A method and system for generating content-to-storyboard description with automatic iterative optimization

Abstract

The present invention provides an automated iterative optimization content-to-storyboard description generation method and system, which relates to the technical field of natural language processing. The method comprises: obtaining a multi-level summary text of a preset text; generating an initial storyboard description text based on the summaries at each level and preset chapters through a storyboard description generation model; obtaining a content quality score of the initial storyboard description text, as well as problem description information, and performing correction training on the storyboard description generation model to obtain a corrected storyboard description generation model; iteratively executing the correction training step until the content quality score of the corrected storyboard description text exceeds the preset quality score, and then outputting it. According to the present invention, the preset chapters can be processed based on the multi-level summary to obtain the storyboard description text of the preset chapters, so that the storyboard description generation model can refer to the contextual content of the full text, thereby reducing the probability of the storyboard description text having problems such as generalization, inconsistency between the preceding and following contexts, misunderstanding, and poor coherence.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and in particular to an automated iteratively optimized content-to-storyboard description generation method and system. Background Art

[0002] In the related art, CN117235523A discloses a method, system, device and storage medium for improving a language generation model. When executing the method, a training text is obtained; the training text is input into a language generation model to determine an output text; the output text is input into an evaluation model to determine evaluation information; the language generation model is trained based on the output text and the evaluation information until the model converges and meets the preset performance requirements. On the basis of traditional standard answer supervised training, a preference evaluation model is introduced to evaluate the text generated by the language generation model. The final model training is iteratively completed based on the output text and the evaluation information of the text. The language generation model is enabled to achieve self-iterative training, encouraging the model to generate content from a variety of perspectives, not just relying on the input training samples, so that the model can have better generation quality and generalization.

[0003] CN112989764A discloses a text generation method, device, equipment and storage medium. The method comprises the following steps: performing a text sequence generation operation on an initial corpus input by a user to obtain a first text sequence, and then using the first text sequence as the initial corpus. The method repeats this step until the number of times the text sequence generation operation is performed reaches a preset number, and then stops this process to obtain multiple first text sequences. The text sequence generation operation comprises a text revision operation and a sampling operation, and the text revision operation comprises at least one of replacement, insertion and deletion. Then, in descending order of matching degree, a preset number of second text sequences are selected and output from the multiple first text sequences. A large number of first text sequences are generated by rapid iteration using a language model, and then second text sequences with higher text quality are screened from the first text sequences by sampling, thereby improving the quantity and quality of text sequences.

[0004] Therefore, in the relevant technology, although natural language text can be processed and text quality can be improved through artificial intelligence-based neural network models, current neural network models are usually limited by the input limitations of the model and cannot process and understand long texts (for example, novels, long scripts, etc.) at one time. As a result, when the model generates content, it is unable to refer to the context of the full text and cannot understand the semantics of the full text, causing the generated content to easily have problems such as generalizations, inconsistencies, misunderstandings, and poor coherence.

[0005] The information disclosed in the background technology section of this application is only intended to deepen the understanding of the general background technology of this application, and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to those skilled in the art. Summary of the invention

[0006] The present invention provides an automated iteratively optimized content-to-storyboard description generation method and system, which can solve the technical problem that when a model generates content, it is unable to refer to the context of the full text and cannot understand the semantics of the full text.

[0007] According to a first aspect of the present invention, there is provided a method for generating content transition storyboard descriptions by automated iterative optimization, comprising:

[0008] Processing a preset text by a natural language processing model to obtain a first summary text of the preset text, a second summary text of each chapter of the preset text, and a third summary text of each paragraph of the chapter;

[0009] The first summary text, the second summary text, the third summary text, and a preset chapter in the preset text are processed by a pre-trained storyboard description generation model to obtain an initial storyboard description text corresponding to the preset chapter;

[0010] According to the preset chapter, the first summary text, the second summary text and the third summary text, the content quality of the initial storyboard description text is evaluated to obtain a content quality score of the initial storyboard description text;

[0011] Reviewing the initial storyboard description text according to the content quality score to obtain problem description information of the initial storyboard description text;

[0012] According to the problem description information of the initial storyboard description text, the storyboard description generation model is subjected to a first correction training to obtain a first corrected storyboard description generation model;

[0013] Obtain the F-th revised storyboard description text through the F-th revised storyboard description generation model, and obtain the content quality score of the F-th revised storyboard description text, where F≥1, and F is a positive integer;

[0014] If the content quality score of the storyboard description text after the F-th revision is higher than or equal to the preset quality score, the storyboard description text after the F-th revision is used as the storyboard description output content;

[0015] Otherwise, obtaining the problem description information of the storyboard description text after the Fth revision according to the content quality score of the storyboard description text after the Fth revision, and performing the F+1th revision training on the storyboard description generation model after the Fth revision to obtain the F+1th revised storyboard description generation model and obtain the F+1th revised storyboard description text;

[0016] In F+1= , the storyboard description text after the F+1th correction is used as the storyboard description output content, where, is the preset number of iterations.

[0017] According to a second aspect of the present invention, there is provided an automated iteratively optimized content transition storyboard description generation system, comprising:

[0018] A summary text module, used to process a preset text through a natural language processing model to obtain a first summary text of the preset text, a second summary text of each chapter of the preset text, and a third summary text of each paragraph of the chapter;

[0019] An initial storyboard description text module is used to process the first summary text, the second summary text, the third summary text and the preset chapters in the preset text through a pre-trained storyboard description generation model to obtain an initial storyboard description text corresponding to the preset chapters;

[0020] A content quality scoring module, configured to evaluate the content quality of the initial storyboard description text according to the preset chapter, the first summary text, the second summary text and the third summary text, and obtain a content quality score of the initial storyboard description text;

[0021] A problem description information module is used to review the initial storyboard description text according to the content quality score to obtain problem description information of the initial storyboard description text;

[0022] A correction training module, used for performing a first correction training on the storyboard description generation model according to the problem description information of the initial storyboard description text, so as to obtain a first corrected storyboard description generation model;

[0023] A scoring module, used to obtain the F-th revised storyboard description text through the F-th revised storyboard description generation model, and obtain a content quality score of the F-th revised storyboard description text, where F≥1, and F is a positive integer;

[0024] An output module, configured to use the storyboard description text revised for the Fth time as storyboard description output content if the content quality score of the storyboard description text revised for the Fth time is higher than or equal to a preset quality score;

[0025] an iterative module, for otherwise obtaining problem description information of the storyboard description text after the Fth revision according to the content quality score of the storyboard description text after the Fth revision, and performing the F+1th revision training on the storyboard description generation model after the Fth revision to obtain the storyboard description generation model after the F+1th revision, and obtaining the storyboard description text after the F+1th revision;

[0026] Iteration completion module, used to complete the F+1= , the storyboard description text after the F+1th correction is used as the storyboard description output content, where, is the preset number of iterations.

[0027] Technical effect: According to the present invention, a first summary text of a preset text, a second summary text of each chapter, and a third summary text of each paragraph can be obtained respectively, and then the preset chapter can be processed based on the above summary texts to obtain an initial storyboard description text of the preset chapter. Based on the summary texts of the full text, chapters, and paragraphs, a storyboard description generation model can refer to the contextual content of the full text, and then the preset chapter is converted into an initial storyboard description text, thereby reducing the probability that the initial storyboard description text will have problems such as generalizing from a particular part to the whole, inconsistency between the preceding and following contexts, misunderstandings, and poor coherence. The initial storyboard description text can be evaluated for quality, problems existing in the initial storyboard description text can be found, and corrections can be made based on the problems, and the process can be iteratively performed until the quality of the corrected storyboard description text meets the requirements, thereby obtaining a high-quality storyboard description text, thereby further reducing the probability that the initial storyboard description text will have problems such as generalizing from a particular part to the whole, inconsistency between the preceding and following contexts, misunderstandings, and poor coherence. When training a natural language processing model, the loss function of the natural language processing model can be set by the semantic similarity and distortion between the first pending summary text and the combined text of the full text summary, and the semantic similarity and distortion between the second pending summary text and the combined text of the chapter summary, and the natural language processing model can be iteratively updated and trained to improve the accuracy of the natural language processing model, thereby obtaining accurate summary texts at multiple levels, so that when subsequently generating storyboard descriptions, accurate summary texts at multiple levels can be used, so that the storyboard description generation model can grasp richer context information as much as possible when it cannot receive the input of long texts, so that the storyboard description is closer to the semantics of the original text, and the probability of generalizing, inconsistency, poor coherence, etc. can be reduced. When determining the accuracy score of the initial storyboard description text, the accuracy of the description of the characters, places and actions in the initial storyboard description text can be solved by comparing the preset chapter summary text with the initial storyboard description text, so as to obtain the accuracy score of the initial storyboard description text, which can objectively and accurately describe the accuracy of the description of the important plots in the initial storyboard description text. When determining the original meaning retention score, the original meaning retention score can be determined by factors such as the semantic similarity between the initial storyboard description text and the original text of the preset chapter, and whether important plots such as characters, places, actions in the initial storyboard description text appear in the original text of the preset chapter, and whether they appear in the second summary text of other chapters. This can accurately and objectively indicate whether the initial storyboard description text can maintain the original meaning of the original text of the preset chapter, and whether there is a defect of important plot spoilers.When determining the semantic error detection score, the semantic accuracy of each sentence in the initial storyboard description text can be determined by comparing the sentences in the initial storyboard description text with the sentences in the original text of the preset chapters sentence by sentence, and whether the important plots such as characters, places, actions in the initial storyboard description text appear in the full text to determine whether there are erroneous plots in the initial storyboard description text generated by the storyboard description generation model based on incorrect understanding, thereby improving the accuracy and objectivity of the semantic error detection score. When reviewing the problems existing in the initial storyboard description text, the problems existing in the initial storyboard description text can be reviewed by the semantic and logical similarity between the initial storyboard description text and the original text, the completeness of the expression of the original text, and the similarity with the original text in emotional expression, so as to obtain accurate problem description information, and provide a data basis for the correction training of the storyboard description generation model and the correction of the initial storyboard description text. When training the storyboard description generation model, the loss function for the first correction training can be determined based on aspects such as the overall distortion of the original text by the initial storyboard description text, the semantic error between the initial storyboard description text and the original text, and the overall difference in emotional expression between the initial storyboard description text and the preset chapters. The weights of the above aspects can be set through the problem description information, so as to focus on the aspects with serious problems and improve the pertinence and efficiency of the training.

[0028] It should be understood that the above general description and the following detailed description are exemplary and explanatory only and do not limit the present invention. Other features and aspects of the present invention will become more apparent from the following detailed description of exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other embodiments can be obtained based on these drawings without creative work.

[0030] Figure 1 A schematic diagram exemplarily shows a flow chart of a method for generating content-to-storyboard descriptions through automated iterative optimization according to an embodiment of the present invention;

[0031] Figure 2 A block diagram of a content transition storyboard description generation system for automated iterative optimization according to an embodiment of the present invention is exemplarily shown. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0034] Figure 1 A flow chart of a method for generating content transition storyboard descriptions according to an embodiment of the present invention is exemplarily shown, wherein the method comprises:

[0035] Step S101, processing a preset text through a natural language processing model to obtain a first summary text of the preset text, a second summary text of each chapter of the preset text, and a third summary text of each paragraph of the chapter;

[0036] Step S102, processing the first summary text, the second summary text, the third summary text and the preset chapters in the preset text by using a pre-trained storyboard description generation model to obtain an initial storyboard description text corresponding to the preset chapters;

[0037] Step S103, evaluating the content quality of the initial storyboard description text according to the preset chapter, the first summary text, the second summary text and the third summary text, to obtain a content quality score of the initial storyboard description text;

[0038] Step S104, reviewing the initial storyboard description text according to the content quality score to obtain problem description information of the initial storyboard description text;

[0039] Step S105, performing a first correction training on the storyboard description generation model according to the problem description information of the initial storyboard description text, to obtain a first corrected storyboard description generation model;

[0040] Step S106, obtaining the F-th revised storyboard description text through the F-th revised storyboard description generation model, and obtaining the content quality score of the F-th revised storyboard description text, where F≥1, and F is a positive integer;

[0041] Step S107, if the content quality score of the storyboard description text after the F-th revision is higher than or equal to the preset quality score, the storyboard description text after the F-th revision is used as the storyboard description output content;

[0042] Step S108, otherwise, obtaining the problem description information of the storyboard description text after the Fth revision according to the content quality score of the storyboard description text after the Fth revision, and performing the F+1th revision training on the storyboard description generation model after the Fth revision to obtain the storyboard description generation model after the F+1th revision, and obtaining the storyboard description text after the F+1th revision;

[0043] Step S109, at F+1= , the storyboard description text after the F+1th correction is used as the storyboard description output content, where, is the preset number of iterations.

[0044] According to the automated iteratively optimized content-to-storyboard description generation method of an embodiment of the present invention, a first summary text of a preset text, a second summary text of each chapter, and a third summary text of each paragraph can be obtained respectively, and then the preset chapter can be processed based on the above summary text to obtain an initial storyboard description text of the preset chapter. Based on the summary texts of the full text, chapters, and paragraphs, a storyboard description generation model can refer to the contextual content of the full text, and then the preset chapter is converted into an initial storyboard description text, thereby reducing the probability that the initial storyboard description text will have problems such as generalizing from a particular part to the whole, inconsistency between the context, misunderstanding, and poor coherence. The initial storyboard description text can be evaluated for quality, problems existing in the initial storyboard description text can be found, and corrections can be made based on the problems existing therein, and the processing can be iteratively performed until the quality of the corrected storyboard description text meets the requirements, thereby obtaining a high-quality storyboard description text, thereby further reducing the probability that the initial storyboard description text will have problems such as generalizing from a particular part to the whole, inconsistency between the context, misunderstanding, and poor coherence.

[0045] According to one embodiment of the present invention, the storyboard description includes a description of the lines, inner monologue, actions, expressions, etc. of characters or other objects in a scene, which can be used as a descriptive language in a script. Novels can be converted into storyboard descriptions so that the staff can shoot movies, TV series, dramas and other works according to the storyboard descriptions. However, the work of converting novels into storyboard descriptions is usually done manually, which is a large workload. Although the neural network model in the related art can perform similar tasks, when the novel is long, the neural network model cannot receive a large amount of input content of the novel. Therefore, it is difficult to directly convert the entire novel into a storyboard description. The method of splitting the novel into chapters and inputting them into the neural network model separately can solve the problem that the neural network model cannot receive a large amount of input content of the novel, but this approach is equivalent to severing the connection between chapters, which easily leads to the problems of generalizing, inconsistency, misunderstanding, and poor coherence in the generated storyboard description.

[0046] According to one embodiment of the present invention, in order to solve the above-mentioned problem, when converting a preset chapter into a storyboard description, summaries at various levels can be processed simultaneously in a step-by-step summarization manner, so that when the storyboard description generation model processes the preset chapter, it can grasp the contextual content through summaries at various levels, thereby establishing a connection with the context, thereby reducing the probability of generalizing, inconsistency between the preceding and following contexts, misunderstandings, poor coherence, and other problems.

[0047] According to one embodiment of the present invention, in step S101, the preset text can be processed by a natural language processing model. For example, the processing can be performed step by step in the order of paragraphs, chapters, and full text to obtain a first summary text of the preset text, a second summary text of each chapter, and a third summary text of each paragraph of the chapter. The above preset text can come from a massive text corpus. To ensure the richness and diversity of the data, text materials of different fields, genres, and styles are collected in the text corpus, including but not limited to novels, scripts, literary works, news reports, encyclopedias, etc. The data scale reaches the level of millions of documents, and the total number of words exceeds billions, which provides sufficient materials for the training of natural language processing and storyboard description generation models, enabling them to fully master the laws and knowledge of natural language.

[0048] According to one embodiment of the present invention, the text corpus covers multiple fields, such as science and technology, humanities, art, society, economy, etc., reflecting the breadth of knowledge. In terms of text genres, there are both realistic news reports and fictional novels, taking into account both factual information and imaginative creativity. In terms of writing style, there are both rigorous and delicate academic works and popular and humorous online articles, showing the diversity of language expression. In addition, a large amount of film and television related corpus, such as movie scripts, storyboard scripts, etc., is also included to make the model closer to the actual film and television creation. Rich and diverse data endows the model with extensive knowledge, diverse language styles and creative thinking ability.

[0049] According to one embodiment of the present invention, in order to ensure the high quality of training data, strict data cleaning and screening measures are also taken for articles in the text corpus. On the one hand, the data is deduplicated to remove redundant content, and at the same time, low-quality and meaningless texts, such as advertisements and spam, are filtered out to improve the signal-to-noise ratio of the corpus. On the other hand, an intelligent data review mechanism is developed to use natural language processing technology to analyze the content of the corpus, identify and remove illegal, vulgar, and erroneous content, and ensure the health and reliability of the data from the source. In addition, a manual sampling link is introduced to conduct quality spot checks on the screened data to further control the data quality. After multiple quality control measures, a high-quality and pure corpus is constructed, laying a solid data foundation for the model.

[0050] According to one embodiment of the present invention, in order to facilitate subsequent model training and application, the format of the preset text can also be standardized. For example, the encoding format of the text is unified, such as UTF-8, to ensure that data from different sources are compatible; at the same time, the structured information of the text is extracted and normalized, such as titles, paragraphs, sentences, etc., to make it conform to a unified data model. In addition, the text is also preprocessed as necessary, such as word segmentation, part-of-speech tagging, named entity recognition, etc., to facilitate subsequent feature extraction and model training. Standardized data formats can improve data processing efficiency and ensure consistency in model training.

[0051] Furthermore, the natural language processing model and storyboard description generation model used can use open source language models as the basis. These models are powerful language models trained on massive corpora, with extensive knowledge reserves and excellent language understanding and generation capabilities. They use a neural network architecture based on deep learning, and learn the basic laws and common sense knowledge of language from large-scale unlabeled corpora through self-supervised learning. At the same time, these models show excellent transfer learning capabilities in downstream tasks, and can adapt to different application scenarios after simple fine-tuning. Therefore, using these pre-trained models as the basis can greatly reduce the cost of training from scratch and achieve breakthroughs in results.

[0052] According to one embodiment of the present invention, a preset text is processed by a natural language processing model to obtain a first summary text of the preset text, a second summary text of each chapter of the preset text, and a third summary text of each paragraph of the chapter, including: processing each paragraph by the natural language processing model to obtain a third pending summary text of each paragraph; processing the chapter and the third pending summary text of each paragraph of the chapter by the natural language processing model to obtain a second pending summary text of the chapter; processing the second pending summary text of each chapter by the natural language processing model to obtain a first pending summary text of the preset text; processing the second pending summary text of the chapter by the natural language processing model to obtain a second pending semantic feature vector of the second pending summary text; combining the third pending summary texts of each paragraph of the chapter to obtain a chapter summary combination text, and processing the chapter summary combination text by the natural language processing model to obtain a chapter combination semantic feature vector; processing the first pending summary text of the preset text by the natural language processing model to obtain a first pending semantic feature vector of the first pending summary text The method comprises the steps of: combining the third undetermined summary texts of each paragraph of each chapter of the preset text to obtain a full-text summary combination text, and processing the full-text summary combination text through the natural language processing model to obtain a full-text combination semantic feature vector; determining the loss function of the natural language processing model according to the second undetermined semantic feature vector, the chapter combination semantic feature vector, the first undetermined semantic feature vector and the full-text combination semantic feature vector; updating and training the natural language processing model through the loss function of the natural language processing model to obtain an updated natural language processing model; iteratively executing the steps of updating and training the natural language processing model until the loss function of the natural language processing model is less than or equal to a first preset threshold value to obtain a trained natural language processing model; processing each paragraph through the trained natural language processing model to obtain a third summary text of each paragraph; processing the chapter and the third summary texts of each paragraph of the chapter through the trained natural language processing model to obtain a second summary text of the chapter; processing the second summary text of each chapter through the trained natural language processing model to obtain a first summary text of the preset text.

[0053] According to one embodiment of the present invention, the text of each paragraph can be received as input through a natural language model to obtain the third pending summary text of each paragraph, and in order to generate the summary of the chapter, in addition to inputting the text of the chapter into the natural language model, the third pending summary text of each paragraph of the chapter can also be input into the natural language processing model. The natural language processing model can generate the second pending summary text of the chapter with reference to the third pending summary text of each paragraph, that is, on the basis that the important content of each paragraph has been extracted into the third pending summary text, the second pending summary text of the chapter is generated, so that the important content in the second pending summary text is consistent with the important content of each paragraph, thereby reducing the probability that the important content in the second pending summary text is inaccurate or there is a large amount of redundant content.

[0054] According to an embodiment of the present invention, since it is usually difficult for a natural language processing model to receive an entire input of a long text, the second summary text of each chapter may be processed by the natural language processing model to obtain a first to-be-determined summary text of a preset text.

[0055] According to one embodiment of the present invention, the first pending summary text, the second pending summary text and the third pending summary text may not be accurate. There may be deviations in the semantic understanding of the original text, or the summary of the original text is not concise and accurate enough. Therefore, the natural language processing model can be updated and optimized to improve the accuracy of the summary.

[0056] According to one embodiment of the present invention, the third pending summary texts of each paragraph of a chapter can be combined to obtain a chapter summary combination text, and the chapter summary combination text can be processed by a natural language processing model to obtain a chapter combination semantic feature vector. The second pending summary text of each chapter can also be processed by a natural language processing model to obtain a first pending summary text of the preset text. The chapter summary combination text is a text obtained by combining the summaries of each paragraph in the chapter, and the second pending summary text is a text obtained by extracting the key content of the entire chapter with reference to the summaries of each paragraph. When the paragraph summary and the chapter summary are extracted accurately, the semantics of the two should be consistent.

[0057] According to one embodiment of the present invention, the third pending summary texts of each paragraph of each chapter of the preset text can be combined to obtain a full-text summary combination text, and the full-text summary combination text can be processed by the natural language processing model to obtain a full-text combination semantic feature vector. The first pending summary text of the preset text can also be processed by the natural language processing model to obtain a first pending semantic feature vector of the first pending summary text. The full-text summary combination text is the text obtained by combining the summaries of each paragraph of the full text, and the first pending summary text is the text obtained by combining the summaries of each chapter, and the second pending summary text of each chapter is the text obtained by extracting the key content of the entire chapter with reference to the summaries of each paragraph. Therefore, when the summaries of the paragraphs, the chapters and the full text are accurately extracted, the semantics of the full-text summary combination text and the first pending summary text should be consistent.

[0058] According to an embodiment of the present invention, the loss function of the natural language processing model can be determined based on the above consistency analysis. The loss function of the natural language processing model is determined based on the second undetermined semantic feature vector, the chapter combination semantic feature vector, the first undetermined semantic feature vector and the full text combination semantic feature vector, including: determining the loss function of the natural language processing model according to formula (1): ,

[0059] (1)

[0060] in, is the second undetermined semantic feature vector of the i-th chapter, The chapter combination semantic feature vector of the i-th chapter, for The transposed vector of for and The cosine similarity of is the first undetermined semantic feature vector, is the combined semantic feature vector of the full text, for The transposed vector of for and The cosine similarity of and is the preset weight, n is the number of chapters in the preset text, i≤n, and i and n are both positive integers.

[0061] According to one embodiment of the present invention, in formula (1), is the cosine similarity between the second undetermined semantic feature vector of the ith chapter and the chapter combination semantic feature vector, which can be used to represent the semantic similarity between the second undetermined summary text and the chapter summary combination text. As mentioned above, the semantics of the second undetermined summary text and the chapter summary combination text should be consistent. Therefore, the cosine similarity between the second undetermined semantic feature vector and the chapter combination semantic feature vector should be high. It can be used as a distortion coefficient between the two, that is, a semantic difference coefficient between the two. The loss function can be set based on the distortion coefficient, so that the loss function is reduced during the training process, and the distortion coefficient is reduced, thereby improving the semantic similarity between the second undetermined summary text and the chapter summary combined text. The distortion coefficient between the second undetermined semantic feature vector and the chapter combined semantic feature vector of each chapter can be solved and averaged to obtain the first item of the loss function of the natural language processing model.

[0062] According to one embodiment of the present invention, is the cosine similarity between the first undetermined semantic feature vector and the full-text combined semantic feature vector, which can be used to represent the semantic similarity between the first undetermined abstract text and the full-text combined abstract text. As mentioned above, the semantics of the first undetermined abstract text and the full-text combined abstract text should be consistent. Therefore, the cosine similarity between the first undetermined semantic feature vector and the full-text combined semantic feature vector should be high. It can be used as a distortion coefficient between the two, that is, a semantic difference coefficient between the two. The loss function can be set based on the distortion coefficient, so that the loss function can be reduced during the training process, and the distortion coefficient can be reduced, thereby improving the semantic similarity between the first to-be-determined summary text and the full-text summary combined text. Therefore, the distortion coefficient between the first to-be-determined semantic feature vector and the full-text combined semantic feature vector can be used as the second item of the loss function of the natural language processing model.

[0063] According to one embodiment of the present invention, the above two items can be weightedly summed to obtain the loss function of the natural language processing model, so that the loss function can be back-propagated and the parameters of the natural language processing model can be adjusted using the gradient descent method to improve the accuracy of the natural language processing model.

[0064] According to one embodiment of the present invention, the above training steps can be iteratively performed, that is, the third pending summary text, the second pending summary text, the first pending summary text, the second pending semantic feature vector, the chapter summary combination text, the chapter combination semantic feature vector, the first pending semantic feature vector, the full-text summary combination text, the full-text combination semantic feature vector and other parameters are obtained again using the natural language processing model after each update, and the above loss function is used to train the natural language processing model again. After the above update training steps are iteratively performed multiple times and the loss function of the natural language processing model is less than or equal to the first preset threshold, a trained natural language processing model is obtained, and the trained natural language processing model can achieve a higher accuracy. The trained natural language processing model can be used to process each paragraph to obtain the third summary text of each paragraph, and the trained natural language processing model can be used to process the chapter and the third summary text of each paragraph of the chapter to obtain the second summary text of the chapter, and the trained natural language processing model can be used to process the second summary text of each chapter to obtain the first summary text of the preset text.

[0065] In this way, the loss function of the natural language processing model can be set by the semantic similarity and distortion between the first pending summary text and the combined text of the full-text summary, as well as the semantic similarity and distortion between the second pending summary text and the combined text of the chapter summary, and the natural language processing model can be iteratively updated and trained to improve the accuracy of the natural language processing model, thereby obtaining accurate summary texts at multiple levels, so that when the storyboard description is subsequently generated, the accurate summary texts at multiple levels can be used, so that the storyboard description generation model can grasp richer contextual information as much as possible when it cannot receive the input of the long text, so that the storyboard description is closer to the semantics of the original text, and the probability of generalizing, inconsistency between the context, poor coherence and other problems is reduced.

[0066] According to one embodiment of the present invention, in step S102, the first summary text, the second summary text, the third summary text and the preset chapter can be processed by a pre-trained storyboard description generation model to obtain an initial storyboard description text corresponding to the preset chapter, that is, the preset chapter can be converted into an initial storyboard description text with reference to the context information provided by the first summary text, the second summary text and the third summary text.

[0067] According to one embodiment of the present invention, before using the storyboard description generation model, the storyboard description generation model can be pre-trained so that the storyboard description generation model can generate a storyboard description text with a format that meets the requirements, a high consistency in content with preset chapters, and a low probability of problems such as generalizations, inconsistencies in the context, and poor coherence.

[0068] According to one embodiment of the present invention, a storyboard description generation model can be pre-trained using sample text. For example, the sample text is a text such as a novel, and a plurality of chapters are selected from the sample text, so that the storyboard description generation model is processed with the sample text to generate sample storyboard description text. The storyboard description generation model can then be trained based on the differences between the sample storyboard description text and manually written storyboard description text in terms of format, level of detail, writing style, and logical and semantic rigor. For example, the differences in character descriptions, scene descriptions, behavior descriptions, line descriptions, emotional expressions, etc. that appear in sample storyboard description texts and manually written storyboard description texts can be statistically analyzed, and the logical errors, semantic errors, format errors, etc. that appear in the sample storyboard description texts can be statistically analyzed. Based on these statistical results, the loss function of the storyboard description generation model is determined, and then the storyboard description generation model is trained through the loss function of the storyboard description generation model. For example, the loss function can be reduced through the gradient descent method, and then the parameters of the storyboard description generation model can be adjusted to improve the accuracy of the storyboard description generation model, so that the differences between the sample storyboard description text generated by the storyboard description generation model and the manually written storyboard description text in terms of format, level of detail, writing style, logical and semantic rigor, etc. are minimized as much as possible. After multiple trainings, the pre-trained storyboard description generation model can be obtained.

[0069] In the example, the sample text may come from the above text corpus, and multiple sample texts may be selected from the text corpus, and the multiple sample texts may be divided into a training set, a validation set, and a test set. When using the sample text in the training set for training, a learning rate decay strategy is adopted, that is, as the number of training rounds increases, the learning rate gradually decreases to achieve a more stable convergence. At the same time, when the performance on the validation set reaches the optimal level and begins to decline, the training is stopped in time to prevent overfitting.

[0070] According to one embodiment of the present invention, in step S103, after the pre-trained storyboard description generation model generates the initial storyboard description text, the content quality of the initial storyboard description text may be evaluated, so as to determine the improvement method of the initial storyboard description text based on the evaluation result, and obtain a more accurate storyboard description text after the improvement. The content quality of the initial storyboard description text may be evaluated based on the original text of the preset chapter, as well as the first summary text, the second summary text, and the third summary text. For example, it may be evaluated whether the initial storyboard description text misinterprets the semantics of the original text, whether there are logical errors, whether there are scenes and characters that have not appeared in the chapter, whether there are plot errors, etc.

[0071] According to an embodiment of the present invention, the content quality of the initial storyboard description text is evaluated according to the preset chapter, the first summary text, the second summary text and the third summary text to obtain a content quality score of the initial storyboard description text, including: determining an accuracy score of the initial storyboard description text according to the initial storyboard description text, the third summary text and the second summary text; determining a key information coverage score of the initial storyboard description text according to the second summary text and the initial storyboard description text; determining an information completeness score according to the initial storyboard description text and the preset chapter; determining an original meaning retention score according to the preset chapter, the second summary text and the initial storyboard description text; determining a semantic error detection score according to the preset chapter, the initial storyboard description text, the first summary text and the second summary text; determining a content quality score of the initial storyboard description text according to the accuracy score, the key information coverage score, the information completeness score, the original meaning retention score and the semantic error detection score.

[0072] According to one embodiment of the present invention, the accuracy score of the initial storyboard description text may be used to evaluate whether the descriptions of scenes, characters, lines, actions, etc. in the initial storyboard description text are consistent with the original text of a preset chapter.

[0073] According to an embodiment of the present invention, the accuracy score of the initial storyboard description text is determined based on the initial storyboard description text, the third summary text and the second summary text, including: searching the first character description text, the first location description text and the first action description text in the initial storyboard description text through the natural language processing model, and determining the first character semantic vector of the first character description text, the first location semantic vector of the first location description text and the first action semantic vector of the first action description text; combining the second summary text of a preset chapter and the third summary text of each paragraph of the preset chapter to obtain a preset chapter summary text; searching the second character description text, the second location description text and the second action description text of the preset chapter summary text through the natural language processing model, and determining the second character semantic vector of the second character description text, the second location semantic vector of the second location description text and the second action semantic vector of the second action description text; and determining the accuracy score of the initial storyboard description text according to formula (2). ,

[0074] (2)

[0075] in, is the second person semantic vector of the jth second person description text, is the first person semantic vector of the kth first person description text, The number of texts describing the first character. The number of description texts for the second character, for The transposed vector of for and The cosine similarity of is the second action semantic vector of the t-th second action description text, For the The first action semantic vector of the first action description text, for The transposed vector of for and The cosine similarity of is the number of the first action description text, The number of description texts for the second action, is the second location semantic vector of the x-th second location description text, is the first place semantic vector of the yth first place description text, for The transposed vector of for and The cosine similarity of is the number of texts describing the first location, is the number of description texts of the second location, max is the maximum value function, , and are preset weights, k, j, ,t,x,y, , , , , and All are positive integers.

[0076] According to one embodiment of the present invention, a natural language processing model can be used to search for a first character description text (e.g., description text of a character's name), a first place description text (e.g., description text of scenes such as indoors, on the street, by the river, on the mountain, etc.), and a first action description text (e.g., an action that a character or other target is doing, such as running, turning around, driving, etc.) in the initial storyboard description text. Furthermore, the encoding layer of the natural language processing model can obtain a first character semantic vector of the first character description text, a first place semantic vector of the first place description text, and a first action semantic vector of the first action description text, so that the above natural language texts can be described by information in vector form.

[0077] According to one embodiment of the present invention, the second summary text of a preset chapter may include brief description information of important characters, important scenes and important actions in the preset chapter, that is, the second summary text may include description texts of a part of the plots with higher importance in the preset chapter, but the second summary text may not include description texts of all the plots. Therefore, the third summary text and the second summary text of each paragraph in the preset chapter can be combined to obtain the preset chapter summary text. The third summary text of each paragraph may include description texts of the plots with higher importance in each chapter, and the richness of the important plots included in the third summary texts of all paragraphs is higher than that of the second summary text. Therefore, when judging whether the initial storyboard description text accurately includes the important plots in the preset chapter, the preset chapter summary text can be used as a comparative reference to determine whether the important characters, important scenes, and important actions in the initial storyboard description text are described correctly and completely. Therefore, the second character description text, the second place description text, and the second action description text of the preset chapter summary text can be searched through a natural language model, and the second character semantic vector of the second character description text, the second place semantic vector of the second place description text, and the second action semantic vector of the second action description text can be determined for comparison with the above-mentioned first character semantic vector, first place semantic vector, and first action semantic vector.

[0078] According to one embodiment of the present invention, in formula (2), It means that among multiple first character semantic vectors, the first character semantic vector with the highest cosine similarity with the jth second character semantic vector is found, and the cosine similarity between the first character semantic vector and the jth second character semantic vector is the value of this formula. If the value of this formula is high, it means that in the initial storyboard description text, there is a first character description text that is consistent with the jth second character description text in the preset chapter summary text. In other words, the description of the jth second character description text in the initial storyboard description text is correct. Otherwise, it means that in the initial storyboard description text, there is no first character description text that is consistent with the jth second character description text in the preset chapter summary text, or the description of the jth second character description text in the initial storyboard description text is wrong. Therefore, is the description accuracy rate of multiple second character description texts in the preset chapter summary text in the initial storyboard description text. Similarly, is the description accuracy rate of multiple second action description texts in the preset chapter summary text in the initial storyboard description text, It is the description accuracy rate of multiple second location description texts in the preset chapter summary text in the initial storyboard description text. The accuracy score of the initial storyboard description text can be obtained by weighted summing the above three description accuracy rates, which can be used to describe the description accuracy rate of important plots in the initial storyboard description text.

[0079] In this way, by comparing the preset chapter summary text with the initial storyboard description text, the accuracy rate of the description of characters, places and actions in the initial storyboard description text can be solved, thereby obtaining the accuracy score of the initial storyboard description text, which can objectively and accurately describe the accuracy rate of the description of important plots in the initial storyboard description text.

[0080] According to one embodiment of the present invention, the key information coverage score of the initial storyboard description text is based on the second summary text as a reference to determine whether the key information in the initial storyboard description text can cover the important plot with high importance in the second summary text. For example, it can be compared whether the information such as characters, actions, locations, lines, etc. in the second summary text can be covered by the information such as objects, actions, locations, lines, etc. in the initial storyboard description text, so as to determine whether the initial storyboard description text can accurately reflect the important information of the preset chapter.

[0081] According to one embodiment of the present invention, the information completeness score of the initial storyboard description text is based on the original text of the preset chapter as a reference to determine whether the key information in the initial storyboard description text can cover the key information in the original text of the preset chapter. That is, compared with the above comparison of only the important plots of the initial storyboard description text with the important plots in the preset chapter summary text, when determining the information completeness score of the initial storyboard description text, the information to be compared is richer and more comprehensive. For example, in addition to the above-mentioned information such as characters, actions, locations, etc., the initial storyboard description text and the information such as weather, time, background introduction, character language, etc. in the preset chapter can also be compared to determine whether the initial storyboard description text can cover the important information in the original text of the preset chapter, thereby judging whether the initial storyboard description text can accurately reflect the original text of the preset chapter.

[0082] According to one embodiment of the present invention, the original meaning retention score can be used to describe whether the initial storyboard description text retains the original meaning of the preset chapter and whether it exceeds the description scope of the preset chapter, that is, whether the plot of other chapters is displayed in the initial storyboard description text to cause defects such as spoilers.

[0083] According to one embodiment of the present invention, determining the original meaning retention score according to the preset chapter, the second summary text and the initial storyboard description text includes: searching the first character description text, the first location description text and the first action description text in the initial storyboard description text through the natural language processing model, and determining the first character semantic vector of the first character description text, the first location semantic vector of the first location description text and the first action semantic vector of the first action description text; determining the initial storyboard description semantic information of the initial storyboard description text through the natural language processing model; determining the chapter semantic information of the preset chapter through the natural language processing model; determining the original meaning retention score according to ... The natural language processing model is used to search for the third character description text, the third place description text and the third action description text in the second summary text of other chapters except the preset chapter, and determine the third character semantic vector of the third character description text, the third place semantic vector of the third place description text and the third action semantic vector of the third action description text; the natural language processing model is used to search for the fourth character description text, the fourth place description text and the fourth action description text in the preset chapter, and determine the fourth character semantic vector of the fourth character description text, the fourth place semantic vector of the fourth place description text and the fourth action semantic vector of the fourth action description text; the original meaning retention score is determined according to formula (3) ,

[0084] (3)

[0085] in, Describe semantic information for the initial storyboard, is the chapter semantic information of the preset chapter, is the first person semantic vector of the kth first person description text, is the fourth person semantic vector of the jth fourth person description text, is the third person semantic vector of the tth third person description text, The number of description texts for the fourth character, The number of description texts for the third person, The number of texts describing the first character. For the The first action semantic vector of the first action description text, is the fourth action semantic vector of the tth fourth action description text, is the third action semantic vector of the yth third action description text, is the number of the first action description text, The number of description texts for the fourth action, The number of description texts for the third action, is the first place semantic vector of the zth first place description text, is the fourth place semantic vector of the u-th fourth place description text, is the third place semantic vector of the vth third place description text, is the number of texts describing the first location, The number of texts describing the third location, The number of texts describing the fourth location, is the first preset similarity, is the second preset similarity, if is the conditional function, max is the maximum value function, , , and are preset weights, k, j, t, , x, y, z, u, v, , , , , , , , and Is a positive integer.

[0086] According to one embodiment of the present invention, the initial storyboard description text as a whole can be used as input to a natural language processing model to obtain initial storyboard description semantic information, and the preset chapter as a whole can be used as input to a natural language processing model to obtain chapter semantic information.

[0087] According to one embodiment of the present invention, important plots in other chapters can be searched in the second summary texts of all chapters except preset chapters, for example, the third character description text, the third place description text and the third action description text in the second summary texts of other chapters, and the third character semantic vector, the third place semantic vector and the third action semantic vector can be extracted respectively.

[0088] According to an embodiment of the present invention, the fourth character description text, the fourth location description text and the fourth action description text that appeared in the original text of the preset chapter can also be searched, and the fourth character semantic vector, the fourth location semantic vector and the fourth action semantic vector can be extracted respectively. Thus, the third character semantic vector, the third location semantic vector and the third action semantic vector, as well as the fourth character semantic vector, the fourth location semantic vector and the fourth action semantic vector can be combined to determine whether important plots that do not appear in the preset chapter but appear in other chapters appear in the initial storyboard description text, thereby causing defects such as spoilers.

[0089] According to one embodiment of the present invention, in formula (3), It is the cosine similarity between the initial storyboard description semantic information and the chapter semantic information, which can be used to represent the overall semantic similarity between the initial storyboard description text and the preset chapter original text, that is, it can be used to determine whether the initial storyboard description text can accurately express the original meaning of the preset chapter original text.

[0090] According to one embodiment of the present invention, in formula (3), the conditional function Can be used to indicate that the condition is met In the case of , the value of the conditional function is the inner conditional function , otherwise it is 0. It means that among the fourth character semantic vectors of multiple fourth character description texts, the fourth character semantic vector with the highest cosine similarity to the first character semantic vector of the kth first character description text is found, and the cosine similarity between the fourth character semantic vector and the first character semantic vector of the kth first character description text is the value of this formula. If the value of this formula is less than the first preset similarity, it means that in the original text of the preset chapter, there is no description similar to the kth first character description text, that is, if in the original text of the preset chapter, there is no description similar to the kth first character description text, then the value of this conditional function is the value of the inner conditional function. If in the original text of the preset chapter, there is a description similar to the kth first character description text, then the value of this conditional function is 0, indicating that there is no error in the kth first character description text. In the inner conditional function, It means that among the third character semantic vectors of multiple third character description texts, the third character semantic vector with the highest cosine similarity to the first character semantic vector of the kth first character description text is found, and the cosine similarity between the third character semantic vector and the first character semantic vector of the kth first character description text is the value of this formula. The value of this formula is greater than or equal to the second preset similarity, indicating that a description similar to the kth first character description text appears in the second summary text of other chapters, that is, in the initial storyboard description text, character descriptions of other chapters appear, and there are defects such as spoilers, then the inner conditional function value is 1, otherwise it is 0. Therefore, when there is no description similar to the kth first character description text in the preset chapter, and there is a description similar to the kth first character description text in other chapters, the function value of the outer conditional function corresponding to the kth first character description text is 1, and the function values ​​of the outer conditional function corresponding to each first character description text are summed to obtain the number of character descriptions with spoilers in the initial storyboard description text. Therefore, It is the ratio of the first character description texts without spoiler defects among all the first character description texts in the initial storyboard description texts.

[0091] According to one embodiment of the present invention, similar to the above, It indicates the ratio of the first action description texts without spoiler defects among all the first action description texts in the initial storyboard description texts. It indicates the ratio of the first location description texts that do not contain spoilers among all the first location description texts in the initial storyboard description texts.

[0092] According to one embodiment of the present invention, the above four items are weighted and summed to obtain an original meaning retention score. The higher the original meaning retention score, the more similar the semantics of the initial storyboard description text is to the original text of the preset chapter, and the lower the probability of defects such as spoilers in important plots.

[0093] In this way, the original meaning retention score can be determined based on factors such as the semantic similarity between the initial storyboard description text and the original text of the preset chapter, and whether important plots such as characters, places, actions in the initial storyboard description text appear in the original text of the preset chapter, and whether they appear in the second summary text of other chapters. This can accurately and objectively indicate whether the initial storyboard description text can maintain the original meaning of the preset chapter, and whether there is a defect of important plot spoilers.

[0094] According to one embodiment of the present invention, the semantic error detection score is used to describe whether there is a single sentence in the initial storyboard description text whose semantics are inconsistent with the original text, and whether there is an erroneous description that deviates from the preset text in the initial storyboard description text. For example, a certain description does not exist in all chapters of the preset text original, and is a text generated by the storyboard description generation model based on erroneous understanding. In this case, the semantic error detection score can be used to detect whether there is such a serious error in the initial storyboard description text.

[0095] According to one embodiment of the present invention, the first character description text, the first location description text and the first action description text in the initial storyboard description text are searched through the natural language processing model, and the first character semantic vector of the first character description text, the first location semantic vector of the first location description text and the first action semantic vector of the first action description text are determined; the first sentence semantic information of each sentence in the initial storyboard description text is determined through the natural language processing model; the second sentence semantic information of each sentence in the preset chapter is determined through the natural language processing model; the first summary text and the second summary text are combined to obtain a full-text summary combined text; and the first summary text is combined through the natural language processing model to obtain a full-text summary combined text. Through the natural language processing model, the fourth character description text, the fourth place description text and the fourth action description text in the preset chapter are searched, and the fourth character semantic vector of the fourth character description text, the fourth place semantic vector of the fourth place description text and the fourth action semantic vector of the fourth action description text are determined; through the natural language processing model, the fifth character description text, the fifth place description text and the fifth action description text in the full text summary combined text are searched, and the fifth character semantic vector of the fifth character description text, the fifth place semantic vector of the fifth place description text and the fifth action semantic vector of the fifth action description text are determined; the semantic error detection score is determined according to formula (4) ,

[0096] (4)

[0097] in, is the semantic information of the first sentence of the hth sentence in the initial storyboard description text, is the semantic information of the second sentence of the gth sentence in the preset chapter, is the number of sentences in the preset chapter, is the number of sentences in the initial storyboard description text, is the first person semantic vector of the kth first person description text, is the fourth person semantic vector of the jth fourth person description text, is the fifth character semantic vector of the tth fifth character description text, The number of description texts for the fourth character, The number of texts describing the fifth character, The number of texts describing the first character. For the The first action semantic vector of the first action description text, is the fourth action semantic vector of the tth fourth action description text, is the fifth action semantic vector of the yth fifth action description text, is the number of the first action description text, The number of description texts for the fourth action, The number of description texts for the fifth action, is the first place semantic vector of the zth first place description text, is the fourth place semantic vector of the u-th fourth place description text, is the fifth place semantic vector of the vth fifth place description text, is the number of texts describing the first location, The number of texts describing the fifth location, The number of texts describing the fourth location, is the first preset similarity, if is the conditional function, max is the maximum value function, , , and are preset weights, k, j, t, , x, y, z, u, v, , , , , , , , and Is a positive integer.

[0098] According to one embodiment of the present invention, each sentence in the initial storyboard description text can be used as an input of a natural language processing model to obtain the first sentence semantic information of each sentence. Each sentence in a preset chapter can be used as an input of a natural language processing model to obtain the second sentence semantic information of each sentence, so that the second sentence semantic information can be used as a reference to determine whether each sentence in the initial storyboard description text is the same or similar in semantics to the original text of the preset chapter. Further, the first summary text and the second summary text can be combined to obtain a full-text summary combination text, in which important plot descriptions in the full text are included, for example, important character descriptions, location descriptions, and action descriptions in the full text. The fifth character description text, the fifth location description text, and the fifth action description text in the full-text summary combination text can be searched through a natural language processing model, and the fifth character semantic vector, the fifth location semantic vector, and the fifth action semantic vector can be extracted to determine whether there are erroneous descriptions in the initial storyboard description text that do not exist in the full text generated by the storyboard description generation model based on erroneous understanding.

[0099] According to one embodiment of the present invention, in formula (4), the conditional function Indicated in When , the conditional function value is 1, otherwise it is 0, where It means that among the second sentence semantic information of multiple sentences in the preset chapter, the second sentence semantic information with the highest cosine similarity to the first sentence semantic information of the h-th sentence in the initial storyboard description text is searched, and the cosine similarity between the second semantic information and the first sentence semantic information of the h-th sentence in the initial storyboard description text is the value of this formula. If the value of this formula is less than the first preset similarity, it means that among the multiple sentences in the preset chapter, there is no sentence with semantics similar to the h-th sentence in the initial storyboard description text. Therefore, the semantics of the h-th sentence in the initial storyboard description text is inconsistent with the original text. The conditional functions corresponding to each sentence in the initial storyboard description text are summed to obtain the number of sentences in the initial storyboard description text whose semantics are inconsistent with the original text. Therefore, It is the ratio of the number of sentences in the initial storyboard description text that are semantically consistent with the original text.

[0100] According to one embodiment of the present invention, in formula (4), the conditional function Indicated in In the case of otherwise it is 0. The meaning of has been explained in the interpretation of formula (3) and will not be repeated here. In the inner condition function, It means that among the fifth character semantic vectors of multiple fifth character description texts, the fifth character semantic vector with the highest cosine similarity to the first character semantic vector of the kth first character description text is found, and the cosine similarity between the fifth character semantic vector and the first character semantic vector of the kth first character description text is the value of the formula. Therefore, It means that in the combined text of the full text summary that describes the key plot of the full text, there is no character description similar to the kth first character description text. In this case, the value of the inner conditional function is 1, otherwise it is 0. Therefore, when there is no character description similar to the kth first character description text in both the preset chapter original text and the combined text of the full text summary, the value of the above conditional function is 1, otherwise it is 0. In other words, when the kth first character description text in the initial storyboard description text is an erroneous description that does not exist in the full text generated by the storyboard description generation model based on erroneous understanding, the conditional function value is 1. The conditional functions corresponding to each first character description text in the initial storyboard description text are summed to obtain the number of erroneous first character description texts in the initial storyboard description text. Therefore, The ratio of the first character description texts that appear in the full text of the preset text among the multiple first character description texts of the initial storyboard description text. is the ratio of the number of first action description texts that appear in the full text of the preset text among the multiple first action description texts of the initial storyboard description text, The ratio of the number of first place description texts that appear in the full text of the preset text among the multiple first place description texts of the initial storyboard description text. The weighted sum of the above four items can be used to obtain the semantic error detection score, which can be used to describe the semantic accuracy of each sentence in the initial storyboard description text and whether there are plots in the initial storyboard description text that are generated by the storyboard description generation model based on incorrect understanding.

[0101] In this way, the semantic accuracy of each sentence in the initial storyboard description text can be determined by comparing the sentences in the initial storyboard description text with the sentences in the preset chapter original text sentence by sentence, and based on whether important plots such as characters, places, actions in the initial storyboard description text appear in the full text, it can be determined whether there are erroneous plots in the initial storyboard description text generated by the storyboard description generation model based on incorrect understanding, thereby improving the accuracy and objectivity of the semantic error detection score.

[0102] According to one embodiment of the present invention, the accuracy score, the key information coverage score, the information completeness score, the original meaning retention score and the semantic error detection score may be weighted and summed to obtain a content quality score for the initial storyboard description text. The higher the content quality score, the more accurate the content of the initial storyboard description text is, and the more it can reflect the semantics of the original text of the preset chapter.

[0103] According to one embodiment of the present invention, in step S104, if the content quality score is lower than a preset quality score, the initial storyboard description text is reviewed to obtain problem description information of the initial storyboard description text, that is, the problems and errors existing in the initial storyboard description text are reviewed, and problem description information is given, and the problem description information can be used to provide an accurate correction method.

[0104] According to one embodiment of the present invention, the initial storyboard description text is reviewed according to the content quality score to obtain problem description information of the initial storyboard description text, including: if the content quality score is lower than a preset quality score threshold, determining the first sentence semantic information of each sentence in the initial storyboard description text through the natural language processing model; determining the second sentence semantic information of each sentence in the preset chapter through the natural language processing model; determining, according to the first sentence semantic information and the second sentence semantic information, a first sentence that is semantically consistent with the sentences in the preset chapter and a second sentence that is semantically inconsistent with the sentences in the preset chapter among multiple sentences in the initial storyboard description text; obtaining a first emotion expression vector of the first sentence and a second emotion expression vector of the first target sentence in the preset chapter that has the highest semantic similarity with the first sentence through a pre-trained sentence sentiment analysis model; determining a second target sentence in the preset chapter other than the first target sentence, and obtaining the second target semantic information of the second target sentence through the natural language processing model; determining a logical review score according to a formula ,

[0105] (5)

[0106] in, is the number of the first sentence, is the number of the second sentence, and is a positive integer;

[0107] Determine the expression completeness review score according to formula (6) ,

[0108] (6)

[0109] in, is the number of the first target sentences, is the number of the second target sentences, and is a positive integer;

[0110] Determine the emotional expression review score according to formula (7) ,

[0111] (7)

[0112] in, is the first sentiment expression vector of the first sentence of the ath sentence, is the second sentiment expression vector of the first target sentence corresponding to the ath first sentence, a is a positive integer, and a≤ ; Determine the problem description information of the initial storyboard description text based on the logic review score, the expression integrity review score and the emotional expression review score.

[0113] According to one embodiment of the present invention, during the review process, the first item of the above formula (4) can be used to determine the first sentence in the multiple sentences of the initial storyboard description text that is consistent with the semantics of the original text of the preset chapter, and the second sentence that is inconsistent with the semantics of the original text of the preset chapter, and count the number of the first sentences and the number of the second sentences respectively. In formula (5), the logical review score can be used to describe the proportion of the first sentences in the multiple sentences of the initial storyboard description text that are consistent with the semantics of the original text of the preset chapter in the multiple sentences of the initial storyboard description text, that is, the semantic correctness of the multiple sentences of the initial storyboard description text.

[0114] According to one embodiment of the present invention, in the original text of the preset chapter, the sentence that is semantically consistent with the above-mentioned first sentence is the first target sentence, and the other sentences are the second target sentences. In formula (6), the expression completeness review score is the proportion of the first target sentence in the sentences in the original text of the preset chapter, which can also represent the proportion of the sentences of the preset chapter that can be accurately described by the initial storyboard description text in the sentences in the original text of the preset chapter, which can be used as the expression completeness review score of the initial storyboard description text.

[0115] According to one embodiment of the present invention, the first emotion expression vector of the first sentence can be used to describe the emotional state described in the first sentence, for example, to describe the joy, excitement, calmness, sadness, anger and other emotions of the character in the first sentence, and can also describe whether the semantic expression in the first sentence is affirmative or negative. The first sentence and the first target sentence have a high degree of semantic similarity, but the emotional expressions of the two may not be the same. For example, in the first target sentence, a plot in which a character excitedly said a certain line is described, while in the first sentence, a plot in which the character calmly said the line is described. The characters, locations, and actions (i.e., speaking actions) of the two are consistent. Therefore, the semantic similarity between the two is high, but there are differences in the emotional expressions of the two. Therefore, the first emotion expression vector of the first sentence and the second emotion expression vector of the first target sentence can be determined respectively by a pre-trained sentence emotion analysis model, and the emotion expression review score can be determined based on formula (7). In formula (7), is the cosine similarity between the first emotion expression vector of the a-th first sentence and the second emotion expression vector of the a-th first target sentence, which can be used to represent the similarity of the emotion expression of the first sentence and the first target sentence. Furthermore, the similarity of the emotion expression between each first sentence and the first target sentence can be averaged, so as to describe the overall accuracy of the emotion expression of each first sentence. Moreover, since the first sentence is only a part of the multiple sentences in the initial storyboard description text, there is also a second sentence in the initial storyboard description text that does not conform to the semantics of the original text. For the second sentence, there is no need to analyze its emotion expression. Therefore, based on the overall accuracy of the emotion expression of each first sentence, multiply it by the logical review score, that is, the proportion of the number of first sentences that conform to the semantics of the original text in the multiple sentences of the initial storyboard description text, and the overall accuracy of the emotion expression of the initial storyboard description text can be obtained.

[0116] According to one embodiment of the present invention, the logic review score, expression integrity review score and emotional expression review score are obtained above, and the problem description information can be determined based on these three scores. For example, if the logic review score is low, it can be indicated in the problem description information that the logic, semantics and other aspects of the initial storyboard description text are quite different from the original text. When the storyboard description text is subsequently regenerated using the revised storyboard description generation model, the corresponding prompt words can be obtained based on the problem description information. For example, if the storyboard description generation model is a language model such as GPT, after it is revised and trained, in addition to inputting the first summary text, the second summary text, the third summary text and the preset chapters in the preset text, prompt words such as "correct the logic at the XXX position", "do not miss the XXX plot", and "the emotional expression at XXX follows the original text" can also be input, thereby generating a more accurate storyboard description text.

[0117] In this way, the problems existing in the initial storyboard description text can be reviewed through the semantic and logical similarity between the initial storyboard description text and the original text, the completeness of the expression of the original text, and the similarity with the original text in emotional expression, so as to obtain accurate problem description information, and provide a data basis for the correction training of the storyboard description generation model and the correction of the initial storyboard description text.

[0118] According to one embodiment of the present invention, in step S105, after obtaining the problem description information of the initial storyboard description text, the storyboard description generation model may be corrected and trained so that the storyboard description generation model can reduce the probability of the problem described in the problem description information, thereby obtaining a corrected storyboard description text.

[0119] According to an embodiment of the present invention, the storyboard description generation model is subjected to a first correction training based on the problem description information of the initial storyboard description text to obtain the first corrected storyboard description generation model, including: determining the loss function of the storyboard description generation model when the first correction training is performed according to formula (8): ,

[0120] (8)

[0121] in, is the semantic information of the first sentence of the rth sentence in the initial storyboard description text, is the second sentence semantic information of the sentence having the highest semantic similarity with the first sentence semantic information of the rth sentence of the initial storyboard description text among the plurality of sentences in the preset chapter, is the first sentiment expression vector of the rth sentence in the initial storyboard description text, is the second emotion expression vector of the sentence with the highest semantic similarity to the semantic information of the first sentence of the rth sentence of the initial storyboard description text among the multiple sentences of the preset chapter, is the second target semantic information of the qth second target sentence, r≤ , q≤ , and r and q are positive integers; the storyboard description generation model is trained for the first time through the loss function during the first correction training to obtain the storyboard description generation model after the first correction.

[0122] According to one embodiment of the present invention, in formula (8), is the cosine similarity between the r-th sentence in the initial storyboard description text and the sentence with the highest similarity to the r-th sentence in the original text of the preset chapter. Therefore, The distortion degree of the rth sentence of the initial storyboard description text to the original text of the preset chapter can be averaged to obtain the overall distortion degree of the initial storyboard description text to the original text. The loss function can be set based on the overall distortion degree to reduce the distortion degree during the training process. As the weight of the overall distortion degree, that is, the higher the overall semantic similarity between the initial storyboard description text and the original text, the lower the distortion degree, the lower the weight; otherwise, the weight is increased, thereby conducting targeted training.

[0123] According to one embodiment of the present invention, is the modulus of the second target semantic information of the second target sentence. In the initial storyboard description text, there is no sentence that is semantically consistent with the second target sentence. Therefore, the modulus of the second target sentence as a whole can be used as part of the loss function. The existence of the second target sentence indicates that there is an error in the semantics between the initial storyboard description text and the original text. The larger the modulus of the second target semantic information, the greater the error in the semantics between the initial storyboard description text and the original text. Therefore, the average value of the modulus of the second target semantic information is used as part of the loss function. During the training process, the modulus of the second target semantic information can be reduced by improving the semantic similarity between the initial storyboard description text and the original text, thereby achieving the purpose of improving the accuracy of the storyboard description generation model. Furthermore, As the weight of this item, that is, the higher the expression completeness review score, the more complete the initial storyboard description text is in expressing the original text, and the lower the possibility of the existence of the second target sentence, the lower the weight. Otherwise, the weight is increased, so as to carry out targeted training.

[0124] According to one embodiment of the present invention, is the similarity of the emotional expression between the rth sentence of the initial storyboard description text and the sentence with the highest similarity in the original text of the preset chapter, is the difference in emotional expression between the two, and the average value of the difference can be taken as the overall difference in emotional expression between the initial storyboard description text and the preset chapter. As the weight of this item, that is, the higher the emotional expression review score is, the smaller the overall difference in emotional expression between the initial storyboard description text and the preset chapter is, the lower the weight is; otherwise, the weight is increased, thereby conducting targeted training.

[0125] According to one embodiment of the present invention, the above three items can be summed to obtain the loss function during the first revised training, and the parameters of the storyboard description generation model can be adjusted by the gradient descent method to improve the accuracy of the storyboard description generation model and obtain the storyboard description generation model after the first revision.

[0126] In this way, the loss function of the first correction training can be determined based on the overall distortion of the original text by the initial storyboard description text, the semantic error between the initial storyboard description text and the original text, and the overall difference in emotional expression between the initial storyboard description text and the preset chapters. The weights of the above aspects can be set through the problem description information, so as to focus on the aspects with serious problems and improve the pertinence and efficiency of the training.

[0127] According to one embodiment of the present invention, the above process of solving multiple scores, reviewing and determining the problem description information, and correcting the training can be used as a posterior design process, and the above multiple scores can be used as an evaluation to determine whether the model is effective in the posterior process, and the problem description information in the review process can be used as a penalty item in the posterior process, that is, a penalty for problems that arise in the model, and in the training process, the above penalty items are used to set optimization goals, that is, minimizing the loss function is used as the optimization goal, thereby completing a model update iteration, and then using the updated iterative model to obtain more accurate processing results.

[0128] According to an embodiment of the present invention, in step S106, the storyboard description generation model after the Fth revision can be used to obtain the storyboard description text after the Fth revision, and the content quality score of the storyboard description text after the Fth revision can be obtained based on a similar method as above, wherein F is a positive integer greater than or equal to 1. In step S107, if the content quality score of the storyboard description text after the Fth revision is higher than or equal to the preset quality score, the storyboard description text after the Fth revision is used as the storyboard description output content; otherwise, in step S108, a review process similar to the above is performed, and the problem description information obtained by the review is used to perform the F+1th revision training on the storyboard description generation model after the Fth revision, to obtain the storyboard description generation model after the F+1 revision, and then obtain the storyboard description text after the F+1 revision based on the storyboard description generation model after the F+1 revision. The above revision process can be iteratively performed until the content quality score of the revised storyboard description text is higher than or equal to the preset quality score, and then the storyboard description output content can be obtained. Alternatively, if the number of iterations is too many and reaches a preset number of iterations, the iterations may be stopped, and in step S109, the storyboard description text corrected for the F+1th time may be used as the storyboard description output content.

[0129] According to one embodiment of the present invention, after the above storyboard description generation model training is completed, the model can also be compressed. In order to enable the model to run efficiently in a resource-limited environment, a variety of model compression technologies can be used to significantly reduce the size and computational overhead of the model while ensuring performance. In the example, the knowledge distillation technology can be used, that is, the large storyboard description generation model is used as the teacher model, and the knowledge and performance of the teacher model are transferred to the small student model. Specifically, the storyboard description generation model is used to process the input text, and the processing results can be used as annotation information, and then the annotation information is used as supervision to train the student model. In this way, the student model can learn the core knowledge of the storyboard description generation model, and the scale of the model is significantly reduced. In the example, the model quantization technology can be used to quantize the weights and activation values ​​of the model, converting them from 32-bit floating point numbers to 8-bit or 16-bit fixed point numbers, thereby reducing the storage and computational overhead of the model. In the example, the model pruning technology can be used to prune and optimize the structure of the model, removing redundant and unimportant neuron connections. Through sparse regularization and importance-based connection pruning strategies, the model structure is significantly simplified while maintaining the expressive power of the model. The pruned model is not only smaller and faster to calculate, but also reduces the risk of overfitting.

[0130] Through the above model compression technology, the model size can be greatly reduced and the operation speed can be greatly improved while the model performance is almost lossless, which greatly reduces the threshold for model deployment and operation, making it easier to apply in edge devices and real-time scenarios. At the same time, the compressed model greatly reduces the demand for computing power and storage, thereby saving computing resources and costs.

[0131] According to the automated iteratively optimized content-to-storyboard description generation method of an embodiment of the present invention, a first summary text of a preset text, a second summary text of each chapter, and a third summary text of each paragraph can be obtained respectively, and then the preset chapter can be processed based on the above summary text to obtain an initial storyboard description text of the preset chapter. Based on the summary texts of the full text, chapters, and paragraphs, a storyboard description generation model can refer to the contextual content of the full text, and then the preset chapter is converted into an initial storyboard description text, thereby reducing the probability that the initial storyboard description text will have problems such as generalizing from a particular part to the whole, inconsistency between the context, misunderstanding, and poor coherence. The initial storyboard description text can be evaluated for quality, problems existing in the initial storyboard description text can be found, and corrections can be made based on the problems existing therein, and the processing can be iteratively performed until the quality of the corrected storyboard description text meets the requirements, thereby obtaining a high-quality storyboard description text, thereby further reducing the probability that the initial storyboard description text will have problems such as generalizing from a particular part to the whole, inconsistency between the context, misunderstanding, and poor coherence. When training a natural language processing model, the loss function of the natural language processing model can be set by the semantic similarity and distortion between the first pending summary text and the combined text of the full text summary, and the semantic similarity and distortion between the second pending summary text and the combined text of the chapter summary, and the natural language processing model can be iteratively updated and trained to improve the accuracy of the natural language processing model, thereby obtaining accurate summary texts at multiple levels, so that when subsequently generating storyboard descriptions, accurate summary texts at multiple levels can be used, so that the storyboard description generation model can grasp richer context information as much as possible when it cannot receive the input of long texts, so that the storyboard description is closer to the semantics of the original text, and the probability of generalizing, inconsistency, poor coherence, etc. can be reduced. When determining the accuracy score of the initial storyboard description text, the accuracy of the description of the characters, places and actions in the initial storyboard description text can be solved by comparing the preset chapter summary text with the initial storyboard description text, so as to obtain the accuracy score of the initial storyboard description text, which can objectively and accurately describe the accuracy of the description of the important plots in the initial storyboard description text. When determining the original meaning retention score, the original meaning retention score can be determined by factors such as the semantic similarity between the initial storyboard description text and the original text of the preset chapter, and whether important plots such as characters, places, actions in the initial storyboard description text appear in the original text of the preset chapter, and whether they appear in the second summary text of other chapters. This can accurately and objectively indicate whether the initial storyboard description text can maintain the original meaning of the original text of the preset chapter, and whether there is a defect of important plot spoilers.When determining the semantic error detection score, the semantic accuracy of each sentence in the initial storyboard description text can be determined by comparing the sentences in the initial storyboard description text with the sentences in the original text of the preset chapters sentence by sentence, and whether the important plots such as characters, places, actions in the initial storyboard description text appear in the full text to determine whether there are erroneous plots in the initial storyboard description text generated by the storyboard description generation model based on incorrect understanding, thereby improving the accuracy and objectivity of the semantic error detection score. When reviewing the problems existing in the initial storyboard description text, the problems existing in the initial storyboard description text can be reviewed by the semantic and logical similarity between the initial storyboard description text and the original text, the completeness of the expression of the original text, and the similarity with the original text in emotional expression, so as to obtain accurate problem description information, and provide a data basis for the correction training of the storyboard description generation model and the correction of the initial storyboard description text. When training the storyboard description generation model, the loss function for the first correction training can be determined based on aspects such as the overall distortion of the original text by the initial storyboard description text, the semantic error between the initial storyboard description text and the original text, and the overall difference in emotional expression between the initial storyboard description text and the preset chapters. The weights of the above aspects can be set through the problem description information, so as to focus on the aspects with serious problems and improve the pertinence and efficiency of the training.

[0132] Figure 2 A block diagram of a content transition storyboard description generation system for automated iterative optimization according to an embodiment of the present invention is exemplarily shown, wherein the system comprises:

[0133] A summary text module, used to process a preset text through a natural language processing model to obtain a first summary text of the preset text, a second summary text of each chapter of the preset text, and a third summary text of each paragraph of the chapter;

[0134] An initial storyboard description text module is used to process the first summary text, the second summary text, the third summary text and the preset chapters in the preset text through a pre-trained storyboard description generation model to obtain an initial storyboard description text corresponding to the preset chapters;

[0135] A content quality scoring module, configured to evaluate the content quality of the initial storyboard description text according to the preset chapter, the first summary text, the second summary text and the third summary text, and obtain a content quality score of the initial storyboard description text;

[0136] A problem description information module is used to review the initial storyboard description text according to the content quality score to obtain problem description information of the initial storyboard description text;

[0137] A correction training module, used for performing a first correction training on the storyboard description generation model according to the problem description information of the initial storyboard description text, so as to obtain a first corrected storyboard description generation model;

[0138] A scoring module, used to obtain the F-th revised storyboard description text through the F-th revised storyboard description generation model, and obtain a content quality score of the F-th revised storyboard description text, where F≥1, and F is a positive integer;

[0139] An output module, configured to use the storyboard description text revised for the Fth time as storyboard description output content if the content quality score of the storyboard description text revised for the Fth time is higher than or equal to a preset quality score;

[0140] an iterative module, for otherwise obtaining problem description information of the storyboard description text after the Fth revision according to the content quality score of the storyboard description text after the Fth revision, and performing the F+1th revision training on the storyboard description generation model after the Fth revision to obtain the storyboard description generation model after the F+1th revision, and obtaining the storyboard description text after the F+1th revision;

[0141] Iteration completion module, used to complete the F+1= , the storyboard description text after the F+1th correction is used as the storyboard description output content, where, is the preset number of iterations.

[0142] It should be understood by those skilled in the art that the embodiments of the present invention described above and shown in the accompanying drawings are only examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functional and structural principles of the present invention have been demonstrated and explained in the embodiments, and the embodiments of the present invention may be deformed or modified in any way without departing from the principles.

[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automated iteratively optimized method for generating content-to-storyboard descriptions, characterized in that: include: Processing a preset text by a natural language processing model to obtain a first summary text of the preset text, a second summary text of each chapter of the preset text, and a third summary text of each paragraph of the chapter; The first summary text, the second summary text, the third summary text, and a preset chapter in the preset text are processed by a pre-trained storyboard description generation model to obtain an initial storyboard description text corresponding to the preset chapter; According to the preset chapter, the first summary text, the second summary text and the third summary text, the content quality of the initial storyboard description text is evaluated to obtain a content quality score of the initial storyboard description text; Reviewing the initial storyboard description text according to the content quality score to obtain problem description information of the initial storyboard description text; According to the problem description information of the initial storyboard description text, the storyboard description generation model is subjected to a first correction training to obtain a first corrected storyboard description generation model; Obtain the F-th revised storyboard description text through the F-th revised storyboard description generation model, and obtain the content quality score of the F-th revised storyboard description text, where F≥1, and F is a positive integer; If the content quality score of the storyboard description text after the F-th revision is higher than or equal to the preset quality score, the storyboard description text after the F-th revision is used as the storyboard description output content; Otherwise, obtaining the problem description information of the storyboard description text after the Fth revision according to the content quality score of the storyboard description text after the Fth revision, and performing the F+1th revision training on the storyboard description generation model after the Fth revision to obtain the F+1th revised storyboard description generation model and obtain the F+1th revised storyboard description text; In F+1= , the storyboard description text after the F+1th correction is used as the storyboard description output content, where, is the preset number of iterations.

2. The method for generating content-to-storyboard descriptions by automated iterative optimization according to claim 1, characterized in that: The preset text is processed by a natural language processing model to obtain a first summary text of the preset text, a second summary text of each chapter of the preset text, and a third summary text of each paragraph of the chapter, including: Process each paragraph using the natural language processing model to obtain a third to-be-determined summary text of each paragraph; Processing the chapter and the third to-be-determined summary text of each paragraph of the chapter by the natural language processing model to obtain a second to-be-determined summary text of the chapter; Processing the second to-be-determined summary text of each chapter by the natural language processing model to obtain the first to-be-determined summary text of the preset text; Processing the second undetermined summary text of the chapter by the natural language processing model to obtain a second undetermined semantic feature vector of the second undetermined summary text; Combining the third undetermined summary texts of the respective paragraphs of the chapter to obtain a chapter summary combination text, and processing the chapter summary combination text by the natural language processing model to obtain a chapter combination semantic feature vector; Processing the first undetermined summary text of the preset text by the natural language processing model to obtain a first undetermined semantic feature vector of the first undetermined summary text; Combining the third undetermined summary texts of each paragraph of each chapter of the preset text to obtain a full-text summary combination text, and processing the full-text summary combination text through the natural language processing model to obtain a full-text combination semantic feature vector; Determining a loss function of the natural language processing model according to the second undetermined semantic feature vector, the chapter combination semantic feature vector, the first undetermined semantic feature vector, and the full-text combination semantic feature vector; Performing update training on the natural language processing model through the loss function of the natural language processing model to obtain an updated natural language processing model; Iteratively executing the step of updating the trained natural language processing model until the loss function of the natural language processing model is less than or equal to a first preset threshold, thereby obtaining a trained natural language processing model; Process each paragraph through the trained natural language processing model to obtain the third summary text of each paragraph; Processing the chapter and the third summary text of each paragraph of the chapter by the trained natural language processing model to obtain a second summary text of the chapter; The second summary text of each chapter is processed by the trained natural language processing model to obtain the first summary text of the preset text.

3. The method for generating content-to-storyboard descriptions by automated iterative optimization according to claim 2, characterized in that: Determining the loss function of the natural language processing model according to the second undetermined semantic feature vector, the chapter combination semantic feature vector, the first undetermined semantic feature vector, and the full-text combination semantic feature vector includes: According to the formula Determining the loss function for your natural language processing model ,in, is the second undetermined semantic feature vector of the i-th chapter, The chapter combination semantic feature vector of the i-th chapter, for The transposed vector of for and The cosine similarity of is the first undetermined semantic feature vector, is the combined semantic feature vector of the full text, for The transposed vector of for and The cosine similarity of and is the preset weight, n is the number of chapters in the preset text, i≤n, and i and n are both positive integers.

4. The method for generating content-to-storyboard descriptions by automated iterative optimization according to claim 1, characterized in that: According to the preset chapter, the first summary text, the second summary text and the third summary text, the content quality of the initial storyboard description text is evaluated to obtain a content quality score of the initial storyboard description text, including: Determining an accuracy score of the initial storyboard description text according to the initial storyboard description text, the third summary text, and the second summary text; Determining a key information coverage score of the initial storyboard description text according to the second summary text and the initial storyboard description text; Determining an information completeness score according to the initial storyboard description text and the preset chapters; Determining a degree of originality preservation score according to the preset chapter, the second summary text, and the initial storyboard description text; Determining a semantic error detection score according to the preset chapter, the initial storyboard description text, the first summary text, and the second summary text; A content quality score of the initial storyboard description text is determined based on the accuracy score, the key information coverage score, the information completeness score, the original meaning retention score, and the semantic error detection score.

5. The method for generating content-to-storyboard descriptions by automated iterative optimization according to claim 4, characterized in that: Determining an accuracy score of the initial storyboard description text according to the initial storyboard description text, the third summary text, and the second summary text, comprising: Using the natural language processing model, searching for a first character description text, a first location description text, and a first action description text in the initial storyboard description text, and determining a first character semantic vector of the first character description text, a first location semantic vector of the first location description text, and a first action semantic vector of the first action description text; Combining the second summary text of the preset chapter and the third summary texts of each paragraph of the preset chapter to obtain a preset chapter summary text; Using the natural language processing model, searching for a second character description text, a second location description text, and a second action description text of the preset chapter summary text, and determining a second character semantic vector of the second character description text, a second location semantic vector of the second location description text, and a second action semantic vector of the second action description text; According to the formula Determine the accuracy score of the initial storyboard description text ,in, is the second person semantic vector of the jth second person description text, is the first person semantic vector of the kth first person description text, The number of texts describing the first character. The number of description texts for the second character, for The transposed vector of for and The cosine similarity of is the second action semantic vector of the t-th second action description text, For the The first action semantic vector of the first action description text, for The transposed vector of for and The cosine similarity of is the number of the first action description text, The number of description texts for the second action, is the second location semantic vector of the x-th second location description text, is the first place semantic vector of the yth first place description text, for The transposed vector of for and The cosine similarity of is the number of texts describing the first location, is the number of description texts of the second location, max is the maximum value function, , and are preset weights, k, j, ,t,x,y, , , , , and All are positive integers.

6. The method for generating content-to-storyboard descriptions by automated iterative optimization according to claim 4, characterized in that: Determining the original meaning preservation score according to the preset chapter, the second summary text and the initial storyboard description text includes: Using the natural language processing model, searching for a first character description text, a first location description text, and a first action description text in the initial storyboard description text, and determining a first character semantic vector of the first character description text, a first location semantic vector of the first location description text, and a first action semantic vector of the first action description text; Determining initial storyboard description semantic information of the initial storyboard description text by using the natural language processing model; Determining chapter semantic information of the preset chapter through the natural language processing model; Using the natural language processing model, searching for a third character description text, a third place description text, and a third action description text in the second summary text of other chapters except the preset chapter, and determining a third character semantic vector of the third character description text, a third place semantic vector of the third place description text, and a third action semantic vector of the third action description text; Searching, by means of the natural language processing model, for a fourth character description text, a fourth location description text, and a fourth action description text in the preset chapter, and determining a fourth character semantic vector of the fourth character description text, a fourth location semantic vector of the fourth location description text, and a fourth action semantic vector of the fourth action description text; According to the formula Determine the original intention retention score ,in, Describe semantic information for the initial storyboard, is the chapter semantic information of the preset chapter, is the first person semantic vector of the kth first person description text, is the fourth person semantic vector of the jth fourth person description text, is the third person semantic vector of the tth third person description text, The number of description texts for the fourth character, The number of description texts for the third person, The number of texts describing the first character. For the The first action semantic vector of the first action description text, is the fourth action semantic vector of the tth fourth action description text, is the third action semantic vector of the yth third action description text, is the number of the first action description text, The number of description texts for the fourth action, The number of description texts for the third action, is the first place semantic vector of the zth first place description text, is the fourth place semantic vector of the u-th fourth place description text, is the third place semantic vector of the vth third place description text, is the number of texts describing the first location, The number of texts describing the third location, The number of texts describing the fourth location, is the first preset similarity, is the second preset similarity, if is the conditional function, max is the maximum value function, , , and are preset weights, k, j, t, , x, y, z, u, v, , , , , , , , and Is a positive integer.

7. The method for generating content-to-storyboard descriptions by automated iterative optimization according to claim 4, characterized in that: Determining a semantic error detection score according to the preset chapter, the initial storyboard description text, the first summary text, and the second summary text includes: Using the natural language processing model, searching for a first character description text, a first location description text, and a first action description text in the initial storyboard description text, and determining a first character semantic vector of the first character description text, a first location semantic vector of the first location description text, and a first action semantic vector of the first action description text; Determining the first sentence semantic information of each sentence in the initial storyboard description text by using the natural language processing model; Determining the second sentence semantic information of each sentence in the preset chapter by the natural language processing model; Combining the first summary text and the second summary text to obtain a full-text summary combined text; Searching, by means of the natural language processing model, for a fourth character description text, a fourth location description text, and a fourth action description text in the preset chapter, and determining a fourth character semantic vector of the fourth character description text, a fourth location semantic vector of the fourth location description text, and a fourth action semantic vector of the fourth action description text; Searching, by means of the natural language processing model, a fifth character description text, a fifth place description text and a fifth action description text in the full-text summary combined text, and determining a fifth character semantic vector of the fifth character description text, a fifth place semantic vector of the fifth place description text and a fifth action semantic vector of the fifth action description text; According to the formula Determining the semantic error detection score ,in, is the semantic information of the first sentence of the hth sentence in the initial storyboard description text, is the semantic information of the second sentence of the gth sentence in the preset chapter, is the number of sentences in the preset chapter, is the number of sentences in the initial storyboard description text, is the first person semantic vector of the kth first person description text, is the fourth person semantic vector of the jth fourth person description text, is the fifth character semantic vector of the tth fifth character description text, The number of description texts for the fourth character, The number of texts describing the fifth character, The number of texts describing the first character. For the The first action semantic vector of the first action description text, is the fourth action semantic vector of the tth fourth action description text, is the fifth action semantic vector of the yth fifth action description text, is the number of the first action description text, The number of description texts for the fourth action, The number of description texts for the fifth action, is the first place semantic vector of the zth first place description text, is the fourth place semantic vector of the u-th fourth place description text, is the fifth place semantic vector of the vth fifth place description text, is the number of texts describing the first location, The number of texts describing the fifth location, The number of texts describing the fourth location, is the first preset similarity, if is the conditional function, max is the maximum value function, , , and are preset weights, k, j, t, , x, y, z, u, v, , , , , , , , and Is a positive integer.

8. The method for generating content-to-storyboard descriptions by automated iterative optimization according to claim 1, characterized in that: According to the content quality score, the initial storyboard description text is reviewed to obtain problem description information of the initial storyboard description text, including: If the content quality score is lower than a preset quality score threshold, determining the first sentence semantic information of each sentence in the initial storyboard description text through the natural language processing model; Determining the second sentence semantic information of each sentence in the preset chapter by the natural language processing model; Determining, according to the first sentence semantic information and the second sentence semantic information, a first sentence that is semantically consistent with the sentences in the preset chapter and a second sentence that is semantically inconsistent with the sentences in the preset chapter among the multiple sentences in the initial storyboard description text; Obtaining, by means of a pre-trained sentence sentiment analysis model, a first sentiment expression vector of the first sentence and a second sentiment expression vector of a first target sentence in the preset chapter that has the highest semantic similarity to the first sentence; Determine a second target sentence other than the first target sentence in the preset chapter, and obtain second target semantic information of the second target sentence through the natural language processing model; According to the formula Determine logical review score ,in, is the number of the first sentence, is the number of the second sentence, and is a positive integer; According to the formula Determining the Expression Completeness Review Score ,in, is the number of the first target sentences, is the number of the second target sentences, and is a positive integer; According to the formula Determine the emotional expression review score ,in, is the first sentiment expression vector of the first sentence of the ath sentence, is the second sentiment expression vector of the first target sentence corresponding to the ath first sentence, a is a positive integer, and a≤ ; According to the logic review score, the expression integrity review score and the emotional expression review score, the problem description information of the initial storyboard description text is determined.

9. The method for generating content-to-storyboard descriptions by automated iterative optimization according to claim 8, characterized in that: According to the problem description information of the initial storyboard description text, the storyboard description generation model is subjected to a first correction training to obtain a first corrected storyboard description generation model, including: According to the formula Determine the loss function for the first round of training of the storyboard description generation model ,in, is the semantic information of the first sentence of the rth sentence in the initial storyboard description text, is the second sentence semantic information of the sentence having the highest semantic similarity with the first sentence semantic information of the rth sentence of the initial storyboard description text among the plurality of sentences in the preset chapter, is the first sentiment expression vector of the rth sentence in the initial storyboard description text, is the second emotion expression vector of the sentence with the highest semantic similarity to the semantic information of the first sentence of the rth sentence of the initial storyboard description text among the multiple sentences of the preset chapter, is the second target semantic information of the qth second target sentence, r≤ , q≤ , and r and q are positive integers; The storyboard description generation model is trained for the first time using the loss function during the first round of revised training to obtain the storyboard description generation model after the first round of revision.

10. An automated iteratively optimized content-to-storyboard description generation system, characterized in that: include: A summary text module, used to process a preset text through a natural language processing model to obtain a first summary text of the preset text, a second summary text of each chapter of the preset text, and a third summary text of each paragraph of the chapter; An initial storyboard description text module is used to process the first summary text, the second summary text, the third summary text and the preset chapters in the preset text through a pre-trained storyboard description generation model to obtain an initial storyboard description text corresponding to the preset chapters; A content quality scoring module, configured to evaluate the content quality of the initial storyboard description text according to the preset chapter, the first summary text, the second summary text and the third summary text, and obtain a content quality score of the initial storyboard description text; A problem description information module is used to review the initial storyboard description text according to the content quality score to obtain problem description information of the initial storyboard description text; A correction training module, used for performing a first correction training on the storyboard description generation model according to the problem description information of the initial storyboard description text, so as to obtain a first corrected storyboard description generation model; A scoring module, used to obtain the F-th revised storyboard description text through the F-th revised storyboard description generation model, and obtain a content quality score of the F-th revised storyboard description text, where F≥1, and F is a positive integer; An output module, configured to use the storyboard description text revised for the Fth time as storyboard description output content if the content quality score of the storyboard description text revised for the Fth time is higher than or equal to a preset quality score; an iterative module, for otherwise obtaining problem description information of the storyboard description text after the Fth revision according to the content quality score of the storyboard description text after the Fth revision, and performing the F+1th revision training on the storyboard description generation model after the Fth revision to obtain the storyboard description generation model after the F+1th revision, and obtaining the storyboard description text after the F+1th revision; Iteration completion module, used to complete the F+1= , the storyboard description text after the F+1th correction is used as the storyboard description output content, where, is the preset number of iterations.

Images