, avoiding subjective biases from human experience and overcoming logical inconsistencies that might result from purely random strategies. This dual closed-loop scheduling based on entropy and causality preserves content diversity while optimizing listener retention, providing a reliable, efficient, and flexible solution for real-time broadcast programming.

[0107] S3: Generate and optimize the connecting phrases for program schedule matching based on the diffusion model, and optimize the connecting phrases.

[0108] Get optimized program schedule Each entry carries an entropy intensity factor. , and joint control factors.

[0109] In the positive phase of each diffusion round, the standard noise figure is... Adjusted to:

[0110]

[0111] in, This represents the adjusted noise figure under joint control factor modulation; The diffusion model is represented by the first... Standard noise figure of the wheel; This represents the weighting coefficient of the combined control factors; Indicates the first Joint control factor for each bridging location.

[0112] exist Introducing more randomness at higher speeds, Maintain text stability at low speeds.

[0113] Returning to the noise reduction stage, Added as gradient correction weights:

[0114]

[0115] in, Indicates the first The diffusion latent variables of the step; The generated number after denoising Step latent variables; The model represents the first time. The noise distribution estimated step by step; This represents the adjustment coefficient of the causal gradient-guided term; Indicates the causal scoring function pair The gradient.

[0116] This causes the diffusion model to automatically bias towards increasing retention rates during generation.

[0117] For each position Perform the following sampling procedure to generate Candidate Connecting Phrases :

[0118] Set the default number of denoising iterations and default sampling step size .

[0119] Compare With preset threshold when At that time, the number of new iterations New Step .

[0120] Otherwise, let , .

[0121] use The above and Full run Secondary reverse diffusion, generating .

[0122] in, This represents the number of inverse diffusion iterations after adaptive update; This indicates the default number of denoising iterations; This indicates the control threshold used to determine whether adjustment is needed. This indicates the adjustment range of the number of iterations corresponding to each unit factor increment; This indicates the updated backdiffusion time step; This indicates the default diffusion sampling time step; This indicates the step size adjustment range corresponding to each unit factor increment.

[0123] For each Record the corresponding Noise ratio , Gradient weights, actual and .

[0124] Candidate Connectives Multidimensional Scoring for Each Calculate the scores for the three items:

[0125] Diversity score Information entropy normalization based on text word frequency distribution; causal-oriented scoring Take the gradient correction term Arithmetic mean; generate depth score : Map the actual number of iterations to [0, 1].

[0126] Then by weight The linear combination yields the overall score:

[0127]

[0128] Candidates with a total score below the minimum passing threshold are removed. Indicates the first A comprehensive score for each candidate phrase; Indicates the first Diversity score of the item; Indicates the first The causal orientation score of the item; Indicates the first The depth score of the generated bar; , , These represent the weighting coefficients for the three types of scores mentioned above.

[0129] The number of listeners at the beginning of a program segment is denoted as... The number of listeners at the end is recorded as Audience retention rate That is, the ratio of the two:

[0130]

[0131] in, Indicates the actual audience retention rate; This indicates the number of listeners at the start of the playback of this segment; This indicates the number of listeners at the end of the playback.

[0132] From past broadcast logs, each bridging location was numbered. Similar environmental labels Collection of retention rate samples based on similar text features (such as word vector clustering results); calculation of the average retention rate of this group of samples. and standard deviation This serves as the basis for subsequent screening and threshold setting.

[0133] For each candidate string phrase Constructing feature vectors: joint control factors ; Context (One-Hot encoding); bridging vector In , , Text features, for Extract the 768-dimensional sentence vector from the pre-trained BERT model. This example uses Chinese BERT-base, with 768-dimensional sentence vectors.

[0134] The Gradient Boosting Regressor (XGBoost) is used, with the following parameters: learning rate 0.1, tree depth 6, number of trees 200; training set size of at least 100,000 historical records, validation set of 20,000 records, and test set of 20,000 records; root mean square error (RMSE) and coefficient of determination are used as the benchmarks. Evaluate model performance, target RMSE 0.05.

[0135] Each candidate Input the trained retention rate model to obtain the predicted retention rate. For all candidate i at position m, if (or more strictly) If the candidate text is not found, then that candidate text will be removed, and the remaining candidates that meet or exceed the historical average retention rate will be retained.

[0136] Candidate Connectivity and its adjacent text , Extract the BERT sentence vectors for each, and denote them as follows: , , Calculate cosine similarity:

[0137]

[0138] Set a consistency threshold .in, Indicates the cosine similarity between the candidate phrase and the previous program; Indicates the cosine similarity between a candidate phrase and the following program; The cosine similarity function represents the similarity between two vectors; Sentence vectors representing candidate concatenation phrases (BERT extraction); Indicates position The program tag vector; Indicates position The program tag vector.

[0139] like or If a candidate fails to meet the given criteria, it is eliminated; only texts that simultaneously satisfy the continuity between the preceding and following paragraphs are retained. The remaining candidates are then sorted from highest to lowest based on their overall score, and the first candidate is selected. and will and corresponding joint control factors Write back to program list entries when After all processing is complete, the final sequence of scripts is output, along with the optimized program list, and submitted to subsequent modules for use.

[0140] in, Indicates position The entropy intensity factor (S2 output); Indicates position The One candidate string phrase; Indicates the final selected connecting phrase; Indicates the bridging location Historical average retention rate; Indicates the bridging location The standard deviation of the retention rate; Indicates the first Prediction retention rate of candidate terms; The minimum threshold representing semantic consistency before and after; This indicates the total number of bridging positions in the program schedule; Indicates position The bridging context feature vector.

[0141] The optimized program list, along with the random entropy intensity and causal feedback factor corresponding to each position, is fed into the diffusion language model. The intensity of the positive noise injection is dynamically adjusted using a joint control factor. During the positive diffusion phase, the noise ratio is automatically adjusted up or down according to the joint control factor, introducing richer random perturbations at nodes requiring enhanced diversity, while maintaining stability at nodes with high coherence requirements. Subsequently, a gradient guidance term is added to the denoising stage, injecting the causal gain signal into the gradient update. Through the direction correction of the noise prediction network output, it automatically biases towards a copywriting style that improves listener retention without excessively sacrificing the naturalness of transitions. This strategy differs from the traditional fixed setting of noise coefficients and gradient weights, achieving fine-grained control over the generation process.

[0142] During the candidate concatenation sampling phase, the system adjusts the number of iteration rounds and step size based on a joint control factor, generating multiple versions of text in parallel and recording the noise ratio, gradient weight, and actual iteration parameters for each candidate. Subsequently, based on a scoring mechanism encompassing word frequency information entropy, causal guidance strength, and iteration depth, all candidate texts are comprehensively scored, eliminating redundant or off-topic texts. Next, text vectors extracted from a pre-trained language model, along with environmental context and bridging features, are input into a retention rate predictor. A regression model then selects the most attractive texts, while semantic coherence is verified between paragraphs to ensure the final selection is not only rich and diverse but also logically consistent.

[0143] Compared to existing methods that rely solely on static templates or uniform diffusion generation, this solution achieves end-to-end adaptive control from noise injection, gradient guidance, sampling step size to multi-level scoring, eliminating reliance on manually set fixed parameters. This results in significant advantages across multiple dimensions, including improved retention rates, generation efficiency, text diversity, and content coherence: retention is reliably improved, the candidate selection process remains highly efficient, and the final copy's match with the program schedule and the listener experience are simultaneously optimized.

[0144] S4: Scores the optimized program schedule and connecting phrases, automatically filters out high-risk content, and generates a safe broadcast program schedule for radio broadcast.

[0145] Read the optimized program schedule generated by S3 Each of the entries Includes: final material identifier, sequence status, bridging content, and final connecting phrase. and joint control factors Synchronously acquire S3 decision logs (each entry is already recorded). ).by Scan program entries in descending order of frequency, prioritizing segments with the strongest randomization and highest risk potential.

[0146] Perform unified cleaning on the block text and the linked text (remove HTML, URLs, and control characters).

[0147] Segment by space or Jieba, and match word units with the level 3 sensitive word database:

[0148] Level A → 30 points per hit, Level B → 20 points per hit, Level C → 10 points per hit.

[0149] A single text entry can be scored up to 30 points, and the score is recorded as follows: Deep CNN classification includes preprocessing and embedding. For Chinese, word-level segmentation is performed, and for English, words are segmented by spaces. A maximum of 200 words are retained, and any insufficient words are padded with pads.

[0150] Use 300-dimensional publicly available pre-trained Chinese FastText word vectors; initialize with random small values ​​for out-of-vocabulary words.

[0151] Generate shape The embedding matrix is ​​used as the network input. Three sets of one-dimensional convolutions are performed in parallel, with kernel widths of 3, 4, and 5, and 128 channels each. Each set of convolutions is followed by ReLU activation, and then global max pooling is performed to obtain a 128-dimensional vector. The three pooling results are concatenated to form a 384-dimensional vector, which is then subjected to Dropout 0.5 to prevent overfitting before being fed into a 256-unit fully connected layer.

[0152] The output layer has 3 nodes, and Softmax generates three probabilities: "compliant / suspected non-compliant / non-compliant". Key parameters include: 300-dimensional word vectors, 384 total convolutional kernels, 256 hidden layers, and a model size of <2MB.

[0153] 500,000 broadcast texts were collected and manually labeled into three categories, with a ratio of 8:1:1 for training / verification / testing.

[0154] The optimizer, Adam, has an initial learning rate of 0.001, a batch size of 128, and a maximum of 20 rounds. It stops early if there is no improvement after 3 rounds on the validation set. The test set targets are: overall accuracy ≥ 95% and non-compliance recall ≥ 90%.

[0155] Model compression: 8-bit quantization + weight pruning, inference latency <10ms / text; outputting probabilities of "non-compliant" models. ,calculate:

[0156]

[0157] Call the copyright library API item by item. If an item contains multiple restricted materials, the highest risk score will be used: 0 points for authorized; 10 points for pending confirmation or anomaly; 20 points for unauthorized. .

[0158] The total risk score is calculated as follows:

[0159]

[0160] in, Indicates the first Risk scoring based on sensitive dictionary hits in the content; This indicates that the CNN model determines the first... The probability value of the content being "non-compliant" ranges from [0, 1]. Indicates the first The machine learning risk score for each item; 50 is a linear mapping coefficient that converts the probability into a scoring range of 0-50; Indicates the first Copyright risk score for each item; Indicates the first The total risk score for each item. The maximum total risk score is 100, and the score is written into the program list item for future reference.

[0161] when At times, if the content is deemed high-risk, candidates are retrieved from the safety reserve based on the type, duration, and style tags of the original material. The most recently updated material that has not been highlighted and has the highest cosine similarity to the original material, along with sufficient stock, is selected for replacement. After replacement, it is ensured that the new material is of high risk. Simultaneously, the connecting phrase is replaced by first searching the safe connecting template library by opening / transition / ending; if no match is found, a general transition phrase is used to postpone the original connecting phrase to the next safe insertion point; the new text is also re-scored until it is low risk.

[0162] when At this time, it is considered medium risk. Risk details are displayed in batches in the WEB review panel list. Reviewers can approve / edit / replace, and the review results are written back to the program schedule in real time, triggering automatic re-scoring. After passing the review, the risk score must also be <40. At that time, it was considered low risk.

[0163] The system writes the following information into the audit database for each replacement / edit / release record: original ID, original total score, new ID, new total score, operation method, timestamp, and operator.

[0164] After automatic replacement and manual review, a safe broadcast program list is obtained. If the risk score of all items is less than 40, the item will be directly pushed to the broadcast control system.

[0165] in, This indicates the final safe broadcast program list, where all program items have a risk score below 40.

[0166] In the broadcast preparation phase, the system first performs a unified cleaning of all program entries and corresponding connecting texts, removing redundant web page tags and link information. Then, a three-level sensitive word database is loaded locally, and each text is scanned word-by-word in a hierarchical manner. Next, a deep convolutional network module performs semantic classification on the processed text. The network structure, while maintaining a model size of less than two megabytes, can complete inference at the millisecond level, and quantization techniques ensure operational efficiency. A copyright verification interface works in parallel, automatically comparing the licensing status of the materials and generating risk markers to block unauthorized or abnormal items at the source. The entire process requires no manual intervention, completing security filtering instantly without the listener's awareness.

[0167] During the content risk assessment phase, the system summarizes the risk information obtained from scanning and classification, and automatically routes content according to three levels of assessment results: "higher risk," "medium risk," and "low risk." For content marked as higher risk, the system immediately retrieves the highest quality and most abundant compliant materials of the same type from the safety material library to replace it, and simultaneously matches suitable transitional text from the linking phrase template library as a substitute. For medium-risk items, the system pushes them to the online review interface in real time. The interface intuitively displays the reasons for the risk and intelligently recommended replacement solutions. After the reviewers complete the confirmation or modification in a very short time, the system automatically recalculates the risk score until compliance is achieved. Low-risk items are directly retained and enter the subsequent broadcast process to ensure that the program's rhythm is not interrupted.

[0168] All operations are automatically recorded in the audit database, including original data identifiers, risk values ​​before and after assessment, replacement or release types, operation time, and operator information. Using a visual monitoring panel, platform administrators can flexibly query and export logs based on program category, processing time, or specific operator, easily completing compliance checks and risk sampling. This closed-loop process achieves fully automated control from security testing to material replacement and review and retention, ensuring the compliance of broadcast content while minimizing the burden of manual review.

[0169] This embodiment also provides an intelligent program scheduling system for broadcast programs, including: a data module for collecting broadcast materials and interactive data, and constructing a structured material pool; a bridging module for generating an initial program list based on program template rules, and injecting random entropy at the junctions of section items to generate bridging content; an optimization module for performing logical verification and dynamic adjustment on the bridging content to obtain an optimized program list; a connecting phrase module for generating connecting phrases that match the optimized program list based on a diffusion model, and optimizing the connecting phrases; and a scoring module for scoring the optimized program list and connecting phrases, automatically filtering out high-risk content, and forming a safe broadcast program list for broadcast.

[0170] Example 2 is an embodiment of the present invention, which provides an intelligent scheduling method for broadcast programs. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0171] First, a 14-day comparative experiment was conducted using a real, operating provincial-level music information radio station. Before the experiment, based on the station's original program recordings, scripts, and interaction logs from the past three years, 5280 music tracks, 1140 short news snippets, 960 traffic updates, 820 weather alerts, 740 advertising materials, and 1360 auxiliary audio elements such as trailers and dialogue were collected and organized. Simultaneously, a total of 452,384 listener interaction records, including song requests, messages, comments, and polls, were compiled. Using a batch tagging script, all materials were labeled with four levels of tags: "type, duration, style, and copyright status." Interaction records were then archived based on "behavioral category, timestamp, and emotional polarity," and finally imported into a PostgreSQL+ElasticSearch hybrid database, forming a 24GB structured material pool. Subsequently, based on the radio station's current programming practices, a reusable program template was designed, comprising six sections: "Hourly News Briefs - Music Interlude - News Segment - Commercials - Weather - Music Returns." An initial program schedule was automatically generated daily from 00:00 to 24:00 in 30-minute increments. The system calculates the interactive diversity entropy in real time. Entropy of content diversity The bridging entropy intensity is obtained at the plate junction. ;when When the threshold is adjusted based on historical statistics, local perturbation and bridging content injection are enabled. For all bridging locations, the system further determines the bridging content based on causal weights. Joint control factor with linear synthesis of entropy intensity It performs version reselection, order fine-tuning, and bridging content enhancement.

[0172] In the concatenation generation stage, a 12-layer Chinese diffusion language model was first fine-tuned using a 1.20GB internal corpus; the system then reads in each concatenation... The positive noise figure is calculated according to... Correction( Samples in the high randomization region show a 23% increase in positive noise; those in the low randomization region maintain the baseline. The model uses... After completing the denoising training, the inference phase generates six candidate linking phrases in parallel for each bridging position. The candidate texts are then scored using a three-component system: diversity, causal orientation, and depth. Texts scoring below a threshold of 0.45 are directly discarded. The remaining texts are then subjected to a dual screening process using a retention rate prediction model (XGBoost, RMSE 0.04) and semantic consistency, ultimately retaining the highest-scoring linking phrase for the program guide. The risk filtering part is executed in parallel using a dictionary, CNN, and copyright information filtering channel. Training accuracy... The non-compliance recall rate was 91.86%. Every night at 23:00, the system starts a batch review process. High-risk (70 points) items are automatically replaced, and medium-risk (40-70 points) items are pushed to the human client. Two reviewers complete the review at an average speed of 1 minute and 48 seconds per item, and finally generate a safe program list that can be directly broadcast and controlled.

[0173] During the 14-day official broadcast, the new system output a total of 6,720 program entries, including 560 bridging points between segments; the average random entropy intensity per bridging point was 0.47. The system generated 3,360 candidate connecting phrases, of which 560 were retained after multi-dimensional screening, a retention rate of 16.67%. Compared with traditional manual scripts, the average length of a single connecting phrase was reduced from 84.13 characters to 62.70 characters. In the program schedule risk assessment phase, 78 high-risk entries (1.16%) were detected, all of which were automatically and seamlessly replaced; 214 medium-risk entries (3.18%) were reviewed manually, of which 158 were retained and 56 were edited; 6,428 low-risk entries (95.66%) required no processing. After the final safe program schedule was broadcast through the broadcast control system, the average retention rate for the entire segment was 88.54%, compared to 76.03% for the control group (traditional manual arrangement), representing an improvement of 16.44%. The program recording error rate decreased from 0.45% to 0.08%; the copyright conflict hit rate decreased from 0.38% to 0.05%. Automated scheduling reduced the time for generating a daily program schedule from an average of 42.70 minutes to 6.85 minutes, saving 83.96% of manual time.

[0174] The above data clearly demonstrates that this AI-driven broadcast orchestration solution outperforms traditional manual processes in multiple dimensions. Firstly, in terms of content security, the system accurately categorizes and promptly addresses high, medium, and low risks, with high-risk items accounting for only 1.16%, a reduction of 62.80% compared to the baseline process. This reduction is attributed to the triple fusion of dictionary-CNN-copyright: dictionary scanning ensures immediate blocking of clearly infringing content, CNN achieves 95.73% accuracy in capturing semantic-level risks, and the copyright interface plugs potential authorization loopholes, ultimately reducing the copyright conflict rate to 0.05%, only one-seventh of that in the manual era. Secondly, the user experience is significantly improved. The system controls bridging randomness and retention orientation through an entropy-causality model: on average... This indicates that moderate randomization was introduced at approximately half of the bridging points, but the causal weights... Timely suppression of excessive jumps increased the retention rate from 76.03% to 88.54%, an equivalent improvement of 16.44%. In an audience pool of 386,000, this translates to an average daily retention of approximately 48,300 more listeners. Simultaneously, the average length of connecting phrases decreased by 25.43%, achieving increased information density and reduced redundancy, indirectly confirming the generative advantages of the diffusion model in balancing diversity and simplicity. Furthermore, operational efficiency improved exponentially. The automatic generation time for a complete program schedule decreased from 42.70 minutes to 6.85 minutes, saving 83.9696 scheduling time; the collaboration between automatic replacement and manual review reduced the average workload per reviewer to 56 items per day, a 60.00% reduction from the baseline of 140 items per day. The saved manpower can be further invested in local interviews and in-depth content production, creating a virtuous cycle. Finally, both content quality and broadcast stability were enhanced. Disorderly behaviors that disrupted continuity were mitigated by sequence disturbances. The automatic bridging significantly reduced thematic deviation between program segments by an average of 18.27%, and decreased the film delay and dead play rate to 0.08%. In the hourly refresh segment, which is of great sensitivity to radio listeners, the automatic bridging kept the switching error within 0.12 seconds, which is 71.43% more robust than manual switching.

[0175] It can be seen that the four-layer linkage mechanism of structured material pool + entropy-causality adaptive scheduling + diffusion script generation + three-channel risk review not only solves the pain points of traditional scheduling mode such as lack of response to real-time interaction, time-consuming and error-prone manual scripts, and lagging violation detection, but also builds a closed-loop link covering "content generation - decision scheduling - security review - broadcast control implementation". Through comparison of multi-dimensional test data, it shows significant and credible advantages in key indicators such as retention rate, timeliness, compliance rate and operating cost.

[0176] This embodiment also provides a computer device applicable to an intelligent scheduling method for broadcast programs, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the intelligent scheduling method for broadcast programs as proposed in the above embodiment.

[0177] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0178] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the intelligent scheduling method for a broadcast program as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0179] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent scheduling of broadcast programs, characterized in that, include: Collect broadcast materials and interactive data, and build a structured material pool; An initial program list is generated based on program template rules, and random entropy is injected at the junction of section items to generate bridging content. Perform logical verification and dynamic adjustment on the bridging content to obtain an optimized program schedule; Generate and optimize program schedule matching based on a diffusion model, and optimize the program schedule matching. The optimized program schedule and connecting phrases are scored, high-risk content is automatically filtered out, and a safe broadcast program schedule is generated for radio broadcasting. The process of generating the initial program list includes extracting corresponding candidate material sets from the structured material pool according to the type, duration, and playback order constraints of each section and section item in the program template. The interaction diversity entropy is calculated based on the proportion of each type of user interaction in the total interaction, and the proportion of each subtype of material in each candidate material set is statistically analyzed to obtain the content diversity index. The content diversity index and the interaction diversity entropy are linearly weighted according to preset weights to obtain the bridging entropy intensity value. When the bridging entropy intensity value is not lower than a preset threshold, a local perturbation is performed on each pair of adjacent plate items; the perturbation includes mapping the bridging entropy intensity value to a perturbation probability according to a preset rule; generating a random number for each pair of adjacent plate items and comparing it with the perturbation probability; if the random number is not higher than the perturbation probability, the order is swapped. And ensure that after the exchange, the number of consecutive occurrences of the same type of section item does not exceed the upper limit, and the program list still conforms to the template order constraints; Before perturbation, if the bridging entropy strength value is not lower than the insertion threshold or the content difference between adjacent segment items is not lower than a preset threshold, bridging content is inserted between adjacent segment items. The bridging content includes at least one of impromptu connecting words, interactive summaries, hot news flashes, or sound effect clips. The length and tone of the bridging content are determined by interpolation of the bridging entropy strength value within their respective preset intervals. The segment item sequence after perturbation and bridging content insertion is used as the initial program list. The process of obtaining the optimized program schedule includes extracting the bridging entropy intensity value for each bridging location from the initial program schedule. The bridging entropy strength value and the interaction data at the corresponding bridging location are input into the causal inference model, and the causal weights for each bridging location are obtained from the causal inference model. The bridging entropy strength value and causal weight are combined proportionally to generate a joint control factor for each bridging position; all bridging positions are traversed sequentially, and each bridging position is optimized and adjusted according to the numerical range of the joint control factor. The optimization and adjustment include version reselection, sequence perturbation, and bridging content enhancement; after traversing and completing the dynamic adjustment of all bridging positions, an optimized program list is generated and output.

2. The intelligent scheduling method for broadcast programs as described in claim 1, characterized in that: The broadcast materials include music, news, weather, traffic conditions, advertisements, stories, dialogue, trailers, long programs, short programs, opening remarks, and closing remarks; the interactive data includes song requests, messages, comments, polls, and audio uploads submitted by users through mini-programs, applications, or websites. The construction of the structured content pool includes tagging the broadcast content, archiving the interactive data, and aggregating all tagged broadcast content and archived interactive data into a database to form a structured content pool that supports multi-condition retrieval, automatic arrangement, and AI content generation.

3. The intelligent scheduling method for broadcast programs as described in claim 2, characterized in that: The concatenation of generating and optimizing program schedule matching includes extracting the joint control factor of each bridging position in the optimized program schedule, and modifying the basic noise coefficient according to the value of the joint control factor before noise injection in the forward diffusion process. In each iteration of reverse denoising, the joint control factor is used as the gradient correction weight to adjust the gradient output of the noise prediction network, and the corrected gradient is superimposed on the denoising result. During the sampling phase, based on the default number of denoising iterations and the default sampling step size of the diffusion model, the joint control factor is compared with a preset threshold. When the joint control factor is higher than the preset threshold, the default number of denoising iterations is increased to the integer part of the product of the default number of iterations plus the difference between the joint control factor and the threshold and the predetermined iteration increment. The sampling step size is shortened by calculating the difference between the joint control factor and the preset threshold, multiplying the difference by a predetermined distance reduction coefficient to obtain the step size reduction magnitude. The result of subtracting the step size magnitude from the default sampling step size is used as the new sampling step size. When the joint control factor is lower than the threshold, the default number of iterations and sampling step size are maintained.

4. The intelligent scheduling method for broadcast programs as described in claim 3, characterized in that: Optimizing the concatenation includes calculating a diversity score, a causal guidance score, and a generation depth score for each candidate concatenation based on the recorded noise injection ratio, gradient correction weights, and sampling parameters. The diversity score is obtained by calculating the information entropy of the word frequency distribution of the candidate concatenation and normalizing it according to the total text length. The causal guidance score is obtained by taking the arithmetic mean of the gradient correction weights for the candidate concatenation in all denoising iterations. The generation depth score is obtained by subtracting the preset minimum number of iterations from the actual number of denoising iterations for the candidate concatenation, dividing the difference by the difference between the maximum number of iterations and the minimum number of iterations, and using the result as the normalized generation depth score. The diversity score, causal orientation score, and generation depth score are linearly combined according to pre-set weights to obtain a comprehensive score, and candidate concatenations with a comprehensive score lower than the minimum passing threshold are eliminated. The audience retention rate is defined as the ratio of the number of remaining listeners at the end of a segment to the number of listeners at the beginning of that segment. Based on historical broadcast data, the average retention rate under different bridging positions, similar contexts, and text features is statistically obtained, and a retention rate prediction model is constructed. Each candidate concatenation phrase, along with its corresponding joint control factor and contextual features, is input into the retention rate prediction model to calculate the expected retention rate of the candidate concatenation phrase, and candidate concatenation phrases with expected retention rates lower than the historical average retention rate are eliminated. The semantic similarity of the remaining candidate concatenation phrases is compared with the preceding and following text, and candidate concatenation phrases with semantic similarity lower than a preset consistency threshold are eliminated. The final remaining concatenation phrases that meet the contextual coherence requirements and their corresponding joint control factors are used as the final broadcast script and submitted together with the optimized program schedule.

5. The intelligent scheduling method for broadcast programs as described in claim 4, characterized in that: The process of creating a safe broadcast program list includes reviewing the content of each segment's text and connecting phrases based on a prohibited word library, a pre-trained deep classification model, and a copyright database interface, and generating a risk score. Risk scores are categorized according to preset grading rules, and high-risk content is automatically replaced while medium-risk content is submitted for manual review. During automatic replacement, matching content is selected from the safety candidate library based on the original material's tags, type, and duration. After manual review, the review results are updated to the program schedule to form a safe broadcast program schedule.

6. An intelligent scheduling system for broadcast programs, based on the intelligent scheduling method for broadcast programs according to any one of claims 1 to 5, characterized in that, include: The data module collects broadcast materials and interactive data, and builds a structured material pool; The bridging module generates an initial program list based on program template rules and injects random entropy at the junctions of section items to generate bridging content; the optimization module performs logical verification and dynamic adjustment on the bridging content to obtain an optimized program list. The linking phrase module generates and optimizes the linking phrases based on the diffusion model to match the program schedule; the scoring module scores the optimized program schedule and linking phrases, automatically filters out high-risk content, and forms a safe broadcast program schedule for broadcast.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the intelligent scheduling method for broadcast programs according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent scheduling method for broadcast programs as described in any one of claims 1 to 5.

Images