Intelligent arrangement method and system for broadcast programs

Abstract

The invention discloses an intelligent broadcast program arrangement method and system, and relates to the technical field of intelligent broadcast program arrangement, and the method comprises the steps: collecting broadcast materials and interaction data, and constructing a structured material pool; generating an initial program list based on a program template rule, and injecting random entropy at the joint of plate items to generate bridging content; performing logic verification and dynamic adjustment on the bridging content to obtain an optimized program list; on the basis of a diffusion model, series words matched with the optimized program list are generated, and the series words are optimized; and scoring the optimized program list and the serial words, automatically screening out high-risk contents, and forming a safe broadcast program list for broadcasting. Second-level compliance verification and safe screening are realized, the manual auditing cost is greatly reduced, and the broadcasting efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent broadcast programming, and in particular to an intelligent programming method and system for a broadcast program. BACKGROUND

[0002] Current broadcast programming mainly relies on pre-defined program templates and manual configuration processes. Users maintain calendarized program templates through a web interface, specify time announcements, dialogues, news, music, advertisements, and clip items in each time period, and select material sources and broadcast methods for each block item. This method can quickly generate daily program schedules, but still relies on fixed concatenated phrases or preset clips at the block item junction, lacking dynamic adjustment capabilities for content transitions.

[0003] In actual operation, program schedules often face uncertainty events such as a surge in audience interaction, breaking news, or road condition information interjections, which can disrupt the original block order, making it difficult to balance coherence and diversity with fixed transitions. At the same time, existing programming lacks real-time logical verification and optimization means for transition content, and cannot actively regulate the connection strategy based on user behavior data or content entropy indicators, resulting in a less smooth program experience. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an intelligent programming method for a broadcast program, which solves the problems of existing broadcast programming systems in dynamic diversified connection, online causal verification and optimization, and automatic compliance risk management, such as the inability to automatically generate diversified transition content, the lack of real-time causal effect evaluation mechanism, and the difficulty in achieving second-level semantic compliance and automatic exclusion of copyright risks.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides an intelligent programming method for a broadcast program, which includes collecting broadcast materials and interaction data, and constructing a structured material pool;

[0008] Generating an initial program schedule based on program template rules, and injecting random entropy to generate bridging content at the block item junction;

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

[0010] Generating concatenated speech matching the optimized program schedule based on a diffusion model, and optimizing the concatenated speech;

[0011] Scoring the optimized program schedule and the concatenated speech, automatically excluding high-risk content, and forming a safe broadcast program schedule for broadcast.

[0012] As a preferred scheme of the intelligent arrangement method of the broadcast program, wherein: the broadcast material includes music, news, weather, traffic, advertisement, story, dialogue, clip, long program, short program, opening speech and closing speech; the interactive data includes song ordering, message, comment, voting and audio uploading submitted by users through applet, application or website;

[0013] The constructing the structured material pool comprises: label processing on the broadcast material, archiving the interactive data, and summarizing all the labeled broadcast material and archived interactive data into a database to form a structured material pool supporting multi-condition retrieval, automatic arrangement and AI content generation.

[0014] As a preferred scheme of the intelligent arrangement method of the broadcast program, wherein: the generating the initial program list comprises: extracting a corresponding candidate material set from the structured material pool according to the type, duration and playing order constraint of each block and block item in the program template; calculating the interactive diversity entropy according to the proportion of each type of user interaction in the total interaction, and calculating the content diversity index by counting the proportion of each sub-type material in each candidate material set; linearly weighting the content diversity index and the interactive diversity entropy according to a preset weight to obtain a bridge entropy intensity value;

[0015] When the bridge entropy intensity value is not less than a preset threshold, performing local disturbance on each pair of adjacent block items; the disturbance comprises: mapping the bridge entropy intensity value to a disturbance probability according to a preset rule; generating a random number for each pair of adjacent block items and comparing it with the disturbance probability, if the random number is not higher than the disturbance probability, then exchanging the order; and ensuring that after the exchange, the number of consecutive appearances of the same type of block item does not exceed the upper limit, and the program list still meets the template order constraint;

[0016] Before performing the disturbance, when the bridge entropy intensity value is not less than an insertion threshold or the content difference degree of adjacent block items is not less than a preset threshold, then inserting bridge content between adjacent block items; the bridge content includes at least one of improvised concatenated words, interactive abstracts, hot news or sound clips, and the length and tone style of the bridge content are determined by interpolation within their respective preset intervals; the sequence of block items after completing the disturbance and bridge content insertion is taken as the initial program list.

[0017] As a preferred scheme of the intelligent arrangement method of the broadcast program, wherein: the obtaining of the optimized program list comprises: extracting the bridge entropy intensity value of each bridge position from the initial program list; inputting the bridge entropy intensity value and the interaction data corresponding to the bridge position into the causal inference model, and obtaining the causal weight for each bridge position from the causal inference model; proportionally combining the bridge entropy intensity value and the causal weight to generate the joint control factor of each bridge position; sequentially traversing all bridge positions, and optimizing and adjusting each bridge position according to the numerical range of the joint control factor, wherein the optimization and adjustment comprises version reselection, sequence disturbance and bridge content enhancement; and after the dynamic adjustment of all bridge positions is completed, the optimized program list is generated and output.

[0018] As a preferred scheme of the intelligent arrangement method of the broadcast program, wherein: the generating of the concatenated language matching the optimized program list comprises: extracting the joint control factor of each bridge position in the optimized program list, and modifying the basic noise coefficient according to the numerical value of the joint control factor before noise injection in the forward diffusion process; in each reverse denoising iteration, 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.

[0019] In the sampling stage, the joint control factor is compared with the preset threshold value based on the default denoising iteration number and the default sampling step length of the diffusion model, when the joint control factor is higher than the preset threshold value, the default denoising iteration number is increased to an integer part of a value obtained by adding the joint control factor and the threshold value difference to the predetermined iteration increment, and the sampling step length of each step is shortened, that is, the difference between the joint control factor and the preset threshold value is calculated, the difference is multiplied by a predetermined distance reduction coefficient to obtain the step length amplitude to be reduced; the result obtained by subtracting the step length amplitude from the default sampling step length is used as the new sampling step length; when the joint control factor is lower than the threshold value, the default iteration number and the sampling step length are maintained.

[0020] As a preferred scheme of the intelligent arrangement method of the broadcast program, wherein: the optimization of the concatenated language comprises: calculating a diversity score, a causal orientation score and a generation depth score for each candidate concatenated language according to the recorded noise injection multiple, the gradient correction weight and the sampling parameter; the diversity score is obtained by calculating the information entropy of the word frequency distribution of the candidate concatenated language and normalizing the total length of the text; the causal orientation score is obtained by calculating the arithmetic mean of the gradient correction weight of the candidate concatenated language in all denoising iterations; the generation depth score is obtained by subtracting the preset minimum iteration number from the actual denoising iteration number of the candidate concatenated language, and then dividing the difference by the difference between the maximum iteration number and the minimum iteration number, and the result is the normalized generation depth score; the diversity score, the causal orientation score and the generation depth score are linearly combined according to the pre-set weight to obtain a comprehensive score, and the candidate concatenated language with a comprehensive score lower than the minimum passing threshold is removed.

[0021] The audience retention rate is defined as the ratio of the remaining audience number at the end of the program segment to the audience number at the beginning of the program segment; based on historical broadcast data, the average retention rate under different bridge positions, similar contexts and text features is obtained, and a retention rate prediction model is constructed; each candidate concatenated language and the corresponding joint control factor and the context feature are input into the retention rate prediction model to calculate the predicted retention rate of the candidate concatenated language, and the candidate concatenated language with a predicted retention rate lower than the historical average retention rate is removed; the remaining candidate concatenated language is compared with the previous segment and the next segment of text in terms of semantic similarity, and the candidate concatenated language with a semantic similarity lower than the preset consistency threshold is removed; the final remaining concatenated language that meets the context continuity and the corresponding joint control factor are used as the final broadcast script, which is submitted together with the optimized program list.

[0022] As a preferred scheme of the intelligent arrangement method of the broadcast program, wherein: the formation of the safe broadcast program list comprises: based on the violation word library, the pre-trained deep classification model and the copyright database interface, the content of each block item text and concatenated language is audited to generate a risk score; the risk score is classified according to the pre-set classification rule, and the high-risk content is automatically replaced and the medium-risk content is submitted for manual audit; during automatic replacement, the original material is replaced by the matching content selected from the safe alternative library according to the label, type and length of the original material; after the manual audit is completed, the audit result is updated to the program list to form the safe broadcast program list.

[0023] In a second aspect, the present application provides an intelligent programming system for a broadcast program, comprising: a data module that collects broadcast materials and interactive data, and constructs a structured material pool; a bridging module that generates an initial program list based on program template rules, and injects random entropy at the junction of block items to generate bridging content; an optimization module that performs logical verification and dynamic adjustment on the bridging content to obtain an optimized program list; a concatenation module that generates concatenation language matching the optimized program list based on a diffusion model, and optimizes the concatenation language; and a scoring module that scores the optimized program list and the concatenation language, automatically filters out high-risk content, and forms a safe broadcast program list for broadcast.

[0024] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the computer program, when executed by the processor, implements any step of the intelligent programming method for a broadcast program according to the first aspect of the present application.

[0025] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the intelligent programming method for a broadcast program according to the first aspect of the present application.

[0026] The present application has the following advantages: by injecting random entropy at the junction of block items, the dynamic diversification of transition content is realized, and the freshness of the audience is improved compared to traditional fixed concatenation language; relying on online causal reasoning to evaluate the impact of transition operation on retention rate in real time, and automatically adjusting the bridging strategy through joint control factors to ensure program coherence while achieving optimal user experience; the improved diffusion model adaptively applies control factors in each noise injection and denoising iteration, so that the AI concatenation language style and length accurately match the context and audience preferences; finally, combining a deep classification model and a copyright database interface to perform risk scoring and automatic replacement on the program list and the concatenation language, completing second-level compliance verification and safe filtering, greatly reducing the cost of manual review and improving the efficiency of broadcast. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0028] Figure 1 A flowchart of an intelligent programming method for a broadcast program. DETAILED DESCRIPTION

[0029] In order to make the above objectives, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0030] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. The present application, however, can be practiced in a variety of ways other than those specifically described herein, and the scope of the present application is not limited to the specific embodiments described herein.

[0031] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an embodiment that is independent or alternative to other embodiments.

[0032] Embodiment 1, refer to Figure 1 For one embodiment of the present application, the embodiment provides an intelligent programming method of a broadcast program, 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, advertisements, stories, dialogues, clips, long programs, short programs, opening remarks and closing remarks; the interactive data includes song requests, messages, comments, votes and uploaded audios submitted by users through applets, applications or websites.

[0035] Building a structured material pool includes: performing tagging processing on the broadcast materials, adding the following label information to each material and storing it as a field attribute: column category (news, music, talk show, advertisement, etc.); emotional style (relaxed, serious, enthusiastic, etc.); play time period adaptation (morning rush hour, night, etc.); content length interval (≤30 seconds, 30 seconds-2 minutes, etc.); AI broadcast readability (only human, AI can broadcast, both); recommended priority (ordinary, priority, strong push); archiving interactive data, structuring and recording song requests, messages, comments, votes and uploaded audios, and adding labels such as interactive type, timestamp and user preference.

[0036] All tagged broadcast materials and archived interactive data are summarized into a database to form a structured material pool that can support multi-condition retrieval, automatic programming and AI content generation, for subsequent program list programming, entropy injection and diffusion content generation links.

[0037] Unified collection and preprocessing of broadcast materials and interactive data, tagging each material with multiple dimensions such as program category, emotional style, play time adaptation, content length interval, AI broadcast readability, and recommendation priority, etc., so that the localized resource library has rich semantic levels for the first time. Compared with the traditional material pool classified only by type or simple length, this scheme can quickly locate the fine-grained content that meets the program requirements in the retrieval stage, greatly reducing the manual screening time. For example, in the morning rush hour, only by querying the "music type" "energetic rhythm" "≤2 minutes" "AI broadcastable" "priority" tags, 120 candidate materials that meet the conditions can be quickly obtained.

[0038] All interactive operations, such as song ordering, messages, comments, voting, and audio uploading, are structured into the same database according to interaction type, timestamp, user ID, and preference tags. By comparing the proportion of each interaction in the total amount, audience engagement indicators can be calculated in real time, providing dynamic parameters for subsequent entropy injection. Unlike the need for manual review of logs or analysis reports in the past, this interaction archiving method forms a closed loop between hot behavior in the time window and program materials, ensuring that the content inserted during the peak interaction period meets user needs and improves the fit between the program and the audience.

[0039] Further, the entire material pool is supported by Elasticsearch+relational database dual engines, supporting multi-condition compound queries and parallel retrieval, and combining with timing tasks and triggers to achieve seamless connection of the "automatic arrangement-entropy injection-diffusion generation" process. In the face of large-scale, multi-format, and high-concurrency scheduling scenarios, the bottleneck of previous systems in real-time, accuracy, and scalability is effectively solved.

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

[0041] Program template : Contains blocks , each block has several block items , and specifies the content type and play order constraints of each item.

[0042] Structured material pool : Prepare multiple versions for each type of music, news, advertising, story, etc., and add a set of tags to each material:

[0043]

[0044] User behavior set : Count of different interaction types, such as song requests, within a specified time window , message counts , comment counts , etc.

[0045] Environmental context features : Such as holiday identifiers, emergency event types, etc.

[0046] Real-time interaction intensity : Block item Corresponding number of related interactions in the last five minutes.

[0047] Wherein, represents a set of program templates; represents the total number of blocks contained in the template; represents the block (such as a music segment, an advertisement segment, a story segment); represents the block item (i.e., the play unit) in the block; represents the block index, with a value range of 1 to ; represents the order index of the block item in the block ; represents the set of all block items under the block ; represents a structured material pool; represents a set of labels attached to each material; type represents the material content type (such as music, advertisement, news, etc.); duration represents the material duration; style represents the material style (such as happy, calm, formal, etc.); copyright represents the material copyright status label; represents a set of user interaction behaviors; represents the number of song request behaviors; represents the number of message behaviors; represents the number of comment behaviors; represents the current external environmental context features.

[0048] Candidate material set extraction extracts, for each block item in the template, a candidate set from the according to its type label:

[0049]

[0050] Wherein, represents the candidate material set corresponding to the block item ; represents a single material in the material pool; type​ representing the content type label; type representing the block item the content type label required; representing the template the first block the first item content.

[0051] Calculate the total number of interactions:

[0052]

[0053] and get the interaction diversity entropy:

[0054]

[0055] where, representing the total number of interactions; representing the number of the class of interaction behavior; representing the user interaction diversity entropy; representing the proportion of the class of interaction behavior in the total interaction; representing the class number of user interaction behavior.

[0056] Group the candidate set by sub-type , and record the number of elements in each group as , and the total number is:

[0057]

[0058] Calculate the content entropy:

[0059]

[0060] where, representing the total number of materials in the candidate set ; represent the number of materials in the candidate set belonging to the sub-type ; represent the index of the sub-type (such as style, language); represent the content entropy of the block item .

[0061] Build a bridge vector:

[0062]

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

[0064] ​​

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

[0066] Candidate set in Version: If , then the probability of random selection is equal; otherwise, the first Version number.

[0067] To adjacent Item Collection ;like , then the probability for each pair of adjacent items is:

[0068]

[0069] generate ,like The two positions are swapped; if the number of consecutive occurrences of the same type exceeds the upper limit before the swap , then skip the exchange.

[0070] in, represents the execution probability of the local perturbation; represents the minimum value function; represents a random variable; Represents a uniform distribution from 0 to 1.

[0071] In each pair of adjacent If or thematic difference , then insert bridging content, which can be impromptu links, interactive summaries, hot news or sound effects trailers.

[0072] The length of the bridge segment is:

[0073]

[0074] Styles can be switched between two preset templates. Interpolation. If , then enter high entropy mode. Expand the version candidate range and perturb the window increased to , the interactive summary or hot news is inserted preferentially.

[0075] wherein, represents the actual generated length of the current bridging content; represents the preset minimum bridging length; represents the preset maximum bridging length; represents the maximum function; represents the intensity threshold for triggering the high-entropy mode; represents the subject difference threshold; represents the version selection threshold; represents the sequence disturbance threshold; represents the bridging insertion threshold.

[0076] The final output program list is:

[0077]

[0078] wherein, represents the optimized complete program list; represents the program content optimized by the th item; represents the total number of items in the optimized program list.

[0079] Read the initial program list, wherein each item contains: material label, bridging vector , entropy intensity factor .

[0080] Obtain real-time audience interaction data ( is the interaction category index) and the current environmental context label .

[0081] Maintain a simplified causal model, nodes include each bridging decision (operation on position ) and effect indicators (audience retention rate). Collect user retention rate or interaction rate data in real time within a sliding window, calculate the incremental causal gain for each position :

[0082]

[0083] wherein, represents the causal gain value of the bridging position ; represents the user retention rate or interaction rate indicator; represents the position a binary variable indicating whether to perform a bridging operation; denotes the mathematical expectation; do denotes the intervention condition for enforcing operations at the bridging position; do denotes the condition for not performing a bridging operation. And through online Bayesian filtering smooth update, suppress noise.

[0084] Map to the causal feedback factor:

[0085]

[0086] Combine with the feedback factor , calculate the joint control factor:

[0087]

[0088] where, denotes the causal feedback factor of the bridging position , the value range is [0, 1]; denotes the hyperparameter for controlling the slope of the causal feedback curve; denotes the joint control factor of the bridging position ; denotes the fusion weight of the entropy factor, the value range is [0, 1]; denotes the entropy intensity factor of the bridging position .

[0089] Define the utility function at each position :

[0090]

[0091] where, denotes the multi-objective utility value of the control factor at position ; denotes the weight hyperparameter of the causal gain term; denotes the penalty weight of the continuity bias term; denotes the reward weight of the diversity promotion term; denotes the adjacent content theme continuity bias at position ; denotes the diversity increment factor at position .

[0092] Approximate the solution for each by gradient ascent:

[0093]

[0094] The iteration number is controlled within 1-3 times to ensure the single-point decision delay is lower than 5ms.

[0095] For each position , according to the optimal joint factor , the following is executed by comparing with a preset threshold value:

[0096] When , version reselection is performed, the current version is removed from the candidate material list at position , and the version with the highest causal gain is selected for replacement.

[0097] When , sequence disturbance is performed, and the next item is swapped with a probability ; if it will lead to continuous over-limit of the same type , it is skipped.

[0098] When , bridge content enhancement is performed, and a bridge section is inserted at position ; the bridge length is determined according to , and the type is preferentially selected as interactive summary or hot news.

[0099] When , a conservative strategy is selected, the version and sequence in S2 are retained, and no additional bridge is inserted. In this embodiment: , , .

[0100] wherein represents the optimal control factor solved at position ; represents the control factor threshold value of version reselection; represents the control factor threshold value of sequence disturbance; represents the control factor threshold value of bridge content enhancement; represents the minimum insertion length of bridge content; represents the maximum insertion length of bridge content; represents the maximum allowable upper limit of continuous occurrence of items of the same type.

[0101] The positions to are sequentially traversed, all bridge positions are sequentially traversed, and the above dynamic adjustment rules are applied and updated in real time; after each update, the decision log is written, and the selected version identifier, whether the sequence is disturbed, the bridge section type and length are recorded; after all position adjustments are completed, the final optimized program list is output: ​

[0102]

[0103] wherein, represents the final optimized program list entry located at the position after traversal and adjustment, each item is accompanied by an updated material identifier, order status, bridging content and , , , .

[0104] In the program generation stage, taking the pre-defined block type and item order in the template as a reference, the system automatically extracts a candidate material set matching each block item, and injects diversified transition content at the junction of adjacent block items according to the random entropy intensity. The injected bridging elements include improvised concatenated words, interactive summaries, hot news or sound effects, and the insertion timing and length of each type of content are dynamically determined by the real-time calculated entropy intensity and content difference, to ensure that the transition is neither harsh nor redundant, and to disrupt the material order in time to create a fresh feeling. This process is completely automated, both respects the play order constraints and provides flexible content insertion in high-interactive periods or thematic nodes, forming the first version of the initial program list.

[0105] Compared with the traditional arrangement method which only relies on static label matching or manual fine-tuning, this scheme realizes controllable disturbance of content order and dynamic generation of multiple version candidates within the same block through random entropy injection, significantly improving program coherence while considering diversified experience. The system detects the number of consecutive occurrences of the same type of content before each disturbance, avoiding the repetition that is easily overlooked in manual operation; when the content difference is large, hot summaries are preferred, avoiding the problem of slow response to sudden events in manual arrangement, so as to achieve more accurate content recommendation and transition during the peak of audience attention.

[0106] After completing the initial bridging, a set of lightweight causal reasoning model is used to analyze the entropy intensity and interaction data at all bridging positions, generate a joint control factor and perform further fine-tuning accordingly. This fine-tuning operation includes version reselection, order fine-tuning and bridging enhancement, each of which is carried out on the premise of ensuring the original template order logic, avoiding subjective bias of manual experience and overcoming the possible logic confusion caused by pure random strategy. Through this double closed-loop scheduling of entropy and causality, both the diversity of content and the audience retention effect are optimized, providing a reliable, efficient and flexible solution for real-time broadcast arrangement.

[0107] S3: generating and optimizing concatenated speech matching the program list based on the diffusion model, and optimizing the concatenated speech.

[0108] obtaining an optimized program list , each entry is accompanied by an entropy intensity factor , and joint control factors.

[0109] In the forward phase of each diffusion round, the standard noise coefficient Adjusted to:

[0110]

[0111] in, represents the adjusted noise figure under the modulation of the joint control factor; Represents the diffusion model The standard noise coefficient of the wheel; Represents the influence weight coefficient of the joint control factor; Indicates the The combined control factor for the bridge positions.

[0112] exist When high, more randomness is introduced. Keep text stable when low.

[0113] When returning to the denoising stage, Added as gradient correction weights:

[0114]

[0115] in, Indicates the The diffusion latent variable of the step; Represents the first Step latent variable; Indicates that the model The noise distribution of the step estimate; represents the adjustment coefficient of the causal gradient guidance term; Represents the causal scoring function pair gradient.

[0116] Automatically bias the diffusion model towards improving retention when generating it.

[0117] For each position , execute the following sampling process to generate Candidate concatenations :

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

[0119] Compare With preset threshold when When the new iteration number , new step size .

[0120] Otherwise, let , .

[0121] Use , the above and full run inverse diffusion, generate .

[0122] wherein, denotes the number of iterations of inverse diffusion after adaptive update; denotes the default number of iterations of denoising; denotes the control threshold for determining whether adjustment is needed; denotes the adjustment amplitude of the number of iterations corresponding to each unit factor increment; denotes the updated time step of inverse diffusion; denotes the default time step of diffusion sampling; denotes the step adjustment amplitude corresponding to each unit factor increment.

[0123] For each piece of record the corresponding , noise ratio , gradient weight, actual and .

[0124] The candidate serial language multi-dimensional score calculates three scores for each piece of

[0125] diversity score : information entropy normalization based on text word frequency distribution; causal orientation score : take the arithmetic mean of the gradient correction term ; generate depth score : map the actual number of iterations to [0, 1].

[0126] Then linearly combine the comprehensive score according to the weight

[0127]

[0128] Eliminate candidates with a total score below the minimum passing threshold. Among them, denotes the comprehensive score of the th candidate serial language; denotes the diversity score of the th; denotes the ​​Causal orientation score of the bar; Causal orientation score of the bar; Generation depth score of the bar; , , The weighted coefficients of the above three types of scores are respectively represented by

[0129] Let the number of listeners at the beginning of a program paragraph be , and the number of listeners at the end be The listener retention rate is the ratio of the two:

[0130]

[0131] Among them, represents the actual listener retention rate; represents the number of listeners at the beginning of the content broadcast; represents the number of listeners at the end of the broadcast.

[0132] From the past broadcast log, for each bridge position number , the same environment label and similar text features (such as word vector clustering results) collect retention rate samples; calculate the average retention rate and standard deviation of this group of samples as the basis for subsequent screening and threshold setting.

[0133] For each candidate serial string , construct a feature vector: joint control factor ; environmental context (One-Hot encoding); bridge vector , , ; text features, extract the 768-dimensional sentence vector of the pre-trained BERT model for . This embodiment uses Chinese BERT-base, and the sentence vector is 768-dimensional.

[0134] Gradient Boosting Regressor (XGBoost) is used, with parameter examples: learning rate 0.1, tree depth 6, tree number 200; training set size at least 100,000 historical records, validation set 20,000, test set 20,000; evaluate model performance with root mean square error (RMSE) and determination coefficient , target RMSE 0.05.

[0135] Input each candidate into the trained retention rate model to get the predicted retention rate . For all candidates i at each position m, if ​(or more strictly ), the candidate text is rejected, and the rest of the candidates that meet or exceed the historical average retention rate are retained.

[0136] BERT sentence vectors are extracted for the candidate concatenated sentence and its adjacent texts , , , , . The cosine similarity is calculated:

[0137]

[0138] Let the consistency threshold be . Among them, represents the cosine similarity between the candidate sentence and the previous program; represents the cosine similarity between the candidate sentence and the next program; represents the cosine similarity function between two vectors; represents the sentence vector (BERT extraction) of the candidate concatenated sentence; represents the program entry vector at position ; represents the program entry vector at position .

[0139] If or , the candidate is rejected; only the text that meets both the coherence of the front and back sections is retained. The remaining candidates are sorted by comprehensive score from high to low, and the first is selected. and the corresponding joint control factor are written back to the program entry When the processing of is completed, the final concatenated script sequence is output, and the optimized program list is submitted to the subsequent module for use.

[0140] Among them, represents the entropy intensity factor (S2 output) at position ; represents the th candidate concatenated sentence at position ; represents the finally selected concatenated sentence; represents the bridging position historical average retention rate; represents the standard deviation of the retention rate at the bridging position ; represents the predicted retention rate of the th candidate sentence; a minimum threshold value representing semantic consistency before and after; a total number of bridge positions in the program list; a position bridge linguistic context feature vector.

[0141] The optimized program list and the random entropy intensity and causal feedback factor corresponding to each position are read into the diffusion language model to jointly control the factor to dynamically adjust the forward noise injection intensity. In the forward diffusion stage, the noise ratio is automatically adjusted up or down according to the level of the joint control factor, so that more rich random disturbance is introduced at nodes that require to enhance diversity, and stability is maintained at nodes with high continuity requirements. Subsequently, a gradient guiding term is added in the denoising link to inject a causal gain signal into the gradient update, and the direction correction output by the noise prediction network is automatically biased towards the script style that can improve audience retention without excessively sacrificing transition naturalness. Such a strategy, which is different from the fixed setting of noise coefficient and gradient weight in traditional methods, realizes fine control of the generation process.

[0142] In the candidate serial language sampling stage, the system adjusts the number of iterations and the step size according to the joint control factor to generate multiple versions of the text in a parallel manner, and records the noise ratio, gradient weight and actual iteration parameters of each candidate. Subsequently, based on the scoring mechanism of word frequency information entropy, causal guidance intensity and iteration depth, all candidate texts are scored comprehensively to eliminate redundant or theme-deviating scripts. Then, the text vector extracted by the pre-trained language model is input into the retention rate predictor together with the environmental context and bridge features, and the most attractive script is selected through the regression model, while the semantic coherence of the front and back sections is checked to ensure that the final selected text is not only rich and diverse, but also logically consistent.

[0143] Compared with the existing methods of generating only by static templates or unified diffusion, this scheme realizes end-to-end adaptive control from noise injection, gradient guidance, sampling step to multi-level scoring, and no longer relies on fixed parameters set by humans. Therefore, it has significant advantages in retention rate improvement, generation efficiency, text diversity and content coherence: the retention effect is reliably improved, the candidate selection process remains efficient, and the matching degree of the final script and the program list and the audience experience are simultaneously optimized.

[0144] S4: Score the optimized program list and serial language, automatically filter out high-risk content, and form a safe broadcast program list for broadcast.

[0145] Read the optimized program list generated in S3 , wherein each entry contains: final material identifier, order status, bridge content, final serial language and joint control factor Synchronize S3 decision log (each record ) with S3 record log. Scan program entries from high to low, prioritize the strongest randomized, highest risk potential passages.

[0146] Clean up plate text and concatenated text (remove HTML, URL, control symbols).

[0147] Match word units with three-level sensitive word library by space or jieba segmentation:

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

[0149] Single text maximum score 30 points, score recorded as Deep CNN classification includes preprocessing and embedding, word-level segmentation for Chinese part, and space segmentation for English part; the longest 200 words are retained, and the insufficient part is filled with PAD.

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

[0151] Generate shape Embedding matrix as network input. Three parallel one-dimensional convolutions, kernel width 3, 4, 5, channel number 128 each; each convolution is followed by ReLU activation, then global maximum pooling to get 128-dimensional vector; three groups of pooling results are concatenated into 384-dimensional, and Dropout 0.5 is used to prevent overfitting, then enter 256-unit fully connected layer.

[0152] Output layer 3 nodes, Softmax produces "compliance / suspected violation / non-compliance" three types of probability. Core parameters include: word vector 300-dimensional, total convolution kernel number 384, hidden layer 256, model size <2MB.

[0153] Collect 500,000 pieces of broadcasted text manually labeled as three categories, 8:1:1 for training / validation / testing.

[0154] Optimizer Adam, initial learning rate 0.001, batch size 128, maximum 20 rounds; if the validation set does not improve for 3 rounds, stop early. Test set target: overall accuracy ≥95%, non-compliance recall rate ≥90%.

[0155] Model compression: 8-bit quantization + weight pruning, inference delay <10ms / text; take the "non-compliance" output probability , calculate:

[0156]

[0157] Call the copyright library interface one by one. If the entry contains multiple restricted materials, the highest risk score is used: authorized 0 points; pending confirmation or abnormal 10 points; unauthorized 20 points. .

[0158] The total risk score is calculated as:

[0159]

[0160] in, Indicates the Risk score based on sensitive dictionary hits in the content; Indicates that the CNN model determines the The probability value of the content being "non-compliant" is in the range of [0, 1]; Indicates the The machine learning risk score of the content; 50 is the linear mapping coefficient, converting the probability into a score range of 0-50; Indicates the The copyright risk score of the content; Indicates the The total risk score of the content. The maximum total risk score is 100, and the score is written into the program list entry for future use.

[0161] when When the risk is high, the candidate is retrieved from the safe alternative library based on the type, duration, and style label of the original material. The material that has been recently updated, has not been marked black, and has the highest cosine similarity with the original material and has sufficient inventory is selected for replacement. After replacement, ensure that the new ; At the same time, the concatenation language is replaced. First, search the safe concatenation template library according to the opening / transition / ending; if there is no match, use the universal transition sentence and postpone the original concatenation language to the next safe socket; the new text is also re-scored until low risk.

[0162] when When the risk is medium, the WEB audit table will display the risk details in batches. The auditor can release / edit / replace the program. The review results will be written back to the program list in real time, triggering automatic re-rating. After the review is passed, the risk score must also be <40. , it is low risk.

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

[0164] After automatic replacement and manual review, a safe broadcast program list is obtained , that is, all items with risk scores <40 will be directly pushed to the broadcast control system.

[0165] in, The final safe broadcast program list indicating that all program item risk scores are lower than 40.

[0166] In the broadcast preparation link, the system first uniformly cleans all program entries and corresponding serial language texts, removes redundant webpage tags and link information, then loads a three-level sensitive word library locally, and performs word-by-word scanning on each text in a hierarchical manner. Next, the deep convolutional network module classifies the processed text in terms of semantics. The network structure can complete inference in milliseconds while keeping the model size less than 2M, and combines with quantization technology to ensure operation efficiency. The copyright verification interface works in parallel, automatically compares the license status of the material and generates a risk mark, blocking unauthorized or abnormal items from the source.

[0167] In the content risk grading stage, the system aggregates the risk information obtained by scanning and classification, and automatically routes according to the evaluation results of "high risk", "medium risk" and "low risk". For high-risk content, immediately call the most optimal and most sufficient compliant material of the same type from the safe material library to replace, and synchronize in the serial language template library to match the appropriate transition text to replace; for medium-risk items, they are pushed to the online review interface in real time, which intuitively displays the risk reasons and the recommended replacement plan. After the reviewer confirms or modifies in a very short time, the system will automatically recalculate the risk score until it is compliant; low-risk items are directly retained and enter the subsequent broadcast process, ensuring uninterrupted program rhythm.

[0168] All link operations are automatically recorded in the audit database, including original data identification, risk label value before and after evaluation, replacement or release type, operation time and operator information. With the help of a visual monitoring panel, platform managers can flexibly query and export logs according to program categories, processing periods or specific operators, easily completing compliance checks and risk sampling. This closed-loop process realizes fully automated control from safety detection to material replacement to review retention, ensuring the compliance of broadcast content and minimizing the pressure on human review.

[0169] The embodiment also provides an intelligent arrangement system for a broadcast program, including: a data module that collects broadcast materials and interactive data and constructs a structured material pool; a bridging module that generates an initial program list based on program template rules and injects random entropy generation bridging content at block item junctions; an optimization module that performs logical verification and dynamic adjustment on the bridging content to obtain an optimized program list; a serial language module that generates serial language matching the optimized program list based on a diffusion model and optimizes the serial language; and a scoring module that scores the optimized program list and the serial language, automatically filters out high-risk content, and forms 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 fourteen-day controlled experiment was conducted on a real-world provincial music information radio station. Based on the station's original program recordings, scripts, and interaction logs from the past three years, the experiment collected and organized 5,280 pieces of music, 1,140 news shorts, 960 traffic information, 820 weather alerts, 740 advertising materials, and 1,360 auxiliary phonemes such as trailers and dialogue lines. A total of 452,384 interactive records, including listener requests, messages, comments, and votes, were also compiled. Using a batch tagging script, all material was annotated with four levels of labels: "genre, duration, style, and copyright status." Interaction records were then archived by "behavior category, timestamp, and sentiment polarity," and finally imported into a PostgreSQL+ElasticSearch hybrid database, creating a 24GB structured material pool. Then, based on the current programming conventions of radio stations, a reusable program template was designed, which includes six sections: "Hourly News - Music Playback - News Interlude - Advertisement - Weather - Music Return". The initial program list was automatically generated at 30-minute granularity for 00:00-24:00 every day. The system calculates the interactive diversity entropy in real time. and content diversity entropy , obtain the bridging entropy strength at the plate junction ;when (The threshold is tuned by historical statistics), local perturbations and bridge content injection are enabled. For all bridge positions, the system is based on causal weights Combined control factor with linear synthesis of entropy intensity , perform version reselection, sequence fine-tuning and bridge content enhancement.

[0172] In the concatenation generation phase, we first use the 1.20GB internal corpus to fine-tune the 12-layer Chinese diffusion language model; the system reads each , the forward noise coefficient is calculated as Correction( ). The forward noise of samples in the high randomization section increases by 23%; the low randomization section maintains the baseline. The denoising training is completed in the first step. In the inference stage, 6 candidate concatenations are generated in parallel for each bridge position. The candidate texts are successively scored on the three components of diversity, causal orientation, and depth. Those below the threshold of 0.45 are directly eliminated. The remaining texts are then screened by the retention rate prediction model (XGBoost, RMSE 0.04) and the semantic consistency of the previous and next texts. Finally, the concatenation with the highest score is retained for the program list. The risk filtering part is executed in parallel using three channels: dictionary + CNN + copyright. Training accuracy The recall rate for non-compliant programs reached 91.86%. Every night at 11:00 PM, the system initiates a batch review process. High-risk items (70 points) are automatically replaced, while medium-risk items (40-70 points) are pushed to the client for manual review. Two reviewers complete the review in an average of 1 minute and 48 seconds per item, ultimately generating a safe program list that can be directly broadcast and controlled.

[0173] During the 14-day official broadcast, the new system generated a total of 6,720 program items, including 560 bridges connecting sections. The average random entropy strength of each bridge was 0.47. The system generated 3,360 candidate linking phrases, of which 560 were retained after multi-dimensional screening, for a retention rate of 16.67%. Compared with traditional manual scripts, the average length of a single linking phrase was reduced from 84.13 words to 62.70 words. During the program risk assessment phase, 78 high-risk items (1.16%) were detected and automatically and seamlessly replaced. Of the 214 medium-risk items (3.18%), 158 were retained after manual review, with 56 edited. 6,428 low-risk items (95.66%) did not require any processing. After the final safe program list was broadcast through the broadcast control system, the average audience retention rate for the entire segment was 88.54%, compared to 76.03% for the control group (traditional manual editing), a 16.44% increase. The program recording error rate dropped from 0.45% to 0.08%, and the copyright conflict hit rate dropped from 0.38% to 0.05%. Automated programming reduced the time required to generate 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 fully demonstrates that the AI-driven broadcast scheduling solution is superior to traditional manual processes in terms of multi-dimensional indicators. First, in the dimension of content security, the system accurately divides the risks into high, medium and low levels and handles them in a timely manner. The proportion of high-risk items is only 1.16%, which is 62.80% less than the baseline process. This reduction is due to the triple fusion of dictionary-CNN-copyright: dictionary scanning ensures that obvious illegal content is immediately blocked, CNN's accuracy rate of 95.73% further captures semantic-level risks, and the copyright interface blocks potential authorization gaps, reducing the final copyright conflict rate to 0.05%, which is only one-seventh of the manual era. Secondly, the user experience is significantly improved. The system controls the bridge randomness and retention orientation through the entropy-causal model: average This indicates that moderate randomization was introduced at about half of the bridge points, but the causal weights The excessive jump is timely suppressed, and the retention rate is increased from 76.03% to 88.54%, which is equivalent to an increase of 16.44%. In a listener pool of 386,000 people, this increase means that an average of about 48,300 listeners are retained daily. At the same time, the average length of the series is shortened by 25.43%, achieving an increase in information density and a reduction in redundancy, which indirectly proves the generation advantage of the diffusion model in balancing diversity and simplicity. Again, the operation efficiency is multiplied. The automatic generation time of the complete program list is shortened from 42.70 minutes to 6.85 minutes, saving 83.9696 of the scheduling time; automatic replacement and manual review cooperation make the average load of the reviewer drop to 56 per day, which is 60.00% less than the benchmark of 140 per day. The saved manpower can be further invested in local interviews and in-depth content production, forming a virtuous cycle. Finally, the content quality and the stability of the broadcast are both enhanced. The disorderly behavior of the injury continuity is significantly reduced due to the order disturbance control, and the theme deviation between program segments is reduced by an average of 18.27%, and the film delay and the dead broadcast rate is reduced to 0.08%. In the hour refresh link of the extremely sensitive radio audience, automatic bridging makes the cut-in error remain within 0.12 seconds, which is 71.43% more stable than manual switching.

[0175] As can be seen, the four-layer linkage mechanism of structured material pool + entropy-causal adaptive scheduling + diffusion script generation + three-channel risk review not only solves the pain points of traditional programming mode, such as no response to real-time interaction, time-consuming and error-prone manual scripts, and lagging rule detection, but also builds a closed-loop link covering "content generation - decision scheduling - safety review - broadcast landing". Through the comparison of multi-dimensional measured data, it shows significant and credible advantages in retention rate, timeliness, compliance rate and operation cost and other key indicators.

[0176] The embodiment also provides a computer device suitable for the case of the intelligent programming method of a broadcast program, which comprises 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 realize the intelligent programming method of the broadcast program proposed in the above embodiment.

[0177] ​The computer device can be a terminal, which comprises a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved by WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0178] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to implement the intelligent arrangement method of a broadcast program as proposed in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic storage, a flash memory, a magnetic disk or an optical disk.

[0179] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A method for intelligently arranging broadcast programs, characterized in that: include: Collect broadcast materials and interactive data, and build a structured material pool; Generate the initial program list based on the program template rules, and inject random entropy at the junction of the section items to generate bridge content; Perform logical verification and dynamic adjustment on the bridge content to obtain an optimized program list; generating a linking phrase that matches the optimized program list based on a diffusion model, and optimizing the linking phrase; Scoring is performed on optimized program lists and linking phrases, automatically filtering out high-risk content, and forming a safe broadcast program list for broadcast.

2. The intelligent broadcast program arrangement method according to claim 1, wherein: The broadcast materials include music, news, weather, traffic conditions, advertisements, stories, lines, trailers, long programs, short programs, opening remarks and closing remarks; the interactive data includes song requests, messages, comments, votes and uploaded audio submitted by users through mini-programs, applications or websites; The construction of the structured material pool includes labeling the broadcast materials, archiving the interactive data, and aggregating all labeled broadcast materials and archived interactive data into a database to form a structured material pool that supports multi-condition retrieval, automatic editing and AI content generation.

3. The intelligent broadcast program arrangement method according to claim 2, wherein: 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 sub-type of material in each candidate material set is counted to obtain the content diversity index; the content diversity index and interaction diversity entropy are linearly weighted according to the preset weight to obtain the bridge entropy strength value; When the bridge entropy strength 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 bridge entropy strength value to a perturbation probability according to a preset rule; generating a random number for each pair of adjacent plate items, comparing it with the perturbation probability, and swapping the order if the random number is not higher than the perturbation probability; And ensure that after the exchange, the number of consecutive appearances of the same type of section items does not exceed the upper limit, and the program list still complies with the template order constraints; Before performing the disturbance, when the bridge entropy strength value is not lower than the insertion threshold or the difference in content between adjacent section items is not lower than the preset threshold, bridge content is inserted between adjacent section items; the bridge content includes at least one of impromptu connecting words, interactive summaries, hot news or sound effects, and the length and tone of the bridge content are determined by interpolation of the bridge entropy strength value within the respective preset intervals; the section item sequence after the disturbance is completed and the bridge content is inserted is used as the initial program list.

4. The intelligent broadcast program arrangement method according to claim 3, wherein: The obtaining of the optimized program list includes extracting a bridge entropy strength value of each bridge position from the initial program list; inputting the bridge entropy strength value and interaction data corresponding to the bridge position into a causal inference model, and obtaining a causal weight for each bridge position from the causal inference model; The bridge entropy intensity value and the causal weight are combined in proportion to generate a joint control factor for each bridge position; all bridge positions are traversed in turn, and each bridge position is optimized and adjusted according to the numerical range of the joint control factor. The optimization adjustment includes version reselection, sequence disturbance and bridge content enhancement; after traversing and completing the dynamic adjustment of all bridge positions, an optimized program list is generated and output.

5. The intelligent broadcast program arrangement method according to claim 4, wherein: Generating a concatenation phrase that matches the optimized program list includes extracting a joint control factor for each bridge position in the optimized program list and modifying the basic noise coefficient according to the value of the joint control factor before injecting noise in the forward diffusion process; In each reverse denoising iteration, 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, the joint control factor is compared with a preset threshold based on the default number of denoising iterations and the default sampling step size of the diffusion model. 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 value obtained by adding the default number of iterations to the product of the difference between the joint control factor and the threshold and the predetermined iteration increment. The sampling step size of each step is shortened by calculating the difference between the joint control factor and the preset threshold, multiplying the difference by a predetermined distance reduction coefficient, and obtaining the step size that needs to be reduced; the result obtained by subtracting the step size 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.

6. The intelligent broadcast program arrangement method according to claim 5, wherein: Optimizing the concatenation includes calculating a diversity score, a causal orientation score, and a generation depth score for each candidate concatenation based on the recorded noise injection rate, gradient correction weight, 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 by the total length of the text; the causal orientation score is obtained by taking the arithmetic average of the gradient correction weights in all denoising iterations for the candidate concatenation; and the generation depth score is obtained by subtracting a preset minimum number of iterations from the actual number of denoising iterations for the candidate concatenation, and dividing the difference by the difference between the maximum number of iterations and the minimum number of iterations, and the result is used 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 below the minimum passing threshold are eliminated; The audience retention rate is defined as the ratio of the number of remaining audience members at the end of a program segment to the number of audience members at the beginning of the program 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 word, the corresponding joint control factor and the contextual features are input into the retention rate prediction model, the expected retention rate of the candidate concatenation word is calculated, and candidate concatenation words with an expected retention rate lower than the historical average retention rate are eliminated; the remaining candidate concatenation words are compared with the previous and next paragraphs of text for semantic similarity, and candidate concatenation words with semantic similarity lower than a preset consistency threshold are eliminated; the final remaining concatenation words that meet the contextual coherence and the corresponding joint control factors are used as the final broadcast script and submitted together with the optimized program list.

7. The intelligent broadcast program arrangement method according to claim 6, wherein: The formation of a safe broadcast program list includes conducting a content review of each section item text and linking phrase based on a violation word library, a pre-trained deep classification model, and a copyright database interface to generate a risk score; Risk scores are classified according to preset classification 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 safe alternative library based on the label, type and duration of the original material. After the manual review is completed, the review results are synchronously updated to the program list to form a safe broadcast program list.

8. An intelligent broadcast program arrangement system, based on the intelligent broadcast program arrangement method according to any one of claims 1 to 7, characterized in that: include: The data module collects broadcast materials and interactive data and builds a structured material pool; The bridging module generates the initial program list based on the program template rules and injects random entropy at the connection between the section items to generate the bridging content; the optimization module performs logical verification and dynamic adjustment on the bridging content to obtain the optimized program list; The linking phrase module generates linking phrases that match the optimized program list based on the diffusion model and optimizes the linking phrases; the scoring module scores the optimized program list and linking phrases, automatically filters out high-risk content, and forms a safe broadcast program list for broadcasting.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the intelligent arrangement method of broadcast programs described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the intelligent arrangement method of broadcast programs described in any one of claims 1 to 7 are implemented.

Images