Incremental training and optimization processing method and system of short drama field corpus

Abstract

The application provides an incremental training and optimization processing method and system of a short play field corpus, relates to the technical field of short play content data processing, and comprises the following steps: through a field self-adaptive intelligent corpus pre-structuring model, original short play field data from an authorized source is integrally processed to generate labeled corpus; the integrally processing comprises automatic data quality screening, labeling of a plot structure label and an emotional context label on text data, and labeling of multi-dimensional visual feature labels on visual materials; based on the labeled corpus, a field instruction fine-tuning and human preference reinforcement learning are performed on a basic generation model to optimize the performance of the basic generation model in narrative generation and cross-modal consistency, and an optimized basic generation model is obtained. The application is suitable for realizing closed-loop optimization and precise improvement of the field adaptability of the generation model of the short play field corpus through multi-dimensional interactive feedback of users.

Description

Technical Field

[0001] This invention relates to the field of short drama content data processing technology, and in particular to the incremental training and optimization processing method and system for a corpus specifically for short dramas. Background Technology

[0002] With the deep application of generative artificial intelligence technology in the field of online short dramas, fine-tuning and incremental optimization of generative models based on dedicated corpora have become the core means to improve the narrative coherence, emotional relevance and cross-modal alignment of short dramas.

[0003] Current corpus construction and training methods in the industry mostly adopt the offline collection, manual annotation, and static storage model. The technical shortcomings they face are: the corpus for short dramas lacks quantitative feature extraction, topological measurement and adaptive semantic distillation mechanisms based on real-time interactive feedback from multiple user dimensions. It is impossible to build a closed-loop optimization system for model-generated content, user interaction behavior, feedback feature quantification, corpus priority screening and incremental training and updating. The iteration of the corpus and the generated model is deviated from the dynamic guidance of real user preferences.

[0004] Because it is impossible to extract key indicators such as peak density, cluster centers, and popularity extremes of user interaction feedback on short story plot structure, emotional context, and multi-dimensional visual features, and it is also impossible to construct feature vector channels under a unified metric space, corpus updates can only rely on manual experience or fixed rule selection, making it difficult to accurately capture real-time changes in user preferences. This results in a long-term static and lagging state, causing the narrative generation and cross-modal consistency optimization effects of the generative model to decay rapidly after fine-tuning. At the same time, the lack of quantitative methods such as topological inclusion degree and metric overlap ratio of feature channels in different iteration cycles makes it impossible to generate reasonable semantic distillation intensity coefficients to guide incremental training. Incremental updates are prone to mixing in low-value, redundant, and noisy data, which reduces training efficiency and fails to form a long-term optimization loop. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide an incremental training and optimization method and system for a corpus dedicated to the field of short dramas. It solves the problems of static lag and deviation from user preferences of existing corpora, realizes accurate incremental updates of the corpus and continuous optimization of model performance, and effectively promotes the large-scale implementation and upgrading of short drama generation technology.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, an incremental training and optimization method for a corpus specifically designed for short dramas, the method comprising: Step 1: Using a domain-adaptive intelligent corpus pre-structured model, the original data from the short drama domain from the authorized source is processed in an integrated manner to generate an annotated corpus; the integrated processing includes automated data quality screening and annotation of text data with plot structure tags and emotional context tags, and annotation of visual materials with multi-dimensional visual feature tags; Step 2: Based on the annotated corpus, perform domain instruction fine-tuning and human preference reinforcement learning on the basic generative model to optimize the performance of the basic generative model in narrative generation and cross-modal consistency, and obtain a performance-optimized basic generative model. Step 3: Generate short drama content using the performance-optimized basic generation model, and simultaneously collect user interaction behavior datasets related to the short drama content; Step 4: Extract the peak density range of interaction data corresponding to the plot structure tags of the short drama content, the high-dimensional clustering center in the emotional context tag vector, and the local maxima in the multimodal visual feature attention heat map from the user interaction behavior dataset as multiple key feedback index points; dynamically construct the corresponding feature vector contour based on the coordinates of the key feedback index points in the multi-dimensional quantity space. Step 5: Calculate the topological inclusion degree and measure overlap ratio between feature vectors at different times; generate semantic distillation intensity coefficients based on the topological inclusion degree and measure overlap ratio. Step 6: Based on the semantic distillation intensity coefficient, prioritize and filter the annotated corpus, and perform incremental training to update the dedicated corpus to achieve closed-loop optimization.

[0007] Secondly, the incremental training and optimization system for a dedicated corpus in the short drama field includes: The corpus preprocessing module is used to process the original data of the short drama domain from the authorized source in an integrated manner through a domain-adaptive intelligent corpus pre-structured model to generate an annotated corpus; the integrated processing includes automated data quality screening and annotation of text data with plot structure tags and emotional context tags, and annotation of visual materials with multi-dimensional visual feature tags; The model optimization module is used to fine-tune the basic generative model with domain instructions and human preference reinforcement learning based on the labeled corpus, so as to optimize the performance of the basic generative model in narrative generation and cross-modal consistency, and obtain a performance-optimized basic generative model. The interactive data acquisition module is used to produce short drama content through a performance-optimized basic generation model and simultaneously collect user interaction behavior datasets related to the short drama content. The feedback feature quantization module is used to extract from the user interaction behavior dataset the peak density range of interaction data corresponding to the plot structure tags of the short drama content, the high-dimensional clustering center in the emotional context tag vector, and the local maxima in the multimodal visual feature attention heat map as multiple key feedback indicator points; based on the coordinates of the key feedback indicator points in the multi-dimensional quantity space, the corresponding feature vector contour is dynamically constructed. The semantic distillation module is used to calculate the topological inclusion degree and measure overlap ratio between feature vectors at different times; based on the topological inclusion degree and measure overlap ratio, the semantic distillation intensity coefficient is generated. The incremental training module is used to prioritize and filter the labeled corpus based on the semantic distillation intensity coefficient, and perform incremental training to update the dedicated corpus, thereby achieving closed-loop optimization.

[0008] The above-described solution of the present invention has at least the following beneficial effects: By employing domain-adaptive corpus preprocessing, quantitative extraction of key indicators from user interaction feedback, feature vector path construction, topological metric and semantic distillation coefficient generation, as well as corpus priority screening and incremental training techniques, this approach effectively overcomes the technical problems of existing short drama-specific corpora being statically lagging, detached from user preferences, unable to form a closed-loop optimization system, and exhibiting diminishing optimization effects, low incremental training efficiency, and susceptibility to low-value data contamination. This results in precise incremental updates of the corpus, improved narrative generation and cross-modal consistency performance of the generation model, increased model training efficiency, and the formation of a long-term iterative closed loop, effectively promoting the large-scale implementation and upgrading of short drama generation technology. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the incremental training and optimization process for a corpus specifically designed for short dramas, provided in an embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of an incremental training and optimization system for a corpus specifically for short dramas, provided in an embodiment of the present invention. Detailed Implementation

[0011] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0012] like Figure 1 As shown, embodiments of the present invention propose an incremental training and optimization method for a corpus specifically designed for short dramas. The method includes the following steps: Step 1: Using a domain-adaptive intelligent corpus pre-structured model, the original data from the short drama domain from the authorized source is processed in an integrated manner to generate an annotated corpus; the integrated processing includes automated data quality screening and annotation of text data with plot structure tags and emotional context tags, and annotation of visual materials with multi-dimensional visual feature tags; Step 2: Based on the annotated corpus, perform domain instruction fine-tuning and human preference reinforcement learning on the basic generative model to optimize the performance of the basic generative model in narrative generation and cross-modal consistency, and obtain a performance-optimized basic generative model. Step 3: Generate short drama content using the performance-optimized basic generation model, and simultaneously collect user interaction behavior datasets related to the short drama content; Step 4: Extract the peak density range of interaction data corresponding to the plot structure tags of the short drama content, the high-dimensional clustering center in the emotional context tag vector, and the local maxima in the multimodal visual feature attention heat map from the user interaction behavior dataset as multiple key feedback index points; dynamically construct the corresponding feature vector contour based on the coordinates of the key feedback index points in the multi-dimensional quantity space. Step 5: Calculate the topological inclusion degree and measure overlap ratio between feature vectors at different times; generate semantic distillation intensity coefficients based on the topological inclusion degree and measure overlap ratio. Step 6: Based on the semantic distillation intensity coefficient, prioritize and filter the annotated corpus, and perform incremental training to update the dedicated corpus to achieve closed-loop optimization.

[0013] In this embodiment of the invention, domain-adaptive intelligent corpus pre-structuring processing improves the accuracy and efficiency of short drama domain corpus annotation. By fine-tuning the basic generation model with domain instructions and strengthening learning based on human preferences, the model's narrative generation capability and cross-modal consistency are effectively optimized. Simultaneously collecting user interaction behavior data and extracting key feedback indicators can accurately capture users' preferences for short drama plots, emotions, and visual features. Dynamically constructing feature vector boundaries and generating semantic distillation intensity coefficients provides a scientific and reasonable basis for corpus selection. Based on the semantic distillation intensity coefficients, corpus priority selection and incremental training are performed, achieving accurate updates and closed-loop optimization of the dedicated corpus, improving model training efficiency, and contributing to the practical upgrade of short drama generation technology.

[0014] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1 involves accessing and parsing the original short drama domain data stream from the authorized source using a domain-adaptive intelligent corpus pre-structured model. Specifically, the domain-adaptive intelligent corpus pre-structured model is an improved version of the BERT-base model combined with a BiLSTM-CNN hybrid model architecture. Its basic architecture originates from mature semantic understanding and multimodal feature extraction models in the field of natural language processing. Targeted architecture optimization and parameter adjustment have been carried out to meet the core requirements of multimodal, temporal, and scenario-based short drama domain corpus in this invention, thus constructing a dedicated model adapted to the preprocessing of short drama corpus. The specific construction process is as follows: The BERT-base model is used as the core text semantic feature extraction module. By fine-tuning the word embedding layer of the model, vocabulary specific to the short drama domain is incorporated, such as short drama narration, transition shots, and exciting plot points, to enhance the model's ability to accurately capture the semantics of short drama texts. A two-layer BiLSTM model module is introduced to adapt to the temporal coherence features of short drama plot segments, better capturing the logical connection between different plot nodes and avoiding semantic breaks in the plot. Three 1D-CNN convolutional layers and pooling layers are added, with convolutional kernel sizes set to 3, 5, and 7 respectively, to initially extract shallow visual features of short drama visual materials, such as character outlines and scene outlines, to achieve a preliminary semantic association between text and visual materials. Finally, a domain-adaptive adaptation layer is added, embedding a short drama domain-specific dictionary, standardized plot structure templates (such as label templates for beginning, development, climax, and ending), and basic rules for visual feature annotation. At the same time, a label expansion interface is reserved to adapt to the corpus processing needs of different types of short dramas in the future, completing the architecture construction of the entire model.

[0015] The training process of this model closely revolves around the corpus processing needs of the short drama domain of this invention, aligning with the actual scenarios of short drama creation and corpus application. Specifically, the training process first selects a sufficient amount of manually annotated sample data from various fields, including short drama scripts, short drama video clips, short drama frames, and short drama dialogue scripts, covering multiple mainstream short drama types such as urban, historical, suspense, and romance. It also incorporates common special scene samples from short dramas, such as the ancient style scenes, dialogue, and visuals of historical dramas, and the foreshadowing plot texts and shots of suspense dramas. Training and validation datasets are constructed, with the ratio of training to validation datasets set at 8:2 to ensure the effectiveness of the training. During training, the basic BERT-BiLSTM-CNN architecture is pre-trained using general Chinese corpus to enable the model to possess basic semantic understanding and feature extraction capabilities. Then, multiple rounds of fine-tuning training are conducted using the constructed short drama domain sample dataset, focusing on optimizing the parameters of the domain adaptive layer and word embedding layer to enhance the model's ability to recognize and match short drama plot structures, emotional expressions, and visual features.

[0016] During training, the AdamW optimizer was used, a reasonable learning rate decay strategy was set, and the cross-entropy loss function was introduced to calculate the model prediction error. A validation dataset was introduced in real time to verify the effect. To address issues such as inaccurate extraction of visual features from short dramas and deviations in understanding the semantics of the plot, the model regularization parameters and the number of iterations were continuously adjusted. Anomalies such as overfitting and recognition bias that occurred during training were eliminated until the model reached the preset targets in terms of accuracy in short drama text parsing, accuracy in preliminary recognition of visual materials, and data processing efficiency, thus completing the model training.

[0017] The advantages and benefits of this improved model are significant. Compared to general corpus processing models, it can accurately adapt to the characteristics of short drama corpora, effectively avoiding problems such as semantic bias, inaccurate feature extraction, and chaotic temporal logic that occur when general models process short, fast-paced plot segments and multimodal corpora in short dramas. Simultaneously, this model possesses the dual capabilities of text semantic understanding and preliminary visual feature extraction, enabling rapid integration with various short drama data sources. It effectively and reasonably adapts to the application scenario of simultaneous processing of scripts, visuals, and dialogue in short drama creation. For example, when processing dialogue text and romantic scenes in sweet romance short dramas, it can accurately capture the emotional inclination in the dialogue and the visual style of the scene, achieving preliminary alignment of text and visual features. Furthermore, the model's domain-adaptive characteristics allow it to flexibly adapt to the corpus processing needs of different types of short dramas, eliminating the need to repeatedly build models for a single short drama type. This significantly reduces model construction and maintenance costs while improving the efficiency and accuracy of subsequent corpus parsing and labeling, laying a solid foundation for the entire corpus preprocessing process. After the model is built and trained, it is connected to the compliant and authorized short drama data source interface through the adaptive intelligent corpus pre-structured model of this domain. It can access multiple types of raw data streams such as short drama text (dialogue, narration, script), video clips, and frame in real time. According to the preset domain data format specifications, it can complete the parsing, splitting and normalization of the data stream, and split the unstructured raw data into structured data fragments that can be processed later.

[0018] Step 1.2 involves using a domain-adaptive intelligent corpus pre-structured model to perform automated quality screening and copyright compliance verification on the raw data stream, resulting in a compliant data set. Specifically, this includes the quality screening module and copyright verification module of the domain-adaptive intelligent corpus pre-structured model working together to apply pre-defined data quality standards for the short drama domain. These standards, considering the characteristics of short dramas—briefness, speed, and coherent plot—are set from three core dimensions: text, visuals, and completeness. Text data must conform to the specifications for short drama dialogue and narration, with no garbled characters, no obvious grammatical errors, and no missing core plot information; visual materials must have clear images and complete frame sequences. The data should be free of blurriness, distortion, and stuttering, with clear and distinguishable audio (if any). Simultaneously, the data must be free of duplication and redundancy, and contain no missing core information. The core semantics must be consistent with the visual information. The entire process of automated filtering is performed on the parsed and normalized raw data stream. Duplicate data is judged based on the comparison of data feature hash values, with a hash value duplication threshold set to no higher than 5%, eliminating completely duplicated or highly similar redundant data. For incomplete data, the focus is on checking the text missing rate and the missing frames of visual materials, eliminating invalid data with more than 30% missing text or missing frames that cause plot breaks. Low-quality data such as blurry images, distorted audio, and garbled text are also filtered out.

[0019] In the copyright compliance verification stage, the copyright database interface is called to compare the ownership registration information and authorization document details of the short drama data. The authorization subject, authorization scope, usage period and usage scenario of each data are verified one by one. The focus is on screening out abnormal data with copyright risks, such as unauthorized edited clips and infringing remakes. For example, for classic scene clips of period dramas, the dual authorization compliance of the screen copyright and the dialogue copyright is additionally verified. The abnormal data is marked and the reasons for the abnormality are recorded in detail. Finally, all the data that has passed the screening and verification are integrated to form a clean, standardized, copyright-free compliance data set that can be directly used for subsequent processing.

[0020] Step 1.3 involves performing narrative structure analysis and emotional context analysis on the text data in the compliant data set to generate corresponding plot structure tags and emotional context tags. Specifically, this includes: for text data in the compliant data set, relying on the text parsing module and sentiment calculation submodule of the domain-adaptive intelligent corpus pre-structured model, combined with a narrative logic template specific to the short drama domain. This template aligns with the core characteristics of short dramas—fast-paced, tightly plotted, and concentrated conflict—and pre-sets standardized narrative node division rules for introduction, development, climax, and ending. It also adapts to the narrative differences of different short drama types such as urban, historical, and suspense dramas, confirming the plot duration proportion, logical connection requirements, and core plot information extraction standards for each node. The text content is then segmented to accurately divide the plot into beginning, development, climax, and ending. The system analyzes key plot points, establishes logical connections between them, and performs narrative structure analysis. It also labels each plot point with corresponding scene and plot type tags; for example, it labels the text data of urban short dramas as "workplace scene - development point" and "family scene - climax point." Through the sentiment computing submodule and a short drama-specific sentiment dictionary, it analyzes the emotional tendency, intensity of emotional fluctuations, and characteristics of emotional continuity sentence by sentence, identifying core emotional expressions such as "exhilarating moments," "angsty moments," and "sweet moments," generating corresponding emotional context tags, including emotional type tags (such as joy, sadness, tension, and sweetness) and emotional intensity tags (such as mild, moderate, and high intensity). For example, it labels the dialogue text of sweet romance short dramas with a "sweet - high intensity" emotional context tag. Finally, it matches complete and accurate plot structure tags and emotional context tags for each piece of text data.

[0021] Step 1.4 involves multimodal feature extraction and alignment of visual materials in the compliant dataset to generate corresponding multidimensional visual feature labels. Specifically, this includes: for visual materials in the compliant dataset, a comprehensive visual feature extraction and alignment process is performed using the multimodal feature extraction submodule and feature alignment module of a domain-adaptive intelligent corpus pre-structured model. The visual feature extraction focuses on: image composition (including long shots, close-ups, and extreme close-ups); scene type (including indoor, outdoor, ancient style, and modern urban scenes); character features (such as makeup, clothing, expressions, and actions); color style (including cool tones, warm tones, and highly saturated tones); and camera language (such as push-in shots, pull-out shots, and panning shots). The core dimensions include, for example, visual materials from suspenseful short dramas, focusing on extracting the tense features in the characters' expressions, the cool color tones in the visuals, and the close-up features in the camera language; in the feature alignment stage, the extracted visual features are double-aligned with the semantic content of the corresponding text data. In terms of time sequence, the visual frames are accurately matched with the corresponding lines and narration based on the timestamps. In terms of semantics, the visual features are matched with the scenes, emotions, and characters described in the text through semantic similarity calculation. For example, the visual scene features of the ancient courtyard in the costume short drama are semantically aligned with the description of the deep courtyard and the antique style in the corresponding text. Finally, a multi-dimensional visual feature label with accurate matching and complete dimensions is generated for each piece of visual material.

[0022] Step 1.5 involves associating and encapsulating the plot structure tags, emotional context tags, and multi-dimensional visual feature tags with the corresponding original data to generate annotated corpus. Specifically, this includes: using the association and encapsulation module of a domain-adaptive intelligent corpus pre-structured model, uniquely identifying and associating the plot structure tags, emotional context tags, and multi-dimensional visual feature tags with the corresponding original data in the compliant data set, and using a data ID and tag ID binding method to ensure that each piece of original data corresponds one-to-one with all its corresponding tags, avoiding tag confusion; after association, the data and tags are uniformly encapsulated according to the set structured corpus format: the encapsulated content includes the original data (text... The system includes the storage path, unique identifier ID, various tag sets, and data type labels (text or visual) for the text and visual materials. It also adds data processing timestamps and verification codes for subsequent data traceability and integrity checks. After encapsulation, each annotated corpus is automatically checked for integrity, identifying issues such as missing tags, incorrect associations, and abnormal encapsulation formats. These are corrected promptly, resulting in a unified format, complete tags, accurate associations, and directly usable for subsequent model training and optimization. This adapts to the incremental training needs of specialized corpora in the short drama domain. For example, annotated corpora for different types of short dramas all use a unified encapsulation format, facilitating efficient subsequent semantic distillation and incremental training.

[0023] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: Construct a domain instruction fine-tuning dataset based on the annotated corpus. This dataset includes short drama narrative structure examples extracted from the annotated corpus, as well as the alignment relationships between text and visual labels in the semantic vector space. Specifically, based on the generated annotated corpus, systematically construct the domain instruction fine-tuning dataset. First, confirm the dataset construction standards. Extract complete narrative fragments and classic plot templates from various mainstream short dramas (urban, historical, suspense, and romantic) proportionally from the annotated corpus as short drama narrative structure examples. Select no fewer than 500 examples for each type of short drama. Each example should have a text length of 1 to 3 minutes, covering the entire process from plot introduction, development, climax, and ending. Ensure that the examples cover different types of plot conflict and emotional expression styles. Then, extract the text data, plot structure labels, emotional context labels, and multi-dimensional visual feature labels from the annotated corpus one by one. Perform semantic encoding processing on these texts and labels respectively. The text data is segmented and encoded according to plot segments to preserve narrative logic. Various visual tags are structured and encoded based on their feature descriptions to highlight visual attributes. The encoded text vectors and visual tag vectors are then uniformly mapped to a 768-dimensional semantic vector space to ensure they are on the same metric dimension for easy comparison. Subsequently, the cosine similarity between the text encoded vector and the corresponding visual tag encoded vector is calculated to quantify the semantic association between them. Each pair of text and visual tag data is compared for similarity, with a similarity threshold of no less than 88% set as the alignment standard. Data pairs with similarity below this threshold or semantic discrepancies (e.g., text describing sweet interactions but visually labeled as cool colors) are marked. All data pairs that do not meet the alignment standard are batch-filtered and removed. At the same time, data pairs that meet the alignment standard are uniquely identified and bound to confirm the accurate correspondence between the recorded text content and various visual tags, ensuring that the subsequent model training can accurately call the association information between text and visual tags.

[0024] Finally, all qualified narrative structure examples and text-visual label alignment relationships were integrated and divided into training and validation subsets in an 8:2 ratio. Each data point was supplemented with exclusive instruction descriptions tailored to the actual creation of short dramas. Instruction types covered plot generation, segment continuation, and label matching generation. For example, a 1.5-minute ancient costume short drama segment was generated by combining ancient style scenes and a sadness-moderate label, following the narrative logic of ancient costume short dramas. At the same time, the dataset format was standardized, and the type, label attributes, and instruction requirements of each data point were labeled. This resulted in a domain instruction fine-tuning dataset that was standardized in format, comprehensive in content, tailored to the needs of the short drama domain, and directly usable for model fine-tuning. This ensured that the dataset could effectively guide the model to learn the characteristics of the short drama domain.

[0025] Step 2.2 involves supervising the basic generative model using a domain-specific instruction fine-tuning dataset. This allows the basic generative model to learn the narrative logic and cross-modal alignment constraints of the short drama domain, resulting in a finely tuned basic generative model. Specifically, this includes: first, determining the selection and improvement of the basic generative model. The basic generative model in this step is an improvement on the GPT-4 architecture. To address the multimodal content generation needs of short dramas, an additional cross-modal semantic alignment layer and a plot pacing control layer are added. The cross-modal semantic alignment layer embeds text-visual semantic mapping rules specific to the short drama domain, enabling the mapping of text content, plot structure tags, emotional context tags, and multi-dimensional visual elements. The precise semantic binding of feature labels and the addition of short drama duration threshold constraints and plot node density control logic to the plot pacing control layer enable dynamic adjustment of plot progression speed and node distribution based on the core duration requirements of 1 to 3-minute short dramas, avoiding issues such as dragging pacing or missing nodes. The improved model retains the original architecture's powerful semantic understanding and content generation capabilities while accurately adapting to the short, fast-paced, and tightly plotted generation characteristics of short dramas. It also supports the linked learning of text and visual labels. The advantages of adopting this improved architecture are that it can reduce redundant content generated by the model, improve the accuracy of plot pacing control and cross-modal consistency, and reduce the iteration cost of subsequent model optimization.

[0026] Subsequently, the constructed domain instruction fine-tuning dataset was input into the improved basic generative model in fixed batches (batch size set to 32, adapting to the batch training requirements of short drama samples) to conduct instruction-supervised fine-tuning. During the fine-tuning process, a hierarchical fine-tuning strategy was adopted, freezing the parameters of the first 10 basic semantic layers of the model to retain general semantic understanding capabilities, and only fine-tuning the parameters of the mid-layer short drama domain adaptation module and the top-layer plot generation output layer. The initial learning rate was set to 1e-5, and a step-wise learning rate decay strategy was adopted, with the learning rate decaying by 10% every 100 iterations. At the same time, an L2 regularization term (regularization coefficient set to 1e-6) was introduced to prevent model overfitting. The input instruction format strictly fits the short drama creation scenario, confirming the annotation of plot theme, emotional intensity, visual style, duration range and core requirements. For example, combining modern workplace scenarios, tension-high visual and emotional tags, a 1-minute short drama segment was generated, requiring the plot to have conflict, concise dialogue, fit the narrative logic of workplace short dramas, and have no redundant dialogue.

[0027] During fine-tuning, every 50 iterations, the effectiveness is verified using a validation subset of the domain instruction fine-tuning dataset. The focus is on checking the narrative coherence (no missing plot nodes, no logical breaks), label matching accuracy (label matching accuracy of no less than 88%), and domain adaptability (conforming to the narrative logic and language style of the corresponding short drama type). For deviations such as broken plots, disjointed labels, and redundant dialogue, the training weights of the corresponding deviation samples are increased to 1.2 times the original weights. At the same time, the learning rate decay rhythm is fine-tuned, and severely deviation samples with label matching accuracy below 60% are removed. The training is iterated repeatedly until the model can stably generate content that conforms to the short drama domain specifications, has accurate label matching, and reasonable plot pacing. Furthermore, the label matching accuracy on the validation subset is consistently no less than 90%, and the narrative coherence rate is no less than 85%. Fine-tuning is then stopped, and the basic generation model after instruction fine-tuning is obtained.

[0028] Step 2.3 involves generating a diverse set of short drama content candidates using the finely tuned basic generation model, and collecting human preference feedback data on these candidates. Specifically, this includes: first, identifying popular demands in the short drama creation and distribution field; then, combining the preferences of different user groups, designing diverse creative instructions for short dramas. These instructions cover core elements such as plot theme, emotional tone, visual style, and duration requirements. For example, an urban romance theme with a modern urban setting, sweet and highly emotional, 1 minute in length, including dialogue between the male and female leads, and visual tags such as warm colors and close-up shots; or a suspenseful detective theme with a rainy night setting. The scene is tense and highly emotional, with a duration of 2 minutes, including foreshadowing. The visual tags are cool colors and close-up shots, ensuring that the creative instructions cover various mainstream short drama types and user preferences. Then, these creative instructions are input one by one into the basic generation model after instruction fine-tuning. The model drives the generation of 5 to 8 different versions of short drama content candidate sets for each creative instruction. Each version includes complete script dialogue, plot node division, emotional thread description, visual scene suggestions and corresponding tags. At the same time, the creative highlights and differences of each version are marked, which fits the application scenario of multi-version screening in actual short drama creation.

[0029] Next, the collection of human preference feedback data was initiated. A feedback team was formed, consisting of 10 to 15 professional short drama creators, 30 to 50 target users, and 5 to 8 experts in the field. Clear feedback evaluation criteria were established, focusing on five core dimensions: plot coherence, emotional relevance, text-visual tag matching, plot appeal, and dialogue naturalness. Each short drama candidate version was scored on a 10-point scale. Specific modification suggestions from the feedback team were collected and categorized into a format of strengths + weaknesses + modification direction. All scoring results, categorized evaluations, and modification suggestions were compiled and archived to form complete, detailed, and practically relevant human preference feedback data, providing a precise basis for subsequent model optimization.

[0030] Step 2.4: Define reward signals based on collected human preference feedback data. Then, using reinforcement learning, optimize the basic generation model after fine-tuning the instructions to improve the alignment between generated content and human preferences. Specifically, this includes: accurately defining reward signals based on collected human preference feedback data. The reward signals are strictly aligned with the evaluation standards and human preferences in the short drama field. The average score from the feedback team is used as the core quantitative basis to confirm the correspondence between scores and reward signals. An average score of 9 to 10 corresponds to the highest reward value, 7 to 8 corresponds to a medium reward value, 5 to 6 corresponds to a basic reward value, and scores below 5 are set with negative rewards. Simultaneously, combining evaluation opinions and modification suggestions, additional reward weights are added by 10% to 20% for candidate content with coherent plot, high emotional relevance, and accurate cross-modal matching. Candidate content with issues such as plot breaks, emotional deviations, visual-text disconnect, and stiff dialogue is penalized. The reward weight is reduced or even eliminated based on the severity of the problem. Specific quantitative standards and calculation methods for reward signals are clearly defined to ensure that the reward signals accurately reflect human preferences. Subsequently, the PPO reinforcement learning algorithm is selected for model policy optimization. This algorithm has a fast convergence speed, adapts to policy adjustments in multimodal generation models, and effectively avoids performance fluctuations during model optimization. The defined reward signals are input into the PPO algorithm to optimize the policy of the basic generation model after fine-tuning the instructions. During optimization, the focus is on adjusting the model's plot generation strategy, emotion expression strategy, and cross-modal alignment strategy. For high-frequency problems in the feedback, such as plot pacing being too fast or too slow, inaccurate emotion expression, and visual tags being disconnected from text, model parameters are adjusted accordingly to guide the model to reduce the generation of content that does not conform to human preferences. Simultaneously, the model optimization effect is monitored in real time, with a small-scale verification every 30 rounds of optimization to ensure that the alignment between the model's generated content and human preferences gradually improves.

[0031] Step 2.5 involves iteratively executing the instruction-supervised fine-tuning and reward-signal-based policy optimization process until the performance of the basic generative model reaches the preset optimization target, thereby obtaining the performance-optimized basic generative model. Specifically, this includes: initiating the iterative execution process of instruction-supervised fine-tuning and reinforcement learning policy optimization; determining the specific process and requirements for each iteration; and using a supplementary subset of the constructed domain instruction fine-tuning dataset to perform supplementary instruction-supervised fine-tuning on the model. The supplementary subset focuses on selecting samples corresponding to the weaknesses exposed in the previous round of reinforcement learning, such as cross-modal matching bias and inaccurate emotional expression data, to consolidate the model's learning of the short drama domain's narrative logic and cross-modal alignment constraints. Simultaneously, parameters such as the learning rate are fine-tuned to avoid model overfitting. After the supplementary fine-tuning is completed, following the processes in steps 2.3 and 2.4, 200 to 300 short drama content candidate sets are regenerated, the latest human preference feedback data is collected, and the quantification standard and weight of the reward signal are updated based on the feedback data to conduct a round of reinforcement learning policy optimization.

[0032] After each iteration, a comprehensive validation is performed using preset model performance evaluation metrics. These metrics include average human preference scores, plot coherence compliance rate, and cross-modal alignment accuracy. The preset optimization targets are an average human preference score of no less than 85 points, a plot coherence compliance rate of no less than 90%, and a cross-modal alignment accuracy of no less than 92%. During the validation process, no fewer than 100 test samples are selected, and the validation is jointly completed by the feedback team and model testers, with detailed records of any model deficiencies. If the model performance does not meet the preset optimization targets, the fine-tuning parameters, reinforcement learning reward weights, and iteration strategies are adjusted based on the validation results. For example, the training weights of emotion expression-related samples are appropriately increased, and the learning rate of the PPO algorithm is adjusted. The above iteration process is then repeated. If the model performance meets the set optimization targets, the iteration stops, and an additional 500 short drama test data samples that were not used in training are selected for final performance validation of the model. This ensures that the model can stably generate content that aligns with human preferences, has strong cross-modal consistency, and conforms to the standards of the short drama domain, ultimately obtaining a performance-optimized basic generation model.

[0033] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1 involves inputting creative instructions into the performance-optimized basic generation model to drive the generation of short drama content containing video, audio, and related metadata. Specifically, this includes designing standardized and diverse creative instructions based on the preferences of various target user groups and mainstream short drama distribution scenarios. These instructions clearly include plot themes, emotional inclinations, visual styles, duration requirements, core plot conflicts, and suggestions for camera language, such as close-up dialogue and close-up emotions. They also specify requirements for dialogue conciseness and emotional expression intensity, covering various mainstream short drama types such as urban, historical, suspense, and romantic comedies. Each instruction is jointly reviewed by 2 to 3 professional short drama creators, focusing on whether the instructions align with the concise and fast-paced characteristics of short dramas, whether the descriptions are clear and unambiguous, and whether the requirements are feasible, ensuring that the instructions meet actual creation and distribution needs. Subsequently, these standardized creative instructions are input one by one into the performance-optimized model in batches of 30. The basic generative model first precisely parses the standardized creative instructions input, extracting core information such as plot theme, emotional tendency, visual style, and duration requirements. Then, combined with the short drama narrative logic learned through training, it plans the plot framework and matches corresponding plot structure tags according to the node division rules of introduction, development, climax, and ending. Subsequently, based on cross-modal alignment rules, it precisely semantically binds the planned plot text, emotional context tags, and multi-dimensional visual feature tags to ensure that the text description, visual presentation, and emotional expression are highly consistent, avoiding the disconnect between tags and content. At the same time, it dynamically allocates the duration of each plot node, controls the conciseness of the dialogue and the speed of plot progression, highlights the core plot conflict, avoids dragging the pace or missing nodes, and strictly adheres to the duration and presentation standards required by the instructions. Finally, it integrates all the above processing results to generate complete short drama content.

[0034] The generated short drama content includes high-definition video with a resolution of no less than 1080P and a frame rate of 25fps, audio with a sampling rate of no less than 44.1kHz and no noise or distortion, clear dialogue, background sounds and sound effects that fit the plot, and detailed associated metadata. The metadata specifically includes a unique identifier for the short drama, a plot synopsis, duration, genre, all corresponding plot structure tags, emotional context tags, multi-dimensional visual feature tags, generation time, and source of instructions. After generation, each piece of short drama content is verified one by one to check whether the video is choppy, whether the picture is clear, and whether the audio is distorted. The completeness and uniqueness of various tags in the metadata are also verified. Invalid content that fails the verification is removed, and qualified content is temporarily archived to ensure that the quality of the short drama content entering the subsequent stages meets the standards.

[0035] Step 3.2 involves deploying the generated short drama content, including video, audio, and associated metadata, to the target content distribution platform, and recording the plot structure tags and emotional context tags associated with the content during publication. Specifically, this includes: performing targeted format adaptation processing on the generated and verified short drama content; adjusting parameters according to the format requirements of different target content distribution platforms; adapting to the common playback formats and encoding standards supported by each platform; unifying the video resolution to 1080P and the audio sampling rate to 44.1kHz to ensure that the content can be played normally on various target platforms without affecting picture and sound quality; subsequently, after completing interface authentication and permission verification through the dedicated open interface call model of each target distribution platform, deploying the adapted short drama content in batches of 50 pieces to various mainstream short drama platforms. During the deployment of the drama content distribution platform, the unique identifier of each short drama is strictly bound to the corresponding platform's release ID. At the same time, the complete tag information associated with each short drama is extracted from the metadata generated in step 3.1, including all plot structure tags, emotional context tags, and corresponding multi-dimensional visual feature tags. This tag information, the unique identifier of the short drama, the platform release ID, the release time, the deployment platform name, and other information are synchronously stored in the tag association table of the dedicated database. After deployment, each short drama is checked one by one to see if it has been successfully launched and if the tags are displayed synchronously. Issues such as launch failures and missing tags are corrected in a timely manner to ensure that the deployment process is complete and traceable, and that each short drama and its associated tags can be presented normally on the corresponding distribution platform, facilitating subsequent interactive data association.

[0036] Step 3.3: Real-time collection of user interaction logs between the user and the published short drama content on the target content distribution platform. These interaction logs include playback, liking, commenting, and sharing behaviors. Specifically, after the short drama content is deployed to the target platform, a real-time interaction behavior collection process is initiated, collecting all user interaction logs between the user and the published short drama content on each distribution platform every 10 seconds. The interaction behaviors specifically cover playback behaviors, including complete playback, mid-play exit, repeated playback, playback progress, and playback duration; liking behaviors, such as regular liking and canceling liking; commenting behaviors, including comment content, comment duration, and comment sentiment; and sharing behaviors, including sharing... The data includes information on channels and sharing frequency; the collected log information is detailed, including the user's unique identifier, interaction timestamp, interaction type, unique identifier of the corresponding short drama content, interaction duration, and related supplementary information such as comment content and sharing channel; during the collection process, an invalid data filtering mechanism is activated in real time to remove accidental likes (likes with a duration of less than 1 second), blank comments, duplicate submissions of the same interaction behavior, and maliciously obtained fake interaction data; valid interaction logs are organized and classified in real time, and stored using a combination of local caching and temporary cloud backup, with cached data synchronized every 30 minutes to prevent data loss, ensuring that the collected data is authentic, complete, valid, and usable for subsequent extraction of key feedback indicators.

[0037] Step 3.4 involves temporally associating and fusing the interaction behavior logs with the plot structure tags, emotional context tags, and visual features recorded at the time of publication to form the user interaction behavior dataset. Specifically, this includes: extracting the unique identifier, platform publication ID, all plot structure tags, emotional context tags, and multi-dimensional visual feature tags associated with each short drama content, and grouping and archiving them according to the unique identifier of the short drama content; then, classifying and summarizing the collected and organized user interaction behavior logs according to the unique identifier of the short drama content to ensure centralized management of all interaction behavior logs for the same short drama; subsequently, accurately temporally associating each interaction behavior log with the corresponding tags of the short drama content according to the timestamp, and combining this with the short drama playback progress nodes to precisely map the user's interaction behavior at a certain moment (such as a like or comment at a certain time) to the corresponding plot segment, plot structure tag, and emotional context tag of the short drama. The system first establishes a point-to-point association between interactive behaviors and tags. After association, the interactive behavior data and tag information are merged and processed to unify the data format, supplement information such as data association identifiers, association time, and data source platform, and establish an interactive behavior-tag association lookup table. Then, a data verification process is initiated to remove invalid data with association errors (such as mismatch between the unique identifier of the short drama corresponding to the interactive behavior and the unique identifier of the tag), disordered time sequence (such as the interaction time being earlier than the short drama release time), and missing data. The merged data is then deduplicated and standardized. Finally, all valid data after association and fusion are integrated to form a user interaction behavior dataset with a unified format, accurate tag association, clear time sequence, complete association information between user interactive behaviors and short drama tags, and covering the interaction characteristics of different short drama types and different user groups. At the same time, the dataset is verified for completeness to ensure that each short drama content has corresponding interactive behavior data.

[0038] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1 involves performing density analysis on the interaction time-series data associated with plot structure tags in the user interaction behavior dataset. This identifies continuous time periods where the interaction frequency exceeds a preset threshold, serving as the peak density interval for the interaction data. Specifically, this includes filtering and removing abnormal interactions and invalid noise data from the user interaction behavior dataset, determining the criteria for abnormal data. Specifically, actions such as playing, liking, and commenting with a single dwell time of less than 2 seconds are considered accidental touches. Repeated likes or shares by the same user on the same short drama more than 10 times within one minute are considered high-frequency, repetitive actions that inflate engagement metrics within a short period. Simultaneously, abnormally high-frequency interaction data occurring during inactive periods (0:00 AM to 6:00 AM) are removed to ensure that the retained data is uniform. This involves providing genuine and valid user interaction data. Then, through a two-way association between the short drama's unique identifier and plot timestamps, it accurately extracts interactive time-series data that is bound to various plot structure tags, such as beginning, development, climax, and ending. This data specifically covers playback progress time-series, including the playback progress distribution per second and the completion rate of each plot segment; like time-series, including the specific time of each like and the corresponding plot node; comment posting time-series, such as the precise time of comment posting and the plot segment associated with the comment content; sharing trigger time-series, including the trigger time of the sharing behavior and the corresponding plot position; and progress bar dragging time-series, including the start and end times of dragging and the plot tag corresponding to the stopping position, ensuring that the binding between time-series data and plot structure tags is accurate.

[0039] Next, considering the 1-3 minute duration of each short drama, a fixed and uniform 5-second time slice was set. The playback cycle of the entire short drama was divided into time slices, and each time slice was associated with a corresponding plot structure tag. Then, the overall interaction frequency of all users within each time slice was counted, and different types of interaction behaviors were assigned differentiated weights: comments had the highest weight, followed by sharing, then likes, and basic playback had the lowest weight. The overall interaction frequency of each time slice was calculated by summing these weights, forming a continuous and uninterrupted interaction density time series. Then, using the average overall interaction frequency of all time slices of the entire short drama as a benchmark, the sequences were first removed from the series. Extremely high values ​​(1.8 times higher than the average) and extremely low values ​​(0.2 times lower than the average) are used to calculate the average of the interaction frequencies of all remaining time slices. This average is then used as the final baseline average. Differentiated floating correction coefficients are set according to the type of short drama: the correction coefficient for romance short dramas such as urban sweet romance and period dramas is set to 1.4, and the correction coefficient for conflict short dramas such as suspense and crime dramas is set to 1.6. The calculation method is: interaction frequency preset threshold = baseline average × correction coefficient. By multiplying the baseline average by the correction coefficient of the corresponding type, the preset threshold of interaction frequency suitable for the current short drama type is calculated to ensure that the threshold can accurately identify plot segments with high user attention.

[0040] Then, the interaction density time series is traversed segment by segment to identify time periods where the overall interaction frequency is consistently higher than the preset threshold, the number of continuous uninterrupted time slices is no less than 6 (i.e., continuous duration is no less than 30 seconds), and the total duration is no more than 1 minute (fitting the short and fast-paced plot characteristics of short dramas). Multiple peak time periods that are adjacent and no more than 5 seconds apart are merged. The start and end boundaries of the merged time periods are calibrated, and time slices with interaction frequencies below the threshold at the boundaries are removed to ensure that the interaction frequency in the entire interval meets the threshold requirements. Finally, for each calibrated time period, the corresponding plot structure tag, plot timestamp accurate to the second, average interaction frequency, and interaction behavior distribution, including the proportion of likes, comments, and shares, are accurately bound. All qualified time periods are sorted according to the interaction frequency to form a standardized, traceable, and accurate peak density interval of interaction data that can reflect the plot segments with high user attention.

[0041] Step 4.2 involves performing dimensionality reduction and clustering analysis on the sentiment vectors associated with sentiment context tags in the user interaction behavior dataset. This extracts the central points of each sentiment category in the high-dimensional vector space, serving as the high-dimensional cluster centers. Specifically, this includes: preprocessing the sentiment-related data in the user interaction behavior dataset, removing invalid sentiment data with empty comments, no effective sentiment direction, or zero interaction volume, and filtering isolated interaction data unrelated to sentiment context tags to ensure the authenticity and validity of subsequent analysis data; then, accurately extracting multi-dimensional sentiment vectors associated with various sentiment context tags such as joy, tension, sadness, sweetness, anger, and warmth from the preprocessed user interaction behavior dataset. These vectors are composed of multiple layers of key features, including the encoded features of user comments after sentiment semantic parsing, the overall completion rate of corresponding sentiment tag segments, the proportion of repeated playbacks to total playbacks, the ratio of interaction conversions to playbacks, and the matching degree between comment sentiment and tag sentiment. These multiple features are fused according to a preset ratio to form a complete multi-dimensional sentiment vector, which comprehensively and realistically reflects users' acceptance and preference for different emotional expressions in short dramas.

[0042] Subsequently, a feature-preserving dimensionality reduction method based on feature contribution was adopted. First, the variance contribution and mutual information value of each feature in the high-dimensional sentiment vector for distinguishing different sentiment categories were calculated. The two sets of values ​​were weighted and summed to obtain the comprehensive contribution of each single-dimensional feature. Then, all features were sorted from high to low based on their comprehensive contribution, prioritizing the retention of the top 80% of core features and removing the bottom 20% of redundant features. This effectively reduced the vector dimension and computational complexity of subsequent clustering analysis without losing core sentiment distinguishing features or reducing sentiment category discriminability. Finally, the set sentiment category labels were used as... Under strict constraints, the dimensionality-reduced sentiment vectors are first divided into different initial groups according to their corresponding sentiment labels. Then, within each initial group, the similarity of vectors is determined by Euclidean distance in the vector space. The smaller the distance, the closer the sentiment features between the vectors. At the same time, a fixed similarity threshold is set in combination with the emotional expression characteristics of short dramas. Sensation vectors within the same initial group whose spatial distribution distance is less than the preset similarity threshold are grouped into the same cluster. The average distance from all vectors in each cluster to the cluster center is calculated, and isolated outlier vectors whose distance from the cluster center exceeds twice the average distance are removed to ensure the feature purity and category consistency of each cluster.

[0043] We calculate the weighted center of all sentiment vectors within each cluster, first defining a unified mathematical notation: Let there be a total of n The number of valid sentiment vector samples, the first i The total effective interactions corresponding to each sample are: The total effective interactions of all samples in the cluster are , No. i The weighting coefficients for each sample are , No. i The reduced-dimensionality sentiment vectors are , d The dimension of the vector after dimensionality reduction. For vector number 1 k The dimensional components, and the final weighted center vector are: Then calculate sequentially according to the standard formula: The first step is to calculate the weighting coefficient of a single sample, using the formula: The second step is to calculate the components of each dimension of the weighted center vector, using the following formula: ;because It can be simplified to After performing calculations dimension by dimension, the weighted agglomeration center vector of the cluster is obtained and determined as the high-dimensional cluster center of the corresponding emotion category.

[0044] Subsequently, each high-dimensional cluster center undergoes a dual validity check. On the one hand, the semantic matching degree between the vector features and the preset sentiment category labels is checked; on the other hand, the rationality of the center vector in spatial distribution is checked to ensure that both indicators meet the preset qualification standards. Then, the high-dimensional cluster centers are bidirectionally bound to the corresponding sentiment context labels and sentiment intensity labels (mild, moderate, and high) and assigned unique identifiers. The spatial coordinates, core feature dimensions, and associated label information of the cluster centers are fully recorded to form traceable, quantifiable, and accurate core indicators that reflect users' sentiment preferences, providing a stable and reliable basis for sentiment features for subsequent multi-dimensional spatial mapping.

[0045] Step 4.3 involves performing gradient detection on the user attention heat distribution map associated with multi-dimensional visual features in the user interaction behavior dataset, locating the pixel coordinates of the attention heat distribution map where the gradient change is zero and the second derivative is negative, as local maxima. Specifically, this includes: firstly, pre-cleaning and standardizing the visual interaction data in the user interaction behavior dataset, removing mis-touch focus data with a single frame dwell time of less than 0.5 seconds, data that repeatedly plays the same frame at high frequency in a short period of time, and blurry visual comment data without clear image direction, retaining only valid visual interaction data with genuine user intent, specifically including the average user dwell time corresponding to each frame, the total number of times a single frame is repeatedly played, the proportion of local image focus area, and the precise frame timestamps bound to comments related to visual features.

[0046] Subsequently, based on the aforementioned effective visual interaction data, a multi-factor weighted fusion method was used to calculate the frame-by-frame user attention intensity value. First, minimum and maximum normalization processing was performed on each factor's data, mapping the original data uniformly to the range of 0 to 1. The normalization rule was: Normalized factor value = (Original factor value in the current frame - Original minimum value of this factor across all frames) ÷ (Original maximum value of this factor across all frames - Original minimum value of this factor across all frames). Then, according to a preset weight allocation rule, fixed differentiated weights were assigned to the frame dwell time factor, repeated playback count factor, local focus percentage factor, and visual comment association factor. The sum of the four weights was... 1. The frame-by-frame user attention heat value is calculated using a weighted summation formula: Frame-by-frame user attention heat value = (normalized value of frame dwell time × corresponding weight) + (normalized value of repeated playback number × corresponding weight) + (normalized value of local focus ratio × corresponding weight) + (normalized value of visual comment association × corresponding weight). This formula yields the standardized attention heat value for each frame. With the frame sequence as the horizontal axis and the standardized heat value as the vertical axis, a continuous and smooth, frame-by-frame user attention heat distribution map is generated. This distribution map can intuitively quantify the distribution of user attention intensity in different frames and different local visual areas.

[0047] Next, a full-frame, point-by-point, frame-by-frame discrete gradient traversal detection is performed on the heat map. First, the mathematical symbols and discrete calculation rules are defined: let the discrete time axis variable be the frame number. t ( t =1 , 2 , … , T, where T is the total number of frames in the short drama), the discrete pixel coordinates of the screen space axis are ( x,y ), H ( t,x,y () is the frame number t Pixel coordinates ( x,y The user attention value at point ) is represented by the first-order forward difference along the time axis as the rate of change of attention over time. The first-order forward difference along the horizontal axis of space is divided into The first-order forward difference along the vertical axis of space is divided into: The discrete first-order gradient vector is denoted as: The second-order forward difference of the time axis is the curvature of the heat change over time: The second-order forward difference classification of spatial axes is the same. The core criterion for determining local maxima is that all components of the first-order gradient are 0 and the second-order difference is less than 0, i.e. , , and A first-order gradient of 0 indicates that the heat value is at a critical point of increasing or decreasing trend, while a second-order difference less than 0 indicates that the critical point is a local peak (heat peak). All points are traversed point by point along the time and space axes. Candidate coordinate points are then selected based on the above criteria.

[0048] Then, outlier verification is performed on the selected local peak coordinate points. Taking each candidate point as the center, three consecutive frames before and after it, along with the surrounding local pixel area, are selected as the neighborhood. The average heat value and heat standard deviation within the neighborhood are calculated. Isolated outlier extreme points with heat values ​​more than twice the average heat value of the neighborhood and without other peak support within the neighborhood are removed. Only valid coordinate points with common group interaction characteristics and continuous and stable neighborhood heat distribution are retained and officially identified as local maxima. Finally, each local maxima is matched with a corresponding precise frame number and playback timestamp. Through unique data identification, the local maxima are precisely bound one-to-one with the multi-dimensional visual feature tags such as scene type, color style, shot composition, character characteristics, and shot language already labeled in step 1. The heat value, spatial location, and temporal location of the maxima are recorded simultaneously, clarifying the core visual feature carriers with the highest user visual attention and providing a complete visual feature basis for subsequent indicator point mapping.

[0049] Step 4.4 maps the peak density range of interactive data, high-dimensional cluster centers, and local maxima points to a unified multi-dimensional quantity space, forming a coordinate set of multiple key feedback indicator points. Specifically, this includes: the unified multi-dimensional quantity space is mathematically defined as the same high-dimensional space as the semantic vector space used to align text and visual labels in Step 2.1, set to 768 dimensions in this embodiment. To achieve the fusion measurement of multi-source heterogeneous feedback indicators, this space is further logically divided into three non-overlapping sub-dimensional regions, each carrying different types of features; a unified multi-dimensional quantity space consistent with the semantic encoding and model training described above is constructed, with the space dimension set to 768 dimensions, completely consistent with the specifications of the semantic vector space described above, and it is further divided into a temporal feature subspace, an emotional semantic feature subspace, and a visual feature subspace. The feature subspace consists of three non-overlapping sub-dimensional regions, each corresponding to a feature dimension of a type of indicator. It possesses a unified dimensional benchmark and measurement rules, ensuring compatibility with the mapping requirements of temporal features, sentiment semantic features, and visual features without dimensional conflicts. Furthermore, dimensionless standardization is applied to the temporal features (such as peak duration and average interaction frequency) corresponding to the peak density intervals of interaction data, the sentiment features (such as the components of the sentiment vector) corresponding to the high-dimensional cluster centers, and the visual features (such as popularity value and spatial coordinate components) corresponding to local maxima. The z-score standardization method is used to calculate the mean and standard deviation of each type of indicator feature value to eliminate differences in numerical range and distribution patterns among different types of indicators, ensuring comparable spatial measurement relationships after mapping the three types of indicators.

[0050] Then, following the preset spatial mapping rules, the key feedback indicators are mapped one by one: time-series features correspond to the specified dimensions of the time-series feature subspace, emotional features correspond to the specified dimensions of the emotional semantic feature subspace, and visual features correspond to the specified dimensions of the visual feature subspace. Based on feature importance priority, the three types of standardized key feedback indicators are mapped to a unified multi-dimensional quantity space, generating their respective spatial coordinates. Simultaneously, the original indicator information corresponding to each coordinate is recorded. Next, all generated spatial coordinates undergo refined processing. First, duplicate coordinates are identified and deduplicated based on the standard that the values ​​of each dimension of the coordinates are completely consistent. Then, the average distance of coordinates of the same type is calculated, and abnormal coordinates deviating from this average distance by more than twice are eliminated. Subsequently, it is verified whether each coordinate falls within the corresponding feature subspace to complete the rationality check. Finally, the verified coordinates are grouped and archived according to three categories: plot structure tags, emotional context tags, and visual feature tags. A unique identifier corresponding to the corresponding tag is added to each group of coordinates, and the original indicator details corresponding to the coordinates are synchronously associated, forming a complete, standardized, traceable set of key feedback indicator point coordinates that can be used for subsequent geometric analysis.

[0051] Step 4.5: Based on the coordinate set, determine the main distribution direction and range of the point set through principal component analysis, and generate a minimum enclosing ellipse accordingly. The boundary defined by the parametric equation of this ellipse is used as the feature vector boundary. Specifically, this includes: performing refined principal component analysis on the coordinate set of key feedback index points; firstly, calculating the variance contribution of each dimension of the coordinate set; sorting the variance contribution from high to low; accumulating the cumulative variance contribution; setting a cumulative variance contribution of ≥85% as the principal component screening threshold; selecting the top two principal component dimensions that meet this threshold, denoted as the first principal component PC1 and the second principal component PC2; and orthogonally projecting all effective index points in the high-dimensional space onto the two-dimensional principal component composed of PC1 and PC2. The plane is used to determine the main extension direction, concentrated distribution area, and overall discrete range of the point set within the plane. Distribution interference caused by weak contribution redundant dimensions is eliminated, and the core distribution parameters and projected coordinates after principal component analysis are recorded. Based on the distribution characteristics and spatial boundaries after principal component projection, a robust fitting algorithm based on least squares is adopted. First, a small number of isolated outliers remaining in the coordinate set are eliminated to avoid outliers affecting the fitting accuracy. Then, the core distribution area of ​​the point set is focused, and all effective key feedback index points are fitted in a bounding manner. During the fitting process, the fitting parameters are continuously adjusted to ensure that the generated ellipse can completely enclose all effective projected point sets while minimizing the range, that is, to obtain the minimum bounding ellipse in the two-dimensional principal component plane.

[0052] Then, the complete geometric parameters of the minimum enclosing ellipse are extracted, and the standard plane parametric equation is given. First, the mathematical notation is defined: Let the center coordinates of the ellipse on the two-dimensional principal component plane be... The length of the major semi-axis is a The length of the minor semi-axis is b The rotation angle between the major axis of the ellipse and the first principal component axis PC1 is . θ (Radian measure), parameter variables are Then the principal component coordinates of any point on the ellipse satisfy the parametric equation system: ; in x ( t Let be the projected coordinates of any point on the ellipse onto the first principal component PC1. y ( t Let be the projected coordinates of any point on the ellipse onto the second principal component PC2. , These are the coordinates of the ellipse center on the PC1 and PC2 axes, respectively, corresponding to the core distribution center of user comprehensive preferences. a Let be the major semi-axis of the ellipse, representing the maximum discrete range of user preferences along the main extension direction. b Let be the minor semi-axis of the ellipse, representing the maximum discrete range of user preferences along the secondary extension direction. θThe rotation angle of the major axis of the ellipse relative to the PC1 axis represents the main distribution direction of user preferences. t The traversal parameter can be set to a value from 0 to 2π to generate a complete closed elliptical boundary.

[0053] Next, the validity of the fitted minimum enclosing ellipse is verified. Each valid projection point is checked to ensure it is completely enclosed by the ellipse, and the fit between the ellipse boundary and the distribution trend of the point set is verified. If any valid points are not enclosed or the fit is insufficient, the fitting parameters are readjusted and the fit is repeated until the verification criteria are met. Finally, the continuous closed ellipse boundary defined by the above geometric parameters and parametric equations is determined as the feature vector enclosure corresponding to this iteration cycle. In this invention, the feature vector enclosure specifically refers to the closed curve uniquely determined by the parametric equation of the minimum enclosing ellipse on the two-dimensional plane spanned by the first two principal components after dimensionality reduction through principal component analysis. This closed curve constitutes the outer envelope boundary of the distribution range of key feedback index points on this plane, used to quantify the overall distribution range of user preferences in a unified low-dimensional subspace. A unique identifier is assigned to this enclosure, binding the current iteration cycle information and principal component analysis parameters, clarifying that this enclosure can quantify the comprehensive preference range of users in the three dimensions of plot, emotion, and vision. This provides a stable and accurate feature boundary for subsequent calculations of topological inclusion and measure overlap ratio, ensuring that the subsequent semantic distillation process accurately matches changes in user preferences.

[0054] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: Obtain the two feature vector boundaries corresponding to the current iteration cycle and the previous iteration cycle respectively; whereby each feature vector boundary is defined by the corresponding elliptic parametric equation, specifically including: first confirming the unified definition rules of the iteration cycle, each iteration cycle is defined according to a preset standard, which can be a fixed time interval (such as once a day) or a fixed amount of user interaction (such as a cumulative preset number of valid user interaction behaviors) as the iteration node, and the current iteration cycle is denoted as the . k The iteration number is denoted as the previous iteration period. k -1 iterations, with each iteration corresponding to an independent and complete user interaction behavior dataset, principal component analysis results, and eigenvector contours, ensuring the independence and comparability of data between iterations; subsequently, the -1 iteration is retrieved. k The iteration and the k The complete information of the feature vector bounding channel corresponding to the -1 iteration is obtained. Each feature vector bounding channel is uniquely defined by the minimum bounding ellipse parametric equation generated by the fitting. During the retrieval process, all geometric parameters of the two ellipses are extracted simultaneously, including the coordinates of the ellipse center on the two-dimensional principal component plane, the length of the major semi-axis, the length of the minor semi-axis, the rotation angle, as well as the corresponding two-dimensional principal component plane projection parameters, iteration period identifier, and other related information.

[0055] Next, a comprehensive integrity and consistency check is performed on the retrieved parameters and parametric equations. This includes verifying not only whether parameters are missing and whether the equation format is consistent with the standard in step 4.5, but also checking the rationality of the parameters, such as whether the length of the major semi-axis is greater than the length of the minor semi-axis, or whether the rotation angle is within the range of (0, π) radians. If any parameter errors or format deviations are found, the backup data for the corresponding iteration period is immediately retrieved for correction. After ensuring that there are no deviations, the ellipse for the current iteration period (denoted as ellipse) is finally determined. k The ellipse of the previous iteration period (denoted as ellipse) k The standard parametric equation of the ellipse (-1), where the ellipse k The parametric equations are as follows: ; oval k The parametric equation for -1 is as follows: ; The meanings of the characters in the formula remain consistent with those in step 4.5, only the subscripts are added. k , k -1 distinguishes the iteration period. t The range is (0, 2π).

[0056] Step 5.2: Based on the ellipse parametric equation of the current iteration period and the ellipse parametric equation of the previous iteration period, calculate the area ratio of the ellipse in the current iteration period to the ellipse in the previous iteration period, and quantify the area ratio as the topological inclusion degree. Specifically, this includes: based on the determined standard parametric equations of the two ellipses, making a preliminary judgment by combining the center coordinates, semi-major axis, semi-minor axis, and rotation angle of the ellipse, and then further refining the determination of the ellipse by traversing the coordinates of key nodes on the ellipse through the parametric equations. k With ellipse k The spatial relationship of -1 can be specifically divided into three cases: ellipse k Completely elliptic k -1 contains, ellipse k With ellipse k -1 partial intersection, ellipse k With ellipse k -1 ensures complete separation, guaranteeing the accuracy of positional relationship determination; subsequently, for different spatial positional relationships, corresponding calculation methods are employed to accurately calculate the ellipse. k Ellipse k The area proportion contained in -1 is formally quantified as the topological coverage, and the quantification formula is as follows: ;in T Represents the topological containment degree, with values ​​ranging from 0 to 1. Represents an ellipse k With ellipse k-1 is the area of ​​the intersection. Represents the current iteration period ellipse k The total area, the formula for the area of ​​an ellipse is: S=πab ,Right now If ellipse k Completely elliptic k -1 indicates that the current user's preference range is completely within the preference range of the previous period. Topological containment T =1; If the two are completely disjointed, it means that the current user preference does not overlap with the preference of the previous period. Topological containment T =0; If the two partially intersect, it indicates that the current user preference overlaps with the previous period's preference. In this case, a numerical integration method is used, combining the parametric equations of the two ellipses for precise integration calculation to obtain the intersection area. Substituting these values ​​into the formula yields a precise area ratio, which serves as the final topological inclusion degree. This metric specifically quantifies the current user's preference range in terms of plot, emotion, and visual dimensions, and the degree to which this preference range is included compared to the previous iteration. T The closer it is to 1, the more closely the current preference matches the preference of the previous cycle, and the smaller the change in preference.

[0057] Step 5.3: Based on the elliptic parametric equations of the current iteration cycle and the previous iteration cycle, calculate the ratio of the intersection area to the union area of ​​the two ellipses, and quantify this ratio as a measure of overlap ratio. Specifically, this includes: continuing to use the two determined elliptic parametric equations and the already precisely calculated intersection area. Next, calculate the area of ​​the union of the two ellipses; before calculating, first confirm the ellipse area. k With ellipse k The total area calculation of -1 is correct, echoing the ellipse area formula mentioned earlier, ensuring... , Calculate accurately, then substitute into the formula for calculating the area of ​​the union. ,in Represents the periodic ellipse of the previous iteration k The total area of ​​-1 can be used to accurately obtain the overall area covered by the two ellipses; then, the overlap ratio is defined as the ratio of the intersection area to the union area, and the quantification formula is: ,in O This represents the overlap ratio, ranging from 0 to 1. If the two ellipses completely overlap, it indicates that the user preferences in the two consecutive periods are completely consistent. Measure the overlap ratio O =1; If the two are completely disjointed, it means that there is no overlap in user preferences between the two periods. O=0; If the two partially intersect or have a single inclusion relationship, the corresponding area value is substituted to accurately calculate the ratio, which is used as the final measure of overlap ratio; This indicator can comprehensively quantify the overall overlap of user preference ranges in two iteration cycles. Compared with the single inclusion perspective of topological inclusion degree, it takes into account the bidirectional overlap characteristics and can more objectively reflect changes in preferences. O The closer the value is to 1, the smaller the change in user preferences between the two periods, and the higher the overlap. O The closer to 0, the greater the change in preferences and the lower the overlap.

[0058] Step 5.4 involves weighted fusion of the topological inclusion degree and the measure overlap ratio to generate a comprehensive quantitative index, which serves as the semantic distillation intensity coefficient. Specifically, this includes: combining the core objectives of semantic distillation, selecting corpora that align with current user preferences while also considering historical preferences; and setting differentiated weighting coefficients for the topological inclusion degree and the measure overlap ratio, where the topological inclusion degree weight... Measure the overlap ratio weight The sum of the weights of the two is strictly 1. The core basis for setting the weights is that topological inclusion focuses on reflecting the one-way relationship where the current preference is included by the preference of the previous cycle, while the overlap ratio measure takes into account the two-way overlap between the preferences of the two cycles, and can more comprehensively reflect the overall trend of preference changes. It has a greater impact on the determination of semantic distillation strength, and therefore is given a higher weight. Subsequently, a weighted summation method is used to calculate the weight of the topological inclusion. T Overlap ratio with measure O The data is fused to generate a comprehensive quantitative index, which serves as the semantic distillation intensity coefficient. The fusion formula is as follows: ,in This represents the semantic distillation intensity coefficient, ranging from 0 to 1; after fusion, the coefficient is... Perform validity verification; if the error is due to calculation error... If the value exceeds the reasonable range of 0 to 1, a linear normalization method is used to correct it to this range, ensuring the validity and reasonableness of the coefficient. This coefficient can directly quantify the change in user preferences between two iterations, and its value directly corresponds to the strength strategy of subsequent semantic distillation. The closer it is to 1, the smaller the change in user preferences and the lower the semantic distillation intensity. There is no need for extensive screening of the corpus; only minor adjustments are required. The closer the value is to 0, the greater the change in user preferences and the higher the semantic distillation intensity. It is necessary to focus on screening corpora that fit the current user preferences and remove historical corpora that deviate significantly from the current preferences. This will provide a precise and practical basis for corpus screening and intensity control in the subsequent semantic distillation process.

[0059] In a preferred embodiment of the present invention, step 6 above may include: Step 6.1: Based on the semantic distillation intensity coefficient, assign dynamic priority weights to each data unit in the annotated corpus; wherein, the higher the semantic distillation intensity coefficient, the greater the priority weight of the corresponding data unit. Specifically, this includes: first determining the semantic distillation intensity coefficient. The core definition of is that its value ranges from 0 to 1. The closer to 1, the smaller the change in user preferences and the lower the semantic distillation intensity. The closer the value is to 0, the greater the change in user preferences and the higher the semantic distillation intensity. Subsequently, all data units of the annotated corpus were analyzed. Each data unit corresponds to annotated corpus related to the plot structure, emotional context, and multi-dimensional visual features of the short drama, and is bound with a unique identifier, feature label, and matching degree value with the feature vector of the current iteration cycle. M , M∈ [0 , 1], M The closer the value is to 1, the higher the match between the data unit and the current user's preference; next, the calculation rules for dynamic priority weights are set to calculate a unique dynamic priority weight for each data unit. W The calculation formula is: ,in This is the basic priority weight for the data unit, fixed at 0.5, representing the basic corpus value of the data unit itself. The semantic distillation intensity coefficient. M The formula represents the matching degree between the data unit and the current feature vector. Its meaning is: when... When the value is relatively large (user preferences change little), the weight W More dependent on matching degree M That is, they tend to choose corpora that match their current preferences; when When the weight is small (user preferences change greatly), W More dependent on basic weights This means it tends to retain corpora with fundamental value; after calculation, all weight values ​​are normalized to the range of 0 to 1. Therefore, the semantic distillation intensity coefficient... Higher and better matching M Higher data units have higher dynamic priority weights. W Generally, the larger the value, the better; after calculation, all weight values ​​are normalized to the range of 0 to 1 to ensure the semantic distillation intensity coefficient. The higher the match, the better. M The higher the data unit, the greater its final dynamic priority weight. W The larger the value, the more accurately the dynamic priority weights of all data units can be set.

[0060] Step 6.2 involves filtering the labeled corpus based on a set dynamic priority weight to select data units with priority weights higher than a preset threshold, forming a high-value training subset. Specifically, this includes: first, setting a preset threshold for the dynamic priority weight based on the size of the dedicated corpus and the needs of incremental model training. W threshold (Initially set to 0.7, dynamically adjustable based on iteration results), this threshold needs to balance the accuracy of corpus selection with the size of the training subset; then, iterate through all data units in the labeled corpus with dynamically assigned priority weights, comparing the weight of each data unit one by one. W With preset threshold W threshold Preliminary screening of weights W ≥ W threshold The data units are then subjected to a second validity check, which verifies the completeness of their annotation information and the authenticity of their matching degree with the current feature vector. Abnormal data units with missing annotations or falsified matching degrees are removed. Finally, the qualified data units are sorted from high to low according to dynamic priority weights, and the complete annotation information, weight values ​​and associated feature labels of each data unit are retained to form a structured, high-value training subset that can be directly used for model training.

[0061] Step 6.3 involves incrementally training the supporting model of the dedicated corpus using a high-value training subset to optimize model parameters and obtain a parameter-optimized supporting model. Specifically, this includes: first, determining the core positioning of the supporting model. This supporting model is essentially a specific application instance or variant of the aforementioned performance-optimized basic generative model in corpus annotation and user preference prediction tasks. It serves as the core supporting model for the dedicated corpus and can be deployed as a short drama user preference prediction supporting model based on the Transformer architecture. It is responsible for connecting to the high-value training subset, optimizing the extraction and matching capabilities of user preference features, and maintaining consistency with the network structure and semantic encoding rules of the previously trained model. Next, the supporting model optimized in the previous iteration is retrieved. During the retrieval process, the model's integrity is verified, checking the completeness of model parameters and the consistency of the network structure to ensure no missing parameters or structural anomalies. Historical training parameters, including learning rate, number of training epochs, and loss value, are recorded to provide a reference for this incremental training. Finally, the complete set of core parameters and constraints for incremental training is determined. The optimizer uses mini-batch gradient descent, and the learning rate is set. η=0.001 (can be dynamically adjusted), training batch size is 32 (to adapt to computing power requirements and avoid memory overflow), training epochs are 10, and constraints are set: freeze the parameters of the model's embedding layer and basic feature extraction layer, and only update the parameters of the model's fully connected layer, attention layer and preference prediction layer. The purpose of this constraint is to retain the effective feature extraction capabilities accumulated by the model in the past training and avoid model overfitting or performance degradation caused by incremental training.

[0062] Subsequently, the high-value training subset underwent data preprocessing, randomly dividing it into training and validation sets at an 8:2 ratio to ensure consistent distribution between the two sets of data. Both sets covered the three feature labels of plot, emotion, and visual, without any class imbalance. During preprocessing, the data was also standardized to unify the data format, facilitating model reading and training. The training set was then input into the supporting model to initiate incremental training. During training, key metrics were monitored in real time, including the cross-entropy loss function value and goodness of fit of the training set. After each round of training, the preference prediction accuracy of the training set was calculated, and an early termination strategy was implemented: if the loss value did not decrease and the accuracy did not improve for three consecutive rounds of training, training was terminated early to avoid ineffective training. After training, the model is comprehensively validated using a validation set. The core validation metrics are user preference prediction accuracy and feature label matching rate. The preset validation thresholds are accuracy ≥ 85% and label matching rate ≥ 88%. If both metrics meet the standards, the model is determined to be the support model after parameter optimization. If one or both metrics fail to meet the standards, the training parameters are adjusted, the learning rate is adjusted to 0.0005 or 0.002, the number of training epochs is increased to 15, and incremental training is restarted until all validation metrics meet the preset standards. After the standards are met, the model parameters are archived, and the current iteration cycle identifier, training parameters, and validation effect data are bound to them for subsequent tracking and iterative optimization.

[0063] Step 6.4 involves synchronously updating the newly labeled data output by the parameter-optimized support model, along with the valid information from the user interaction behavior dataset generated in this iteration, to the dedicated corpus, completing the closed-loop optimization process. This specifically includes: first, finely filtering and processing the newly labeled data output by the parameter-optimized support model to confirm the specific type of the new labeled data, including plot structure labels, emotional context labels, and multi-dimensional visual feature labels added by the model to the unlabeled short drama corpus, as well as the confidence scores for each label. The confidence scores are determined by the model based on specific... The feature matching degree is automatically calculated; then, the confidence level of the newly labeled data is verified. The confidence level threshold is set to 0.8, and the confidence level of each newly labeled data is checked one by one. Only the valid new labeled data with a confidence level ≥ 0.8 is retained, and the low-quality labeled data with a confidence level below 0.8 is removed to avoid low-quality data from polluting the dedicated corpus; the retained valid new labeled data is standardized by binding a unique label identifier, current iteration cycle information, model version information and corresponding confidence score to it, and it is archived according to feature labels to ensure that the data format is consistent with the existing data in the dedicated corpus.

[0064] Next, the effective information in the user interaction behavior dataset generated in this iteration was sorted out. During the screening process, abnormal interaction data and invalid noise data, such as accidental touch behavior and data scalping behavior, were strictly removed. The effective information extracted specifically includes: peak density intervals of interaction data and corresponding plot structure labels, high-dimensional cluster centers and corresponding emotional context labels, local maxima and corresponding visual feature labels, as well as feature vectors surrounding complete geometric parameters, topological inclusion degree and measure overlap ratio. This effective information was classified according to data type, and the corresponding iteration cycle identifier and calculation basis were added. After standardization, structured data was formed. Subsequently, the update process of the dedicated corpus was started. The processed effective new labeled data was inserted into the corresponding category directories of the dedicated corpus according to category: plot labeled corpus directory, emotional labeled corpus directory, and visual labeled corpus directory. The effective information of user interaction behavior in this iteration was stored in the interaction data archive directory of the corpus. The version number and iteration cycle identifier of the corpus were updated synchronously, and the correlation between data was improved, such as binding the new labeled data with the corresponding user interaction features to ensure data traceability during subsequent retrieval.

[0065] Finally, a full integrity and consistency check was performed on the updated dedicated corpus. The check process consisted of three steps: First, the integrity of the newly added data was checked, verifying whether data identifiers, annotation information, and parameter information were missing; second, the consistency of the data was checked, verifying whether the format and terminology of the newly added data were consistent with the existing data in the corpus, and whether the feature tags matched; third, the uniqueness of the data was checked, removing duplicate data to ensure that the corpus was free of redundancy. After the check passed, the corpus update log was recorded, specifying key information such as update time, updated data volume, and iteration cycle, completing the update of the dedicated corpus. Thus, a closed-loop optimization process was formed, encompassing user interaction data collection and analysis, feature vector path construction, topological inclusion degree and measure overlap ratio calculation, semantic distillation intensity coefficient generation, high-value corpus screening, incremental training of supporting models, and dedicated corpus updates. This ensured that both the dedicated corpus and supporting models could continuously adapt to changes in user preferences, improving the accuracy of subsequent short drama content generation and preference prediction.

[0066] like Figure 2 As shown, embodiments of the present invention also provide an incremental training and optimization system for a corpus specifically designed for short dramas, comprising: The corpus preprocessing module is used to process the original data of the short drama domain from the authorized source in an integrated manner through a domain-adaptive intelligent corpus pre-structured model to generate an annotated corpus; the integrated processing includes automated data quality screening and annotation of text data with plot structure tags and emotional context tags, and annotation of visual materials with multi-dimensional visual feature tags; The model optimization module is used to fine-tune the basic generative model with domain instructions and human preference reinforcement learning based on the labeled corpus, so as to optimize the performance of the basic generative model in narrative generation and cross-modal consistency, and obtain a performance-optimized basic generative model. The interactive data acquisition module is used to produce short drama content through a performance-optimized basic generation model and simultaneously collect user interaction behavior datasets related to the short drama content. The feedback feature quantization module is used to extract from the user interaction behavior dataset the peak density range of interaction data corresponding to the plot structure tags of the short drama content, the high-dimensional clustering center in the emotional context tag vector, and the local maxima in the multimodal visual feature attention heat map as multiple key feedback indicator points; based on the coordinates of the key feedback indicator points in the multi-dimensional quantity space, the corresponding feature vector contour is dynamically constructed. The semantic distillation module is used to calculate the topological inclusion degree and measure overlap ratio between feature vectors at different times; based on the topological inclusion degree and measure overlap ratio, the semantic distillation intensity coefficient is generated. The incremental training module is used to prioritize and filter the labeled corpus based on the semantic distillation intensity coefficient, and perform incremental training to update the dedicated corpus, thereby achieving closed-loop optimization. The above description is a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An incremental training and optimization method for a corpus specifically designed for short dramas, characterized in that... The method includes: By using a domain-adaptive intelligent corpus pre-structured model, the original data of short dramas from authorized sources are processed in an integrated manner to generate labeled corpus; the integrated processing includes automated data quality screening and labeling text data with plot structure tags and emotional context tags, and labeling visual materials with multi-dimensional visual feature tags; Based on the annotated corpus, the basic generative model is fine-tuned with domain instructions and reinforced with human preferences to optimize its performance in narrative generation and cross-modal consistency, thus obtaining a performance-optimized basic generative model. Short drama content is generated using a performance-optimized basic generation model, and user interaction behavior datasets related to the short drama content are collected simultaneously. From the user interaction behavior dataset, the peak density range of interaction data corresponding to the plot structure tags of short drama content, the high-dimensional clustering center in the emotional context tag vector, and the local maxima in the multimodal visual feature attention heat map are extracted as multiple key feedback index points; based on the coordinates of the key feedback index points in the multi-dimensional quantity space, the corresponding feature vector contour is dynamically constructed. Calculate the topological inclusion degree and measure overlap ratio among feature vectors at different times; generate semantic distillation intensity coefficients based on the topological inclusion degree and measure overlap ratio. Based on the semantic distillation intensity coefficient, the labeled corpus is prioritized and filtered, and incremental training is performed to update the dedicated corpus, thus achieving closed-loop optimization.

2. The incremental training and optimization method for a corpus specifically for short dramas according to claim 1, characterized in that, A domain-adaptive intelligent corpus pre-structured model is used to perform integrated processing on raw short drama domain data from authorized sources to generate labeled corpora. This integrated processing includes automated data quality screening and labeling text data with plot structure and emotional context tags, as well as labeling visual materials with multi-dimensional visual feature tags, including: By using a domain-adaptive intelligent corpus pre-structured model, we can access and parse the original data stream of short dramas from authorized sources. The domain-adaptive intelligent corpus pre-structured model performs automated quality screening and copyright compliance verification on the original data stream to obtain a compliant data set. Narrative structure analysis and emotional context analysis are performed on text data in the compliant dataset to generate corresponding plot structure tags and emotional context tags. Multimodal feature extraction and alignment are performed on visual materials in the compliant data set to generate corresponding multidimensional visual feature labels; The plot structure tags, emotional context tags, and multi-dimensional visual feature tags are associated with and encapsulated with the corresponding original data to generate annotated corpus.

3. The incremental training and optimization method for a corpus specifically for short dramas according to claim 2, characterized in that, Based on the annotated corpus, the basic generative model is fine-tuned using domain instructions and reinforced with human preferences to optimize its performance in narrative generation and cross-modal consistency, resulting in a performance-optimized basic generative model, including: A domain instruction fine-tuning dataset was constructed based on the annotated corpus. The domain instruction fine-tuning dataset includes short drama narrative structure examples extracted from the annotated corpus and the alignment relationship between text and visual labels in the semantic vector space. The basic generative model is fine-tuned with instruction supervision by fine-tuning the domain instruction fine-tuning dataset, so that the basic generative model learns the narrative logic and cross-modal alignment constraints of the short drama domain, and obtains the instruction-fine-tuned basic generative model. A diverse set of short drama content candidates is generated by fine-tuning the basic generative model, and human preference feedback data on the short drama content candidates is collected. A reward signal is defined based on the collected human preference feedback data. Based on the defined reward signal, a reinforcement learning approach is used to optimize the basic generative model after fine-tuning the instructions, so as to improve the alignment between its generated content and human preferences. The instructions are executed iteratively to supervise fine-tuning and policy optimization based on reward signals until the performance of the basic generative model reaches the preset optimization target, thereby obtaining the performance-optimized basic generative model.

4. The incremental training and optimization method for a corpus specifically for short dramas according to claim 3, characterized in that, Short drama content is generated using a performance-optimized basic generative model, and a dataset of user interaction behaviors associated with the short drama content is collected simultaneously, including: Creative instructions are input into the performance-optimized base generation model to drive the generation of short drama content that includes video, audio, and associated metadata. The generated short drama content, which includes video, audio, and associated metadata, is deployed to the target content distribution platform, and the plot structure tags and emotional context tags associated with the content are recorded when it is published. Real-time collection of user interaction logs with the published short drama content on the target content distribution platform, wherein the interaction behavior includes playback, likes, comments and sharing behaviors; The interaction behavior logs are temporally correlated and fused with the plot structure tags, emotional context tags, and visual features recorded at the time of publication to form the user interaction behavior dataset.

5. The incremental training and optimization method for a corpus specifically for short dramas according to claim 4, characterized in that, We extracted the peak density range of interactive data corresponding to the plot structure tags of the short drama, the high-dimensional clustering center in the emotional context tag vector, and the local maxima in the multimodal visual feature attention heat map as multiple key feedback indicator points. Based on the coordinates of the key feedback index points in the multi-dimensional quantity space, the corresponding feature vector contour is dynamically constructed, including: Density analysis is performed on the interaction time series data associated with plot structure tags in the user interaction behavior dataset to identify continuous time periods where the interaction frequency exceeds a preset threshold, which are taken as the peak density interval of the interaction data. We perform dimensionality reduction and clustering analysis on the sentiment vectors associated with sentiment context tags in the user interaction behavior dataset, and extract the agglomeration center points of each sentiment category in the high-dimensional vector space as high-dimensional cluster centers. Gradient detection is performed on the user attention heat distribution map associated with multi-dimensional visual features in the user interaction behavior dataset. The pixel coordinates points in the attention heat distribution map where the gradient change is zero and the second derivative is negative are located as local maxima points. The peak density range of interactive data, high-dimensional cluster centers, and local maxima points are mapped to a unified multi-dimensional quantity space to form a set of coordinates for multiple key feedback indicator points. Based on the coordinate set, the main distribution direction and range of the point set are determined by principal component analysis. Based on this, a minimum enclosing ellipse is generated, and the boundary defined by the parametric equation of this ellipse is used as the feature vector enclosing path.

6. The incremental training and optimization method for a corpus specifically for short dramas according to claim 5, characterized in that, Calculate the topological inclusion degree and measure overlap ratio among feature vectors at different time periods; based on the topological inclusion degree and measure overlap ratio, generate semantic distillation intensity coefficients, including: Obtain the two eigenvector contours corresponding to the current iteration cycle and the previous iteration cycle respectively; where each eigenvector contour is defined by the corresponding elliptic parametric equation; Based on the elliptic parametric equation of the current iteration period and the elliptic parametric equation of the previous iteration period, calculate the area ratio of the ellipse of the current iteration period to that of the ellipse of the previous iteration period, and quantify the area ratio as the topological inclusion degree. Based on the elliptic parametric equations of the current iteration cycle and the elliptic parametric equations of the previous iteration cycle, the ratio of the intersection area to the union area of ​​the two ellipses is calculated, and the ratio is quantified into a measure of overlap ratio. The topological inclusion degree and the measure overlap ratio are weighted and fused to generate a comprehensive quantitative index, which serves as the semantic distillation intensity coefficient.

7. The incremental training and optimization method for a corpus specifically for short dramas according to claim 6, characterized in that, Based on the semantic distillation intensity coefficient, the annotated corpus is prioritized and filtered, and incremental training is performed to update the dedicated corpus, achieving closed-loop optimization, including: Based on the semantic distillation intensity coefficient, a dynamic priority weight is assigned to each data unit in the annotated corpus; wherein, the higher the semantic distillation intensity coefficient, the greater the priority weight of the corresponding data unit. The labeled corpus is filtered based on the set dynamic priority weights to select data units with priority weights higher than a preset threshold, forming a high-value training subset; Incremental training of the support model for the dedicated corpus is performed using a high-value training subset to optimize the model parameters and obtain a parameter-optimized support model. The newly labeled data output by the optimized support model, along with the effective information from the user interaction behavior dataset generated in this iteration, are synchronously updated to the dedicated corpus to complete the closed-loop optimization process.

8. An incremental training and optimization system for a corpus specifically designed for short dramas, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The corpus preprocessing module is used to process the original short drama domain data from the authorized source in an integrated manner through a domain-adaptive intelligent corpus pre-structured model to generate annotated corpus. The integrated processing includes automated data quality screening and labeling text data with plot structure tags and emotional context tags, as well as labeling visual materials with multi-dimensional visual feature tags; The model optimization module is used to fine-tune the basic generative model with domain instructions and human preference reinforcement learning based on the labeled corpus, so as to optimize the performance of the basic generative model in narrative generation and cross-modal consistency, and obtain a performance-optimized basic generative model. The interactive data acquisition module is used to produce short drama content through a performance-optimized basic generation model and simultaneously collect user interaction behavior datasets related to the short drama content. The feedback feature quantization module is used to extract from the user interaction behavior dataset the peak density range of interaction data corresponding to the plot structure tags of the short drama content, the high-dimensional clustering center in the emotional context tag vector, and the local maxima in the multimodal visual feature attention heat map as multiple key feedback indicator points; based on the coordinates of the key feedback indicator points in the multi-dimensional quantity space, the corresponding feature vector contour is dynamically constructed. The semantic distillation module is used to calculate the topological inclusion degree and measure overlap ratio between feature vectors at different times; based on the topological inclusion degree and measure overlap ratio, the semantic distillation intensity coefficient is generated; the incremental training module is used to prioritize and filter the labeled corpus based on the semantic distillation intensity coefficient, and perform incremental training to update the dedicated corpus to achieve closed-loop optimization.

Images