Big data processing-based short play subtitle processing method and device, and medium

Abstract

The invention discloses a short play subtitle processing method and device based on big data processing and a medium, and relates to the technical field of intelligent subtitle generation, and the method comprises the steps: carrying out the big data analysis of a user behavior data set, obtaining a timestamp compression coefficient value and a behavior feature vector, and carrying out the initial correction through a TensorFlow Lite lightweight model, and obtaining the initial correction subtitle data; performing incremental federated learning processing on the optimized subtitle data sequence to obtain a federated optimization parameter set; generating a standardized subtitle file based on the federal optimization parameter set, distributing the standardized subtitle file to a short play channel through an application program interface, and outputting a direct play subtitle file; by introducing an incremental federal learning updating mechanism based on an optimized subtitle data sequence, continuous optimization and cross-device coevolution on the premise of protecting user privacy are realized, and the long-term adaptive capacity of the system to a complex viewing scene is also enhanced.

Description

Technical Field

[0001] This invention relates to the field of intelligent subtitle generation technology, and in particular to a method, device and medium for processing subtitles for short dramas based on big data processing. Background Technology

[0002] With the explosive growth of short videos and short dramas on mobile internet platforms, subtitle processing technology has become a core element in enhancing user experience. Developments in related technologies primarily focus on adaptive subtitle optimization methods based on user behavior data. For example, existing technologies (such as using supervised learning models to make initial adjustments to subtitles by collecting metrics like user viewing time, click-through rates, and pause behavior) have gradually matured. These systems typically employ offline data processing frameworks, such as the Apache Spark batch analysis platform, to integrate user behavior logs and generate subtitle correction parameters, which are then applied to dynamically adjust subtitle timing and semantic expression. Academic research indicates that such methods can improve the synchronization between subtitles and user attention, optimizing information delivery efficiency.

[0003] The key issue with existing technologies lies in the limitations of dynamic calibration mechanisms for user attention thresholds. Existing methods are typically based on preset thresholds or simplified statistical models (such as mean or variance analysis). When processing multi-source heterogeneous user behavior data, it is difficult to accurately identify the cognitive load threshold. The heterogeneity of timestamps and multimodal features in user behavior datasets are not adequately modeled, which can easily lead to threshold bias, thereby reducing the accuracy and real-time response of subtitle semantic correction. This limitation directly affects the user experience of short drama subtitle output, such as the frequent occurrence of subtitles being out of sync with the user's cognitive rhythm. There is an urgent need to introduce more refined decision calibration technologies while maintaining the existing privacy protection framework. Summary of the Invention

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

[0005] Therefore, this invention provides a short drama subtitle processing method based on big data processing to solve the problems of insufficient dynamic accuracy and privacy in attention threshold recognition.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a short drama subtitle processing method based on big data processing, comprising: collecting big data on user viewing behavior, cleaning and formatting it to generate a user behavior dataset, and calculating an initial attention index value; based on the user behavior dataset and the initial attention index value, identifying the cognitive load threshold through a Gaussian mixture model, performing dynamic calibration by combining federated learning, outputting an attention threshold, and comparing it with the initial attention index value to obtain a decision instruction; performing big data analysis on the user behavior dataset to obtain a timestamp compression coefficient value and a behavior feature vector, and performing initial calibration through a TensorFlow Lite lightweight model to obtain initial calibrated subtitle data; performing semantic calibration on the initial calibrated subtitle data using the decision instruction to output an optimized subtitle data sequence; performing incremental federated learning processing on the optimized subtitle data sequence to obtain a federated optimization parameter set; generating a standardized subtitle file based on the federated optimization parameter set, and distributing it to the short drama playback channel through an application programming interface to output a directly playable subtitle file.

[0007] As a preferred embodiment of the short drama subtitle processing method based on big data processing described in this invention, the specific steps for calculating the initial attention index value are as follows: Collect big data on user viewing behavior, identify and remove abnormal high-speed scrolling caused by accidental device touches in the big data of user viewing behavior through a sliding filter with an adaptive window width, and fill in the missing points of the big data of user viewing behavior through a time series interpolation algorithm to obtain a user behavior dataset; The dynamic weight parameters of the user behavior dataset are loaded through federated learning, and the initial attention index value is calculated using a multi-index weighted calculation formula.

[0008] As a preferred embodiment of the short drama subtitle processing method based on big data processing described in this invention, the specific steps for obtaining the decision instruction are as follows: Based on the average pause frequency, average replay ratio, and average swipe speed in the user behavior dataset, and combined with the initial attention index value, a standardized multidimensional feature matrix is ​​constructed. The standardized multidimensional feature matrix is ​​input into the Gaussian mixture model for cluster analysis to identify the boundary points of high cognitive load clusters, generate an initial attention threshold, and calculate the absolute difference between the attention index value of the terminal device and the initial attention threshold to obtain the local attention bias. The local attention bias is aggregated through a federated learning framework, and the initial attention threshold is dynamically calibrated using a gradient descent algorithm to generate the attention threshold. The initial attention index value is intelligently compared with the attention threshold. When the real-time attention index exceeds the attention threshold, a compression mode decision instruction is obtained. When the real-time attention index is less than the attention threshold, a correction mode decision instruction is obtained.

[0009] As a preferred embodiment of the short drama subtitle processing method based on big data processing described in this invention, the specific steps for obtaining the timestamp compression coefficient value and the behavior feature vector are as follows: Extract the average pause frequency, average replay ratio, and average swipe speed from the user behavior dataset, and calculate the weighted attention entropy index; The weighted attention entropy exponent is transformed using the Sigmoid function, and the timestamp compression coefficient value is output. The fluctuation amplitude of pause frequency is analyzed by sliding window standard deviation analysis, the time derivative of replay ratio is calculated to capture the trend of comprehension ability change, and the sliding speed is standardized to output behavioral feature vector.

[0010] As a preferred embodiment of the short drama subtitle processing method based on big data processing described in this invention, the initial correction of subtitle data is obtained by using a lightweight TensorFlow Lite model. The specific steps are as follows: Tensor normalization is performed on the timestamp compression coefficient value and the behavior feature vector to obtain the standardized timestamp compression coefficient value and the standardized behavior feature vector. The original subtitle text is collected, and the standardized behavioral feature vector is input into the TensorFlow Lite lightweight model. The semantics are optimized through the LSTM layer, cultural hot words are replaced in the original subtitle text, and the subtitle duration interval is dynamically scaled based on the standardized timestamp compression coefficient value to obtain the initial corrected subtitle data.

[0011] As a preferred embodiment of the short drama subtitle processing method based on big data processing described in this invention, the specific steps for outputting the optimized subtitle data sequence are as follows: When the decision instruction is a compression mode decision instruction, the semantic core of the initial correction caption data is located through a multi-head attention mechanism, and the summary ratio is dynamically calculated based on the standardized timestamp compression coefficient value to generate compressed text; When the decision instruction is a correction mode decision instruction, dialect replacement is performed by matching the dialect dictionary through cultural adaptation. At the same time, the cognitive enhancer inserts explanatory sentences and adds visual markers when it detects professional terms, thus obtaining rich text caption data. The compressed text and rich text caption data are combined with the original duration and standardized timestamp compression coefficient value of the initial corrected captions, and a structured spatiotemporal encapsulation package is generated through a differentiated duration calculation algorithm. The structured spatiotemporal encapsulation package is discretized into independent subtitle entities according to the timeline, and the start and end timestamps are extracted from the start and end fields of the structured spatiotemporal encapsulation package respectively to obtain the start and end timestamps of each independent subtitle entity. By utilizing the start and end timestamps of each independent caption entity, a comprehension buffer time is added to excessively short texts, and visual marker symbols that cross independent caption entities are forcibly extended to obtain a fixed extension amount for visual marker symbols. The independent caption entities are then serialized into an ordered data array according to the JSON-LD specification to obtain an optimized caption data sequence.

[0012] As a preferred embodiment of the short drama subtitle processing method based on big data processing described in this invention, the specific steps for obtaining the federated optimization parameter set are as follows: Visual markers, cognitive load levels, and processing modes are extracted from the optimized caption data sequence and quantified into dialect weight coefficients, cognitive level values, and Boolean flags for processing modes, respectively, to generate a cognitively enhanced feature matrix. Using a cognitively enhanced feature matrix to drive a lightweight TensorFlow Lite model, the theoretical playback duration of each independent subtitle entity in a structured spatiotemporal encapsulation package is predicted in time series. The error loss between the predicted duration and the actual duration of the entity is calculated, and the backpropagation algorithm is used to generate a local gradient tensor. The local gradient tensor is injected with quantum noise and obfuscated. Homomorphic encryption is used to encapsulate it into a secure data packet. The visual marker symbol density of the optimized subtitle data sequence is extracted. The product of visual marker density and fixed extension of visual marker is used as the cognitive compensation factor. The gradient data in the security data package is decrypted, and the weighted average gradient is calculated by combining the knowledge entropy index. The parameters of the TensorFlow Lite lightweight model and the cognitive compensation factor are dynamically updated to form a federated optimization parameter set.

[0013] As a preferred embodiment of the short drama subtitle processing method based on big data processing described in this invention, the specific steps for outputting the subtitle file for direct playback are as follows: Based on the federated optimization parameter set, the temporal prediction weights, dialect conversion thresholds, and cognitive compensation factors are separated to activate the subtitle generation function; the dialect and standard language are automatically switched according to the dialect conversion threshold, the cognitive compensation factor is used to extend the display duration of professional terms, and visual markers are integrated to generate subtitle data packages; The subtitle data package is waveform-matched and aligned with the original audio of the short drama, distributed through the application programming interface, and after the subtitle data package is loaded on the playback channel, a subtitle file that can be played directly is output.

[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements any step of the short drama subtitle processing method based on big data processing as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the short drama subtitle processing method based on big data processing as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By constructing a dynamic attention threshold generation mechanism based on Gaussian mixture model and federated learning, the invention achieves accurate identification and personalized calibration of the user's cognitive load threshold, overcoming the problem of insufficient adaptability caused by the reliance on fixed thresholds in traditional methods, and significantly improving the matching accuracy between subtitle adjustment strategies and user attention states; by introducing an incremental federated learning update mechanism based on optimized subtitle data sequences, the semantic correction results are transformed into cognitively enhanced features and drive the secure aggregation and update of model parameters, which not only achieves continuous optimization and cross-device collaborative evolution of the model while protecting user privacy, but also enhances the system's long-term adaptability to complex viewing scenarios. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a short drama subtitle processing method based on big data processing.

[0019] Figure 2 A flowchart for generating the attention threshold.

[0020] Figure 3 A flowchart for generating a structured spatiotemporal encapsulation package.

[0021] Figure 4 A flowchart for forming a set of federal optimization parameters. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a short drama subtitle processing method based on big data processing, including the following steps: S1. Collect big data on user viewing behavior, clean and format it to generate a user behavior dataset, and calculate the initial attention index value.

[0026] S1.1. Collect big data on user viewing behavior, identify and remove abnormal high-speed scrolling caused by device mis-touch in the big data of user viewing behavior through a sliding filter with adaptive window width, and fill in the missing points of the big data of user viewing behavior through time series interpolation algorithm to obtain user behavior dataset.

[0027] Specifically, user pause frequency, replay ratio, and scrolling speed are collected to obtain big data on user viewing behavior: When collecting user pause frequency, the number of pause operations is counted according to a fixed duration window (e.g., 1 minute) to obtain the average user pause frequency; when collecting replay ratio, replay operation events are marked with a fixed sampling frequency (e.g., once per second) to obtain the average proportion of the number of times the event occurs to the total number of samples, forming the average replay ratio; when collecting scrolling speed, the instantaneous speed of the displacement distance and time interval of each scrolling operation is measured to obtain the average scrolling speed. The user pause frequency, average user pause frequency, replay ratio, average replay ratio, scrolling speed, and average scrolling speed are integrated to form the user viewing behavior big data. To analyze user viewing behavior big data, an adaptive window-width sliding filter is used. This filter dynamically adjusts the window width based on the sliding speed sequence to obtain a moving average and standard deviation, identifying high-speed sliding events with abnormal sliding speeds. These events are caused by accidental device touches. The data points corresponding to these abnormal high-speed sliding events are directly removed from the user viewing behavior big data. For the user viewing behavior big data with missing data after removing the anomalies, a time series interpolation algorithm is applied. This algorithm locates the missing points based on the timestamp sequence and uses adjacent valid data points for interpolation calculations to fill in all missing data points, forming a user behavior dataset.

[0028] S1.2. Load the dynamic weight parameters of the user behavior dataset through federated learning, and calculate the initial attention index value through a multi-index weighted calculation formula.

[0029] Specifically, based on three metrics in the user behavior dataset—the average pause frequency, the average replay ratio, and the average swipe speed—the initial attention index is calculated using a multi-metric weighted calculation formula.

[0030] It should be noted that the expression for calculating the initial attention index value is: ; in, This represents the initial attention index value. This indicates the total number of playback segments within the statistics window. Indicates the index of the playback segment. Indicates the first The pause frequency of each playback segment, Indicates pause frequency Dynamic weight parameters, This indicates the total number of times the playback operation occurred. Indicates the index of the playback operation. Indicates the first Playback ratio for each playback operation Indicates playback ratio Dynamic weight parameters, This indicates the total number of sliding operations. Indicates the index of the sliding operation. Indicates the first The speed of each swipe operation Indicates sliding speed The dynamic weight parameters.

[0031] S2. Based on the user behavior dataset and the initial attention index value, the cognitive load threshold is identified through a Gaussian mixture model, and dynamic calibration is performed by combining federated learning. The attention threshold is output and compared with the initial attention index value to obtain the decision instruction.

[0032] S2.1. Based on the average pause frequency, average replay ratio, and average swipe speed in the user behavior dataset, and combined with the initial attention index value, construct a standardized multidimensional feature matrix.

[0033] Specifically, the average pause frequency, average replay ratio, and average swipe speed recorded in the user behavior dataset are extracted as basic behavioral features; the initial attention index value is used as the cognitive state feature; the above four features are arranged in a fixed order into a four-dimensional vector (average pause frequency, average replay ratio, average swipe speed, and initial attention index value); z-score standardization is performed on each feature dimension (e.g., the pause frequency average dimension: the difference between the average pause frequency and the historical pause frequency average of the same dimension); all standardized feature vectors are arranged in rows to form a standardized multi-dimensional feature matrix.

[0034] S2.2. Input the standardized multidimensional feature matrix into the Gaussian mixture model for cluster analysis, identify the boundary points of high cognitive load clusters, generate the initial attention threshold, calculate the absolute difference between the initial attention index value of the terminal device and the initial attention threshold, and obtain the local attention bias.

[0035] Specifically, the standardized multidimensional feature matrix is ​​input into the Gaussian mixture model to perform cluster analysis: the probability density value of each sample point in each Gaussian component is obtained; the expectation step is iteratively executed through the expectation-maximization algorithm to obtain the posterior probability of the sample point belonging to each component; based on the posterior probability of each component, the mixing coefficient, mean vector, and covariance matrix of the Gaussian component are updated by maximization; after the iteration terminates, the Gaussian component with the highest mean of the feature dimension of the initial attention index value is taken as the high cognitive load cluster, the convex hull boundary of the feature space of the sample set of the high cognitive load cluster is obtained, and the boundary point with the largest Euclidean distance from the cluster center point in the convex hull boundary point set is selected as the key boundary point; the coordinate values ​​of the key boundary point on the feature axis of the initial attention index value of the standardized multidimensional feature matrix are read, and the coordinate values ​​on the feature axis of the initial attention index value are mapped to the initial attention threshold; the absolute difference between the initial attention index value and the initial attention threshold is taken as the local attention bias.

[0036] It should be noted that the Gaussian mixture model is trained using historical user behavior data as the training sample set. The Gaussian mixture model parameters are initialized by determining the number of Gaussian components, the weight coefficients of each Gaussian component, the mean vector, and the covariance matrix. The expectation-maximization algorithm iterative process is executed: for each training sample point, the posterior probability of belonging to the Kth Gaussian component is calculated. The posterior probability is calculated as the product of the current Gaussian component weight and the probability density value of the current Gaussian component relative to the sample point. The quotient of this product and the weighted sum of the probability densities of all components is then used as the posterior probability of the Kth Gaussian component. Based on the posterior probability, the parameters are updated, and the parameters of each Gaussian component are obtained. After each iteration, the log-likelihood function value (the total log probability of all samples) is calculated. If the change in the log-likelihood function value between two consecutive iterations is less than the convergence threshold (0.001 in the example), the training is terminated. The trained Gaussian mixture model is then obtained.

[0037] It should be noted that the convergence threshold is obtained by multiplying the logarithm of the sample size by the standard deviation of the data noise by a factor of 10.

[0038] S2.3. Aggregate local attention biases using a federated learning framework and dynamically calibrate the initial attention threshold using a gradient descent algorithm to generate the attention threshold.

[0039] Specifically, the cloud aggregation server of the federated learning framework receives a set of local attention bias values ​​uploaded by multiple terminal devices (each terminal device only uploads a single scalar value of the local attention bias); a sample-weighted average algorithm is used to calculate the global bias; the gradient descent algorithm is used to dynamically calibrate the initial attention threshold, the initial attention threshold is set as the optimization variable, the updated optimization variable is calculated using the gradient descent algorithm formula, and the updated optimization variable is output as the attention threshold.

[0040] It should be noted that the expression for the optimized variable to be updated is: ; in, This represents the updated optimization variable. Indicates the initial attention threshold. This represents the learning rate in gradient descent. This represents the global deviation.

[0041] S2.4. Intelligently compare the initial attention index value with the attention threshold. When the real-time attention index exceeds the attention threshold, a compression mode decision instruction is obtained. When the real-time attention index is less than the attention threshold, a correction mode decision instruction is obtained.

[0042] Specifically, the real-time attention index value is read and compared with the attention threshold value. If the real-time attention index value is significantly greater than the attention threshold value, the Boolean flag value of the compression mode decision instruction is output as true. If the real-time attention index value is less than or equal to the attention threshold value, the Boolean flag value of the correction mode decision instruction is output as false. The output decision instruction result must include the decision mode flag field (flag 1 for compression mode and flag 0 for correction mode).

[0043] Beneficial effects: By extracting the convex hull boundary points of high-dimensional behavioral features using Gaussian mixture models and mapping them to attention thresholds, and introducing a federated learning mechanism to dynamically calibrate attention, unsupervised clustering and federated optimization are coupled. The geometric features of boundary points solve the problem of quantifying the critical value of cognitive load, and cross-device gradient descent enables the attention threshold to have dynamic adaptability. Existing technologies judge cognitive state based on a fixed threshold set by a single behavioral feature (such as triggering an alarm only when the pause frequency exceeds a preset value), or use a centralized machine learning model to train a static attention classifier with full user data and generate decision instructions by comparing the real-time attention index with a fixed threshold through hard rules. In contrast, by mining the geometric boundaries of multi-dimensional behavioral features using Gaussian mixture models, the critical point of cognitive load is accurately quantified, and cross-device dynamic calibration is achieved by integrating federated learning. Compared with traditional methods that rely on a single feature or static model, this significantly improves the personalization, adaptability, and discrimination accuracy of attention thresholds, and effectively solves the problems of lagging critical state recognition and insufficient generalization ability.

[0044] S3. Perform big data analysis on the user behavior dataset to obtain timestamp compression coefficient values ​​and behavior feature vectors.

[0045] S3.1. Extract the average pause frequency, average replay ratio, and average swipe speed from the user behavior dataset, and calculate the weighted attention entropy index.

[0046] Specifically, the sum of the absolute values ​​of the three means—the average pause frequency, the average replay ratio, and the average sliding speed—is used to obtain the probability normalization of each mean: the average pause frequency probability is obtained by quotienting the average pause frequency and the sum of the absolute values ​​of the three means; the average replay ratio probability is obtained by quotienting the average replay ratio and the sum of the absolute values ​​of the three means; and the average sliding speed probability is obtained by quotienting the average sliding speed and the sum of the absolute values ​​of the three means. The weighted attention entropy index is then calculated based on the average sliding speed probability, the average replay ratio probability, and the average pause frequency probability using a weighted attention entropy index.

[0047] It should be noted that the expression for calculating the weighted attention entropy exponent is: ; in, This represents the weighted attention entropy index. Indicates pause frequency Normalized probability value Playback ratio Normalized probability value Sliding speed Normalized probability value Indicates the pause frequency. Indicates the playback ratio. This indicates the sliding speed.

[0048] S3.2. Use the Sigmoid function to transform the weighted attention entropy exponent and output the timestamp compression coefficient value.

[0049] Specifically, the weighted attention entropy index is input into the Sigmoid function to perform a non-linear mapping, resulting in a timestamp compression coefficient value.

[0050] It should be noted that the expression for calculating the timestamp compression factor is as follows: ; in, This represents the timestamp compression factor value. Represents the natural constant. This represents the steepness adjustment factor of the Sigmoid curve.

[0051] S3.3. By analyzing the standard deviation of the sliding window, the fluctuation amplitude of the pause frequency is analyzed, the time derivative of the replay ratio is calculated to capture the trend of changes in comprehension ability, and the sliding speed is standardized to output the behavioral feature vector.

[0052] Specifically, operations are performed based on the pause frequency, replay ratio, and sliding speed in the user behavior dataset: the standard deviation of the pause frequency sequence is obtained using a fixed time window to obtain the fluctuation amplitude value; the time derivative of the replay ratio time series is calculated using the backward difference method to output the change trend value; the mean and standard deviation parameters of the sliding speed benchmark for the current device type are called to perform Z-score standardization on the original sliding speed sequence to output the standardized sliding speed value; the fluctuation amplitude value, change trend value, and standardized sliding speed value at the latest time point are extracted and combined into a behavior feature vector.

[0053] S4. Perform initial calibration using the TensorFlow Lite lightweight model to obtain initial calibrated caption data.

[0054] S4.1. Perform tensor normalization on the timestamp compression coefficient value and the behavior feature vector to obtain the standardized timestamp compression coefficient value and the standardized behavior feature vector.

[0055] Specifically, the original timestamp compression coefficient values ​​are extracted, and Z-score standardization is performed based on the mean and standard deviation of the historical timestamp compression coefficient value dataset to obtain the standardized timestamp compression coefficient values. Independent Z-score standardization is then applied to the three dimensions of the behavioral feature vector: the pause frequency fluctuation amplitude dimension is obtained by the quotient of the difference between the pause frequency fluctuation amplitude and the historical mean pause frequency fluctuation amplitude, and the historical standard deviation of the pause frequency fluctuation amplitude; the playback ratio deviation dimension is obtained by the difference between the playback ratio time derivative and the historical mean playback ratio time derivative, and the playback ratio time derivative dimension is obtained by the quotient of the playback ratio deviation dimension and the historical standard deviation of the playback ratio time derivative; the dynamic deviation of the sliding speed dimension is obtained by the difference between the standardized sliding speed and the historical mean standardized sliding speed, and the standardized sliding speed dimension is obtained by the quotient of the dynamic deviation of the sliding speed and the historical standard deviation of the sliding speed. Finally, the standardized timestamp compression coefficient values ​​and the standardized behavioral feature vector are recombined and output.

[0056] S4.2. Collect the original subtitle text, input the standardized behavioral feature vector into the TensorFlow Lite lightweight model, perform semantic optimization through the LSTM layer, replace cultural hot words in the original subtitle text, and dynamically scale the subtitle duration interval based on the standardized timestamp compression coefficient value to obtain the initial corrected subtitle data.

[0057] Specifically, the process involves collecting raw subtitle text data, inputting standardized behavioral feature vectors into a TensorFlow Lite lightweight model, and processing them through an LSTM layer. The LSTM layer outputs a cultural adaptation parameter index value, which is used to match target words from a local cultural hot word library to replace corresponding words in the raw subtitle text. Simultaneously, standardized timestamp compression coefficient values ​​are read, and the scaled subtitle duration is obtained by multiplying the original subtitle sentence's baseline duration by the standardized timestamp compression coefficient value. The duration distribution corresponding to the drop in user comprehension accuracy to a critical value under different subtitle durations is statistically analyzed, and the 5th percentile is used as the baseline value for the minimum guarantee threshold. Device response latency compensation is added to generate a minimum guarantee threshold. When the scaled subtitle duration is lower than the minimum guarantee threshold (1.0 second in the example), the minimum guarantee mechanism is activated to force an extension to the minimum guarantee threshold. Initial corrected subtitle data is then generated.

[0058] It should be noted that, to train the TensorFlow Lite lightweight model, a training dataset of 100,000 labeled samples was constructed. The input data consisted of standardized behavioral feature vectors, and the label data consisted of manually labeled cultural adaptation parameter index values. A neural network structure containing LSTM layers and fully connected layers was constructed, and the index prediction probability distribution was generated through the Softmax output layer. The cross-entropy loss function was used to measure the prediction error. The Adam optimizer was configured (with an example learning rate of 0.001) for parameter updates, and the training terminated after 200 iterations. The trained TensorFlow Lite lightweight model was then obtained.

[0059] S5. Use decision instructions to perform semantic correction on the initial corrected subtitle data and output an optimized subtitle data sequence.

[0060] S5.1. When the decision instruction is a compression mode decision instruction, the semantic core of the initial correction caption data is located through a multi-head attention mechanism, and the summary ratio is dynamically calculated based on the standardized timestamp compression coefficient value to generate compressed text.

[0061] Specifically, when the decision instruction is a compression mode decision instruction: the attention weight score of each word is obtained through a multi-head attention mechanism, and the continuous word sequence with the highest attention weight score is located as the semantic core; the standardized timestamp compression coefficient value is extracted, and the summary ratio (the difference between the standardized timestamp compression coefficient value and 1) is calculated through the Sigmoid function mapping. The target length is determined based on the semantic core and the summary ratio, and the keyword combination is extracted in descending order of attention weight score through a pointer generation network to generate compressed text.

[0062] It should be noted that the training process for the pointer generation network is as follows: A dataset of historical short drama subtitle texts is collected to construct training sample pairs. Each sample pair contains the complete original subtitle text and a manually annotated compressed text summary. The text is segmented and a vocabulary is constructed. The original subtitle text sequence is input into the encoder LSTM layer to generate a hidden state vector. The decoder LSTM layer obtains the weight distribution of the encoder's hidden state through an attention mechanism. Combined with the pointer network mechanism, it dynamically selects whether to copy words from the source text or generate words from the vocabulary. Maximum likelihood estimation is used to optimize the objective function, which is a negative log-likelihood loss function plus a coverage loss term. The pointer generation network parameters are updated through backpropagation until the validation set ROUGE score converges, and the trained pointer generation network is output.

[0063] S5.2. When the decision instruction is a correction mode decision instruction, dialect replacement is performed by matching the dialect dictionary through cultural adaptation. At the same time, the cognitive enhancer inserts explanatory short sentences and adds visual markers when it detects professional terms, thus obtaining rich text caption data.

[0064] Specifically, when the decision instruction is a correction mode decision instruction: read the initial correction caption data text field, call the device's GPS positioning data to match the regional dialect lexicon, traverse the text to replace the target words with dialect equivalents; simultaneously start the cognitive enhancer to scan the text, when a professional term is detected, insert an explanatory sentence after the professional term, and add a visual marker symbol at the beginning of the professional term; output rich text caption data containing dialect replacement text, inserted explanatory sentences, and visual marker symbols.

[0065] S5.3. Combine the compressed text and rich text subtitle data with the original duration and standardized timestamp compression coefficient value of the initial corrected subtitles, and generate a structured spatiotemporal encapsulation package through a differentiated duration calculation algorithm.

[0066] Specifically, when the decision instruction is a compression mode decision instruction, the compressed text string is read; when the decision instruction is a correction mode decision instruction, the rich text subtitle data string is read. Differential duration calculation is performed by combining the original subtitle single-sentence baseline duration field and the standardized timestamp compression coefficient value field of the initial correction subtitle data: the product of the original subtitle single-sentence baseline duration and the standardized timestamp compression coefficient value yields the calculated compressed duration in the compression mode scenario; the calculated corrected duration in the correction mode scenario is the sum of the original subtitle single-sentence baseline duration, the fixed extension of the visual marker symbol, and the incremental duration of the explanation sentence. The text string and the calculated duration value are encapsulated into a structured spatiotemporal encapsulation data entity, containing a text content field, a start timestamp field, an end timestamp field, a processing mode marker field, and a visual marker symbol position index field (only the symbol offset is recorded in the correction mode).

[0067] S5.4. Discretize the structured spatiotemporal encapsulation package into independent subtitle entities according to the timeline, and extract the start timestamp and end timestamp from the start field and end field of the structured spatiotemporal encapsulation package respectively to obtain the start and end timestamps of each independent subtitle entity.

[0068] Specifically, the structured spatiotemporal encapsulation sequence is traversed in chronological order, and discretization is performed on each structured spatiotemporal encapsulation: the start timestamp value is directly read from the start field of the structured spatiotemporal encapsulation, and the end timestamp value is directly read from the end field of the structured spatiotemporal encapsulation; the start timestamp value and the end timestamp value are assigned to the start timestamp field and the end timestamp field of the independent subtitle entity, respectively; at the same time, the text content field, processing mode mark field, and visual mark symbol position index field of the structured spatiotemporal encapsulation are copied to the corresponding independent subtitle entity field; the start and end timestamps of each independent subtitle entity are generated.

[0069] S5.5. Using the start and end timestamps of each independent caption entity, the understanding buffer time is increased for excessively short texts, and the visual marker symbols that cross independent caption entities are forcibly extended to obtain a fixed extension amount for the visual marker symbols. The independent caption entities are serialized into an ordered data array according to the JSON-LD specification to obtain an optimized caption data sequence.

[0070] Specifically, the operation involves iterating through the sequence of independent subtitle entities and performing the following steps: When the number of characters in the text content field of an independent subtitle entity is lower than the character threshold (calculated based on average reading speed and the minimum human eye fixation unit); simultaneously, the visual marker position index field is checked. If the visual marker position index field of the current independent subtitle entity exists and the visual marker position index field of the next independent subtitle entity is consecutive (e.g., current end index 5, next start index 6), the termination timestamp of the current independent subtitle entity is forcibly extended; the adjusted independent subtitle entities are serialized into an ordered data array according to the JSON-LD specification. Each entity is encapsulated as a JSON object containing a start timestamp field, an end timestamp field, a text content field, a processing mode marker field, and a visual marker position index field, forming an optimized subtitle data sequence.

[0071] It should be noted that the processing time per word is calculated based on the average reading speed, combined with the minimum fixation unit of the human eye.

[0072] S6. Perform incremental federated learning on the optimized subtitle data sequence to obtain the federated optimization parameter set.

[0073] S6.1. Extract visual marker symbols, cognitive load levels, and processing modes from the optimized caption data sequence, and quantify them into dialect weight coefficients, cognitive level values, and Boolean flags for processing modes, respectively, to generate a cognitively enhanced feature matrix.

[0074] Specifically, the process iterates through each independent subtitle in the subtitle data sequence: counts the occurrence frequency of the marker symbol in the visual marker symbol position index field of the independent subtitle entity, and obtains the dialect weight coefficient by multiplying the occurrence frequency of the marker symbol by the dialect heat factor (0.15 in the example); reads the difference between the end timestamp and the start timestamp of the independent subtitle entity as the actual duration, classifies the cognitive load level according to the duration range, and outputs the cognitive level value; extracts the processing mode marker field value and directly maps it to the processing mode Boolean flag (compression mode = 0, correction mode = 1); combines the dialect weight coefficient value, the cognitive level value, and the processing mode Boolean flag in a fixed order into a three-dimensional row vector, stacks the row vectors of all independent subtitle entities, and forms a cognitive enhancement feature matrix.

[0075] S6.2. Using a cognitively enhanced feature matrix to drive a lightweight TensorFlow Lite model to perform time-series prediction of the theoretical playback duration of each independent subtitle entity in the structured spatiotemporal encapsulation package, calculate the error loss between the predicted duration and the actual duration of the entity, and use the backpropagation algorithm to generate a local gradient tensor.

[0076] Specifically, the cognitively enhanced feature matrix is ​​input into the TensorFlow Lite lightweight model for forward propagation: the fully connected layer of the TensorFlow Lite lightweight model calculates the theoretical playback duration prediction value of the independent subtitle entity based on the dialect weight coefficient component, the cognitive level numerical component, and the processing mode Boolean flag component; the difference between the end timestamp and the start timestamp of the corresponding independent subtitle entity in the optimized subtitle data sequence is read as the actual duration of the entity; the mean squared error loss function value is obtained by squared the difference between the theoretical playback duration prediction value and the actual duration of the entity; the backpropagation algorithm is executed to obtain the partial derivatives of the trainable parameters of the TensorFlow Lite lightweight model and generate the local gradient tensor.

[0077] S6.3. The local gradient tensor is injected with quantum noise and obfuscated. Homomorphic encryption is used to encapsulate it into a secure data packet. The visual marker symbol density of the optimized subtitle data sequence is extracted.

[0078] Specifically, a quantum noise injection operation is performed on the local gradient tensor: a random matrix with the same dimension as the local gradient tensor (the element values ​​follow a Gaussian distribution with a mean of 0 and a standard deviation equal to the quantum noise intensity parameter, with an example noise intensity of 0.05) is generated using the np.random.normal function of the NumPy library. This random matrix is ​​then superimposed on the element values ​​of the local gradient tensor to achieve obfuscation. The obfuscated local gradient tensor is then encrypted using the Paillier homomorphic encryption algorithm and encapsulated into a homomorphically encrypted secure data packet. Simultaneously, each independent subtitle entity in the optimized subtitle data sequence is traversed, and the number of occurrences of the marker symbols in the visual marker symbol position index field is counted. The visual marker symbol density is obtained by dividing the total number of symbol occurrences by the total number of entities in the optimized subtitle data sequence.

[0079] S6.4. The product of the visual marker density and the fixed extension of the visual marker is used as the cognitive compensation factor.

[0080] Specifically, the visual marker density value and the visual marker fixed extension value are multiplied: the visual marker density scalar value of the optimized caption data sequence is read, and based on the visual marker fixed extension (0.3 seconds in the example), the product of the visual marker density and the visual marker fixed extension is used as the cognitive compensation factor.

[0081] S6.5. Decrypt the gradient data in the security data packet, calculate the weighted average gradient by combining the knowledge entropy index, dynamically update the TensorFlow Lite lightweight model parameters and cognitive compensation factors, and form a federated optimization parameter set.

[0082] Specifically, the encrypted gradient data in the security data packet is decrypted to obtain the original gradient value. The knowledge entropy index of each terminal device is extracted as a weight coefficient, and a weighted average gradient is calculated. Based on the weighted average gradient, the gradient descent algorithm is executed to update the TensorFlow Lite lightweight model parameters, and the cognitive compensation factor is adjusted synchronously. The updated TensorFlowLite lightweight model parameters and the cognitive compensation factor are integrated to form a federated optimization parameter set.

[0083] It should be noted that the expression for calculating the weighted average gradient is: ; in, This represents the weighted average gradient. Indicates the first Knowledge entropy index of terminal devices. Indicates the first The local gradient vector of the device. Represents the device index variable. This represents the total number of terminal devices participating in federated learning.

[0084] Beneficial Effects: This paper proposes a feature matrix that includes dialect weights, cognitive hierarchy, and processing modes to drive temporal prediction in the TensorFlow Lite lightweight model. Quantum noise obfuscation and homomorphic encryption are introduced to ensure gradient security. A visually labeled density-driven cognitive compensation factor is proposed, combined with a knowledge entropy-weighted gradient aggregation mechanism, to achieve collaborative dynamic optimization of TensorFlow Lite lightweight model parameters and compensation strategies. Existing technologies typically set subtitle playback duration uniformly based on fixed rules or centralized machine learning models, using single behavioral features (such as viewing time or click-through rate) to trigger simple feedback mechanisms. This lacks dynamic perception and personalized modeling of user cognitive load, poses privacy risks, and makes it difficult to achieve fine-grained, adaptive subtitle optimization. In contrast, this innovative solution combines multi-dimensional cognitive feature modeling with lightweight temporal prediction to achieve refined dynamic adaptation of subtitle playback duration. The introduction of quantum noise and homomorphic encryption ensures data security during federated learning, enhancing privacy protection capabilities. The paper also introduces a novel visually labeled density-driven cognitive compensation factor, combined with a knowledge entropy-weighted aggregation mechanism, enabling TensorFlow Lite to achieve collaborative dynamic optimization of model parameters and compensation strategies. The Lite lightweight model features cross-device adaptive optimization capabilities, significantly improving the personalization, security, and cognitive friendliness of subtitle presentation.

[0085] S7. Generate standardized subtitle files based on the federated optimization parameter set, and distribute them to the short drama playback channel through the application programming interface, outputting subtitle files for direct playback.

[0086] S7.1. Based on the federated optimization parameter set, separate the temporal prediction weights, dialect conversion thresholds, and cognitive compensation factors to activate the subtitle generation function; automatically switch between dialects and standard language according to the dialect conversion threshold, use the cognitive compensation factor to extend the display duration of professional terms, and integrate visual marker symbols to generate subtitle data packages.

[0087] Specifically, based on the federated optimization parameter set, the temporal prediction weight parameter, dialect conversion threshold parameter, and cognitive compensation factor parameter are separated; the temporal prediction weight parameter is called to activate the temporal prediction capability of the subtitle generation function; the original subtitle text data of the short drama is read, and when the dialect heat value of the device's GPS positioning matching area exceeds the dialect conversion threshold parameter, the dialect dictionary is called to perform dialect replacement; otherwise, the standard language text is retained; when detecting professional terms, explanatory short sentences are inserted, and the cognitive compensation factor parameter is applied to extend the display duration of professional terms; visual markers are added at the beginning of the explanatory short sentences; the processed text content field, visual marker position index field, and adjusted duration field are encapsulated into a subtitle data package JSON structure to obtain the subtitle data package.

[0088] It should be noted that the dialect usage frequency data of multiple terminal devices is aggregated through the federated learning framework, and the dialect conversion threshold parameter is dynamically generated by combining knowledge entropy weighted calculation.

[0089] S7.2. Match and align the subtitle data package with the original audio of the short drama using waveform matching, distribute it through the application programming interface, and output a subtitle file that can be played directly after loading the subtitle data package on the playback channel.

[0090] Specifically, based on the adjusted duration and text content fields in the subtitle data packet, a dynamic time warping algorithm is invoked to match the original audio waveform of the short drama: Mel frequency cepstral coefficient feature extraction is performed on the original audio to generate an acoustic feature sequence, and audio segments are simultaneously segmented according to the start and end timestamp fields of the subtitle data packet; the minimum dynamic time warping distance between the Mel frequency cepstral coefficient sequence of the audio segment and the forced alignment path of the subtitle text is obtained, and the start timestamp field value of the subtitle data packet is finely adjusted to align the phoneme boundaries (example: offset ±0.2 seconds); the timestamp-calibrated subtitle data packet is transmitted to the short drama playback channel server through the application programming interface; the playback channel loading module parses the subtitle data packet to generate a WebVTT format file (containing calibrated timestamps, text content, and visual marker position metadata), and outputs a subtitle file that can be directly embedded in the video player.

[0091] This embodiment also provides a computer device applicable to the short drama subtitle processing method based on big data processing, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the short drama subtitle processing method based on big data processing as proposed in the above embodiment.

[0092] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0093] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the short drama subtitle processing method based on big data processing as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0094] In summary, this invention achieves accurate identification and personalized calibration of the user's cognitive load threshold by constructing a dynamic attention threshold generation mechanism based on Gaussian mixture model and federated learning. This overcomes the lack of adaptability caused by the reliance on fixed thresholds in traditional methods and significantly improves the matching accuracy between subtitle adjustment strategies and user attention states. Furthermore, by introducing an incremental federated learning update mechanism based on optimized subtitle data sequences, the semantic correction results are transformed into cognitively enhanced features and drive the secure aggregation and update of model parameters. This not only enables continuous optimization and cross-device collaborative evolution of the model while protecting user privacy, but also enhances the system's long-term adaptability to complex viewing scenarios.

[0095] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for processing subtitles for short dramas based on big data processing, characterized in that: include, Collect big data on user viewing behavior, clean and format it to generate a user behavior dataset, and calculate the initial attention index value; Based on user behavior datasets and initial attention index values, a Gaussian mixture model is used to identify the cognitive load threshold. Federated learning is then used for dynamic calibration, and an attention threshold is output. This threshold is then compared with the initial attention index value to obtain decision instructions. Big data analysis was performed on the user behavior dataset to obtain timestamp compression coefficient values ​​and behavior feature vectors. Initial correction was then performed using a TensorFlow Lite lightweight model to obtain initial corrected caption data. The initial corrected caption data is semantically corrected using decision instructions, and an optimized caption data sequence is output. Incremental federated learning is applied to the optimized subtitle data sequence to obtain a set of federated optimization parameters. Standardized subtitle files are generated based on the federated optimization parameter set and distributed to short drama playback channels via application programming interface, outputting subtitle files for direct playback.

2. The short drama subtitle processing method based on big data processing as described in claim 1, characterized in that: The specific steps for calculating the initial attention index value are as follows: Collect big data on user viewing behavior, identify and remove abnormal high-speed scrolling caused by accidental device touches in the big data of user viewing behavior through an adaptive window width sliding filter, and fill in the missing points of the big data of user viewing behavior through time series interpolation algorithm to obtain user behavior dataset; The dynamic weight parameters of the user behavior dataset are loaded through federated learning, and the initial attention index value is calculated using a multi-index weighted calculation formula.

3. The short drama subtitle processing method based on big data processing as described in claim 1, characterized in that: The specific steps for obtaining the decision instruction are as follows: Based on the average pause frequency, average replay ratio, and average swipe speed in the user behavior dataset, and combined with the initial attention index value, a standardized multidimensional feature matrix is ​​constructed. The standardized multidimensional feature matrix is ​​input into the Gaussian mixture model for cluster analysis to identify the boundary points of high cognitive load clusters, generate an initial attention threshold, and calculate the absolute difference between the attention index value of the terminal device and the initial attention threshold to obtain the local attention bias. The local attention bias is aggregated through a federated learning framework, and the initial attention threshold is dynamically calibrated using a gradient descent algorithm to generate the attention threshold. The initial attention index value is intelligently compared with the attention threshold. When the real-time attention index exceeds the attention threshold, a compression mode decision instruction is obtained; when the real-time attention index is less than the attention threshold, a correction mode decision instruction is obtained.

4. The short drama subtitle processing method based on big data processing as described in claim 1, characterized in that: The specific steps for obtaining the timestamp compression coefficient value and the behavioral feature vector are as follows: Extract the average pause frequency, average replay ratio, and average swipe speed from the user behavior dataset, and calculate the weighted attention entropy index; The weighted attention entropy exponent is transformed using the Sigmoid function, and the timestamp compression coefficient value is output. The fluctuation amplitude of pause frequency is analyzed by sliding window standard deviation analysis, the time derivative of replay ratio is calculated to capture the trend of comprehension ability change, and the sliding speed is standardized to output behavioral feature vector.

5. The short drama subtitle processing method based on big data processing as described in claim 1, characterized in that: The specific steps for obtaining the initial corrected subtitle data are as follows. Tensor normalization is performed on the timestamp compression coefficient value and the behavior feature vector to obtain the standardized timestamp compression coefficient value and the standardized behavior feature vector. The original subtitle text is collected, and the standardized behavioral feature vector is input into the TensorFlow Lite lightweight model. The semantics are optimized through the LSTM layer, cultural hot words are replaced in the original subtitle text, and the subtitle duration interval is dynamically scaled based on the standardized timestamp compression coefficient value to obtain the initial corrected subtitle data.

6. The short drama subtitle processing method based on big data processing as described in claim 1, characterized in that: The specific steps for optimizing the output subtitle data sequence are as follows. When the decision instruction is a compression mode decision instruction, the semantic core of the initial correction caption data is located through a multi-head attention mechanism, and the summary ratio is dynamically calculated based on the standardized timestamp compression coefficient value to generate compressed text; When the decision instruction is a correction mode decision instruction, dialect replacement is performed by matching the dialect dictionary through cultural adaptation. At the same time, the cognitive enhancer inserts explanatory short sentences and adds visual markers when it detects professional terms, thus obtaining rich text caption data. The compressed text and rich text caption data are combined with the original duration and standardized timestamp compression coefficient value of the initial corrected captions, and a structured spatiotemporal encapsulation package is generated through a differentiated duration calculation algorithm. The structured spatiotemporal encapsulation package is discretized into independent subtitle entities according to the timeline, and the start and end timestamps are extracted from the start and end fields of the structured spatiotemporal encapsulation package respectively to obtain the start and end timestamps of each independent subtitle entity. By utilizing the start and end timestamps of each independent caption entity, a comprehension buffer time is added to excessively short texts, and visual marker symbols that cross independent caption entities are forcibly extended to obtain a fixed extension amount for visual marker symbols. The independent caption entities are then serialized into an ordered data array according to the JSON-LD specification to obtain an optimized caption data sequence.

7. The short drama subtitle processing method based on big data processing as described in claim 1, characterized in that: The specific steps for obtaining the federated optimization parameter set are as follows: Visual markers, cognitive load levels, and processing modes are extracted from the optimized caption data sequence and quantified into dialect weight coefficients, cognitive level values, and Boolean flags for processing modes, respectively, to generate a cognitively enhanced feature matrix. Using a cognitively enhanced feature matrix to drive a lightweight TensorFlow Lite model, the theoretical playback duration of each independent subtitle entity in a structured spatiotemporal encapsulation package is predicted in time series. The error loss between the predicted duration and the actual duration of the entity is calculated, and the backpropagation algorithm is used to generate a local gradient tensor. The local gradient tensor is injected with quantum noise and obfuscated. Homomorphic encryption is used to encapsulate it into a secure data packet. The visual marker symbol density of the optimized subtitle data sequence is extracted. The product of visual marker density and fixed extension of visual marker is used as the cognitive compensation factor. The gradient data in the security data package is decrypted, and the weighted average gradient is calculated by combining the knowledge entropy index. The parameters of the TensorFlow Lite lightweight model and the cognitive compensation factor are dynamically updated to form a federated optimization parameter set.

8. The short drama subtitle processing method based on big data processing as described in claim 1, characterized in that: The output directly plays the subtitle file. The specific steps are as follows: Based on the federated optimization parameter set, the temporal prediction weights, dialect conversion thresholds, and cognitive compensation factors are separated to activate the subtitle generation function; the dialect and standard language are automatically switched according to the dialect conversion threshold, the cognitive compensation factor is used to extend the display duration of professional terms, and visual markers are integrated to generate subtitle data packages; The subtitle data package is waveform-matched and aligned with the original audio of the short drama, distributed through the application programming interface, and after the subtitle data package is loaded on the playback channel, a subtitle file that can be played directly is output.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the short drama subtitle processing method based on big data processing as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the short drama subtitle processing method based on big data processing as described in any one of claims 1 to 8.

Images