AI-based script creation system and method

Abstract

The invention discloses an AI-based script creation system and method, and the method comprises the steps: receiving an initial cue word, and carrying out the recognition and extraction of script core elements of the initial cue word; inputting the script core elements into a pre-trained script generation model, and generating an initial script by the script generation model; performing text similarity, plot structure similarity and character relationship similarity calculation on the initial script and copyright scripts in a script database, and outputting the initial script only when the similarities are smaller than a preset percentage; according to the method, the overall creation period is shortened, the cost is reduced, the integrity of the script and the copyright security are both considered, and the copyright risk of the script is reduced while the script creation efficiency and quality are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a script creation system and method based on AI. BACKGROUND

[0002] The traditional script creation highly depends on the experience of professional screenwriters, and has the problems of complex process, long cycle and high cost.

[0003] Some existing AI script tools can only generate text fragments, and have the defects of incomplete generated content structure, lack of multi-character and multi-scene interaction ability, and difficulty in direct application to video production process.

[0004] At the same time, the generated script may have substantial similarity to certain copyrighted works (such as highly similar plot structure and character setting), which may lead to user liability for using the script.

[0005] Therefore, there is an urgent need for a new script generation scheme to improve script creation efficiency and quality while reducing the copyright risk of the script. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a script creation system and method based on AI, which can improve script creation efficiency and quality while reducing the copyright risk of the script.

[0007] In order to solve the above technical problems, the technical scheme adopted by the present application is: A script creation system based on AI, comprising: A preprocessing module: receiving an initial prompt word and identifying and extracting script core elements from the initial prompt word; A script generation module: inputting the script core elements into a pre-trained script generation model to generate an initial script from the script generation model; The script generation model adopts a large language model, which is trained based on a public script with no copyright effect and a non-script field text: Obtain a public script with no copyright effect and a non-script field text; Obtain pre-labeled transferable elements in the non-script field text, and generate corresponding script expressions based on an NLP model; Construct a mixed corpus according to the public script and the script expressions of the transferable elements, and unsupervisedly pre-train the large language model; Supervise the fine-tuning of the large language model using the public script to obtain the script generation model; A copyright verification module: calculating the text similarity, plot structure similarity and character relationship similarity between the initial script and the copyrighted script in a script database; The initial script is output only when all similarities are less than a preset percentage; otherwise, the script generation module is returned to regenerate the script.

[0008] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows: An AI-based scriptwriting method, comprising the following steps: S1. Receive the initial prompt words and identify and extract the core elements of the script from the initial prompt words; S2. Input the core elements of the script into the pre-trained script generation model, and the script generation model generates an initial script; The script generation model employs a large language model, trained on a mixture of publicly available scripts without copyright validity and non-script domain text. Obtain publicly available scripts without copyright protection, as well as texts outside the script field; Extract pre-annotated transferable elements from non-script domain texts and generate corresponding script-like expressions based on NLP models; A hybrid corpus is constructed based on the script-like expressions of the publicly available scripts and the transferable elements, and the large language model is pre-trained in unsupervised manner. The script generation model is obtained by supervising the large language model using the publicly available scripts. S3. Calculate the text similarity, plot structure similarity, and character relationship similarity between the initial script and the copyrighted scripts in the script database; S4. Output the initial script only when all similarities are less than the preset percentage; otherwise, return to step S2.

[0009] The beneficial effects of this invention are as follows: This invention provides an AI-based script creation system and method that automatically generates scripts by extracting core elements from prompts, replacing manual initial conception and reducing reliance on professional expertise. The script generation model employs a large language model trained on a mixture of copyright-free publicly available scripts and non-script texts. This mixed corpus covers multi-scene interactions and plot logic, generating a structurally complete initial script (including scenes, dialogue, and actions), thus addressing the shortcomings of existing AI-generated scripts that are fragmented and lack interaction. Furthermore, copyright similarity is verified from three dimensions: text, plot structure, and character relationships. Only scripts that meet the criteria are output, avoiding liability for substantial similarity. Overall, this shortens the creation cycle, reduces costs, and simultaneously ensures script integrity and copyright security, improving script creation efficiency and quality while reducing copyright risks. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an AI-based scriptwriting method according to an embodiment of the present invention. Figure 2 This is a schematic diagram of an AI-based scriptwriting system according to an embodiment of the present invention; Label Explanation: 1. An AI-based script creation system; 2. Preprocessing module; 3. Script generation module; 4. Copyright verification module. Detailed Implementation

[0011] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0012] The AI-based script creation system and method described above are applicable to script generation, and are further illustrated below through specific implementation methods: Please refer to Figure 1 One embodiment of the present invention is as follows: An AI-based scriptwriting method, comprising the following steps: S1. Receive the initial prompt words and identify and extract the core elements of the script from the initial prompt words.

[0013] In this embodiment, a two-stage NLP model of "intent recognition + entity extraction" (the base model is BERT-Large, and the fine-tuning dataset is the script creation prompt word annotation library, which contains more than 100,000 manually annotated "prompt word-core element" samples) is used to identify and extract the core elements of the script.

[0014] The core elements of a script include the following dimensions: Theme and Direction: Clearly define the core theme of the script (e.g., "campus cooperation" or "workplace teamwork"). Role Framework: Initially define the core roles (number, identity, core objectives, such as "2 high school students: Role A (club leader, objective: to salvage a failed event), Role B (technical backbone, objective: to prove the value of the solution)"). Scope of scenarios: Identify key scenarios (such as "campus library, club activity room, playground"); Core Conflict: Identify the focus of the conflict (e.g., "A club activity is on the verge of cancellation due to a procedural error, and the two main characters need to replan within two days"). Emotional tone: Determine the style of the script (e.g., "youthful inspiration, light comedy, heartwarming and healing").

[0015] S2. Input the above-mentioned core elements of the script into the pre-trained script generation model, and the script generation model generates the initial script.

[0016] In this embodiment, a specially trained large language model transforms the core elements extracted by S1 into a standardized, structurally complete initial script containing multi-character and multi-scene interactions, replacing the traditional screenwriter's "from scratch" conception stage and shortening the creation cycle. This enables the model to generate a complete script containing "scene titles, character lists, dialogue, action cues, and plot node annotations."

[0017] The script generation model employs a large language model, trained on a mixture of publicly available scripts without copyright validity and non-script domain text. Obtain publicly available scripts without copyright protection, as well as texts outside the scriptwriting field.

[0018] In this embodiment, the acquisition of publicly available scripts without copyright validity includes, but is not limited to, the following two types of script resources: ① classic scripts that have entered the public domain (such as play scripts before 1923, and officially released public service scripts); ② non-commercial scripts authorized by users (obtained through platform agreements, clearly indicating no copyright disputes), totaling more than 100,000 scripts in the basic library.

[0019] Non-script texts include, but are not limited to, copyright-free novels (such as old-school online novels and public domain literary works), everyday dialogue data (publicly available campus / workplace dialogue datasets), and news event descriptions (copyright-free press releases).

[0020] Extract pre-annotated transferable elements from non-script domain texts and generate corresponding script-like expressions based on an NLP model.

[0021] In this embodiment, for non-script text, a hybrid model of "BERT+Seq2Seq" is used to extract elements that can be transformed into script interactions from non-script text, including: character interaction type (such as "comfort, argument, collaboration"), scene details (such as "coffee shop: window seat, laptop on the table"), and emotion-related actions (such as "angry: clenching fists, raising voice").

[0022] Convert the extracted elements into a standard script format, example: The original novel text reads: "Lin Xiao saw Chen Yang frowning at his computer in the coffee shop, walked over and patted him on the shoulder." The converted script excerpt reads: "[Scene: Campus Café - Afternoon] Lin Xiao (carrying a coffee, walking towards a window seat): Are you struggling with the program again? (pats Chen Yang's shoulder lightly) Chen Yang (looks up, rubbing his temples): The live streaming system keeps crashing, tomorrow's exhibition is going to be a disaster..."

[0023] A hybrid corpus is constructed based on the publicly available scripts and the script-like expressions of the transferable elements, and the large language model is then subjected to unsupervised pre-training.

[0024] In this embodiment, the "uncopyrighted public script" and the "transformed script expression" are mixed in a ratio of 7:3, with a total scale of more than 500,000 script fragments, ensuring that the model covers three core capabilities: "multi-character interaction", "multi-scene switching", and "plot logic".

[0025] In this embodiment, the large language model used is the Llama2-70B model. Taking it as the base model, unsupervised training is carried out for 1 million steps with a mixed corpus, aiming to let the model learn: ① script format specifications (writing of scene annotations, character names + lines, and action prompts); ② plot logic coherence (such as the temporal logic of "conflict occurs -> character attempts to solve -> a turn occurs -> conflict is resolved"); ③ multi-character dialogue styles (tone differences of characters with different identities, such as the colloquialism of students and the formality of teachers).

[0026] The large language model is fine-tuned with supervision using the said public script to obtain the said script generation model.

[0027] In this embodiment, a "core element - standard script" annotation dataset (more than 10,000 samples, annotated by professional screenwriters) is constructed. Using the core elements extracted in S1 as the input and the standard script as the output, the pre-trained LLM is fine-tuned to optimize the conversion accuracy of "element -> script". For example: when the input is "core conflict: equipment failure in the club exhibition", the model needs to generate a complete plot chain of "the character discovers the failure -> discusses solutions -> divides the work and executes", rather than scattered fragments.

[0028] S3. Calculate the text similarity, plot structure similarity, and character relationship similarity between the said initial script and the copyrighted scripts in the script database; Step S3 includes: S31. Calculate the overlap degree of the texts of the said initial script and each copyrighted script through cosine similarity to obtain the text similarity.

[0029] Step S31 includes the steps: S311. Based on the NLP-based text structured extraction technology, disassemble the said initial script and the said copyrighted script into three types of sub-texts: line segments, scene descriptions, and action prompts.

[0030] In this embodiment, the calculation of text similarity includes: Text preprocessing: Both the initial script and the copyrighted script (from the public copyright database, such as the film and television script copyright registration library) are disassembled into three types of sub-texts: "line segments, scene descriptions, action prompts", and meaningless words (such as "de, le") are removed.

[0031] S312. Use the Sentence-BERT model to encode each type of sub-text to obtain text vectors.

[0032] In this embodiment, vector encoding is performed using the Sentence-BERT model, which encodes each type of subtext into a 768-dimensional vector (ensuring semantic representation rather than literal matching).

[0033] S313. For the text vectors of the same type corresponding to the initial script and the copyright script, calculate the cosine similarity respectively, and perform a weighted calculation on the cosine similarity of the three types of sub-texts to obtain the text similarity.

[0034] In this embodiment, the similarity calculation is as follows: the cosine similarity of the three types of sub-text vectors of the initial script and the single copyrighted script is calculated respectively, and then weighted according to the weight of "dialogue 40% + scene 30% + action 30%" to obtain the final text similarity (range 0-100%).

[0035] Traditional text similarity calculation methods are prone to misjudgment due to "overall text similarity but identical key sub-texts (such as dialogue)," failing to accurately reflect copyright risks. This embodiment's text similarity calculation process first decomposes the initial script and the copyrighted script into three categories of sub-texts: dialogue, scene descriptions, and action cues. It then selectively extracts copyright-sensitive parts (e.g., dialogue is the core expression of the script, while scenes and actions are key to style). Next, it uses Sentence-BERT encoding to achieve semantic-level matching rather than literal matching, avoiding missed judgments due to "synonym substitution with semantic similarity." Finally, it weights the sub-texts according to their importance (dialogue 40%, scene 30%, action 30%), making the similarity results more aligned with copyright determination logic. This solves the problems of insufficient detail and unbalanced weighting in text similarity calculations, improving the accuracy of text-level copyright verification.

[0036] S32. Construct the plot graph of the initial script, and calculate the similarity between the plot graph of the initial script and the plot graph of each copyrighted script through a pre-trained graph neural network to obtain the plot structure similarity.

[0037] (1) Plot graph construction: Nodes: Use the ACE event extraction model to extract key events from the initial script (e.g., "Event 1: Li Ran discovers a data error (Character: Li Ran, Time: Morning); Event 2: Zhao Kai denies the error (Character: Zhao Kai, Time: Morning)"). Edges: Mark the logical relationship between events (cause and effect: error discovery -> negation; sequence: event 1 -> event 2); (2) Calculation of map similarity: Pre-trained Graph Attention Network (GAT): Trained with plot graphs of over 10,000 copyrighted scripts to learn plot structure representations.

[0038] GAT (Graph Attention Network) is an AI model specifically designed to process "graph data"—it can "focus on key connections in the graph" just like a human: for example, in a plot graph, the "causal relationship between 'discovering an error' and 'being refuted'" is more important than the "temporal relationship between two unrelated events." GAT will prioritize this core connection through "attention weights," capturing the logical structure of the plot more accurately than ordinary graph models.

[0039] Using plot graphs from a large number of copyrighted scripts as training data, these graphs are pre-labeled with "event nodes (e.g., 'Li Ran discovers a data error')" and "logical edges (e.g., 'cause and effect: error discovery -> denial')". GAT learns from this data the "general structural patterns of script plots": for example, "conflict-based scripts often follow the chronological logic of 'discovering a problem -> attempting to solve it -> a turning point -> solving the problem'," and "in workplace scripts, 'subordinate reporting -> leader making a decision' is a common causal relationship." After training, GAT can quickly identify "which events are core nodes" and "which logical edges are key relationships" in the plot graph of any new script, laying the foundation for subsequent comparisons.

[0040] Input the plot graphs of the initial script and the copyrighted script, calculate "node similarity (event content matching)" and "edge similarity (logical relationship matching)", and weight them to obtain the plot structure similarity (range 0-100%).

[0041] In this embodiment, the ACE event extraction model is first used to extract event nodes and label logical relationships (causality, temporal sequence), transforming the abstract plot into a structured graph, thus solving the problem of plot being difficult to visualize and quantify. Then, a pre-trained graph attention network (GAT) is used to calculate similarity. GAT can prioritize core logical relationships (such as "discovering a problem -> solving a problem"), and captures the key framework of the plot more accurately than ordinary graph models, avoiding misjudgments caused by interference from secondary events.

[0042] S33. Construct the character interaction matrix of the initial script, and perform matrix similarity calculation with the character interaction matrix of each copyrighted script to obtain the character relationship similarity.

[0043] In this embodiment, the unstructured character relationships in the script are transformed into a structured matrix through the NLP technology chain (role recognition -> interaction extraction -> quantitative statistics), realizing the digital mapping of "role-interaction-frequency".

[0044] Character interaction matrix construction: Example as follows: Rows / columns: Core roles (Li Ran, Zhao Kai, General Manager Wang); Elements: Interaction type + frequency (e.g., "Li Ran & Zhao Kai: Questioning twice, persuading once; Zhao Kai & General Manager Wang: Reporting once"), Example matrix: Li Ran Zhao Kai Wang Zong Li Ran - Question (2), Persuade (1) - Zhao Kai Deny (2), Recognize (1) - Report (1) Wang Zong - Receive (1) - Matrix similarity calculation: The difference between matrix elements (interaction type + frequency) is calculated using the Manhattan distance formula and converted into a similarity score (range 0-100%).

[0045] In this embodiment, an NLP model is first used to identify the types of character interactions (questioning, persuasion, etc.) and quantify their frequency, constructing a "character-interaction-frequency" character interaction matrix to transform abstract character relationships into structured data. Then, Manhattan distance is used to calculate the differences between matrix elements, accurately comparing interaction details (e.g., the difference between "questioning twice, persuading once" and "questioning once, persuading twice"). This solves the problems of difficulty in quantifying character relationships and subjective similarity judgments, accurately identifying copyright risks at the interaction pattern level, and further reducing the possibility of liability due to similar character relationships.

[0046] Text similarity detection checks for literal and semantic overlap, avoiding identical dialogue and scene descriptions; plot structure similarity uses a plot graph comparison framework to prevent similarities in core plot chains such as "conflict-resolution"; character relationship similarity focuses on character interactions to avoid similar character relationship patterns (such as "questioning-cooperation"). These three dimensions cover the main risks of copyright infringement, providing a more comprehensive and accurate assessment than single-text verification, further reducing the risk of users being held liable for similarities in script structure and character relationships.

[0047] S4. Output the initial script only when all similarities are less than the preset percentage; otherwise, return to step S2.

[0048] In this embodiment, if the text similarity is <30%, the plot structure similarity is <40%, and the character relationship similarity is <40%, then the initial script is output. Furthermore, a copyright similarity detection report can be generated based on the values ​​of each similarity metric.

[0049] S41. When all similarities are less than a preset percentage, a semantic parsing model is used to extract the connecting elements information, including character design, storyboard production, and video generation, from the initial script and output them along with the initial script.

[0050] In this embodiment, after the script's copyright compliance is achieved, the semantic parsing model extracts the necessary connecting information for character design, storyboard production, and video generation. These elements can be directly integrated into subsequent video production stages (such as character design elements for modeling and storyboard elements for storyboard design), eliminating the need for manual secondary processing. This solves the problem of the disconnect between the script and video production processes, achieving a smooth connection between "script generation and video production," reducing time and labor costs in intermediate stages, improving overall creative efficiency, and enabling AI-generated scripts to directly empower video production.

[0051] The extraction of connecting elements in character design (connecting character modeling) includes: Based on the visual description extraction model, visual features are analyzed from the initial script to obtain information on the connecting elements of the character design.

[0052] In this embodiment, visual features are analyzed from the action cues and dialogue context of the script based on a visual description extraction model (such as BLIP-2). For example, the appearance features of "carrying a camera bag" are extracted from "Lin Xiao ran in carrying a camera bag".

[0053] By employing visual description extraction models (such as BLIP-2), visual features are directly analyzed from action cues in the script (such as "running in carrying a camera bag") and dialogue context, forming a clear basis for character design (such as "carrying a camera bag"). This solves the problem of lacking accurate textual basis for character modeling, avoids deviations between modeling and script content, makes character modeling more efficient and aligned with script settings, further strengthens the connection between the script and the pre-production stage of video production (character modeling), and reduces manual adjustment costs.

[0054] The extraction of transitional information (connecting storyboard designs) in storyboard production includes: The semantic segmentation model identifies objects and their positional relationships in the scene, and matches them with a pre-set camera angle library based on the script plot to obtain information on the connecting elements for storyboard production.

[0055] In this embodiment, a "scene analysis + shot language mapping" model is adopted. First, the semantic segmentation model identifies objects (bookshelves, long tables) and their positional relationships in the scene. Then, based on the script plot (dialogue / actions), a preset shot angle library is matched (e.g., "comforting action -> close-up, scene panorama -> overhead shot"). This solves the problem of low efficiency caused by reliance on manual interpretation in storyboard design, making the storyboard more consistent with the script scene and plot, and shortening the storyboard design cycle in the early stages of video production.

[0056] The extraction of transition element information (from the video generation tool) includes: Based on the time series annotation model, the duration is calculated by combining the number of words in the dialogue and the complexity of the actions. The visual parameter library is then used to match the emotion tags to obtain the connection element information for video generation.

[0057] In this embodiment, based on a time-series annotation model, and combined with the number of words in the dialogue and the complexity of the actions to calculate the duration, action instructions are generated (e.g., "Lin Xiao pats Chen Yang's shoulder -> converted to [Action: lightly pats shoulder with palm, force: light, duration: 1 second]"), as well as dialogue duration annotations (e.g., "Chen Yang's line 'Try it' -> duration: 2 seconds, used to match lip-sync animation"). Visual effect parameters are generated by combining emotion tag matching with a visual parameter library (e.g., "Character's dejected scene -> color tone: cool (RGB: 200, 220, 255), brightness: reduced by 10%). This solves the problem of unfounded video generation parameter settings and disconnect from the script, ensuring that the duration and visual effects of the generated video accurately match the script content, reducing manual adjustments, achieving seamless integration from script to video generation, and improving video production efficiency.

[0058] In this embodiment, if the text similarity is <30%, the plot structure similarity is <40%, and the character relationship similarity is <40%, an initial script, a connection element table, and a copyright detection report are output, which users can directly use for video production.

[0059] If any type of similarity exceeds the limit: return to step S2, triggering the model's differential iteration—the model automatically adjusts the content of high-risk dimensions: Text exceeding limits: Replace identical lines (e.g., change "data error" to "data logic deviation"); Exceeding the acceptable limits: Adjust the order of events (e.g., change "first denial -> then find evidence" to "first find evidence -> then communicate"). Character relationships exceeded limits: Modify interaction types (e.g., change "question" to "suggestion"); After each iteration, S2->S3 is executed again until all similarities are below the threshold.

[0060] Embodiment 2 of the present invention is as follows: According to another aspect of the invention, Figure 2 This is a schematic diagram illustrating an AI-based scriptwriting system according to an embodiment of the present invention.

[0061] An AI-based scriptwriting system 1, comprising: Preprocessing module 2: Receives initial prompts and identifies and extracts the core elements of the script from the initial prompts; Script generation module 3: Input the core elements of the script into the pre-trained script generation model, and the script generation model generates the initial script; The script generation model employs a large language model, trained on a mixture of publicly available scripts without copyright validity and non-script domain text. Obtain publicly available scripts without copyright protection, as well as texts outside the script field; Extract pre-annotated transferable elements from non-script domain texts and generate corresponding script-like expressions based on NLP models; A hybrid corpus is constructed based on the script-like expressions of the publicly available scripts and the transferable elements, and the large language model is pre-trained in unsupervised manner. The script generation model is obtained by supervising the large language model using the publicly available scripts. Copyright verification module 4: Calculates the text similarity, plot structure similarity, and character relationship similarity between the initial script and the copyrighted scripts in the script database; The initial script is output only when all similarities are less than a preset percentage; otherwise, the script generation module 3 is returned to regenerate the script.

[0062] It uses the preprocessing module 2, the script generation module 3, and the copyright verification module 4 to jointly implement the steps in the AI-based script creation method of the above embodiment 1.

[0063] This invention presents an AI-based scriptwriting system and method that systematically addresses the aforementioned pain points through a collaborative process encompassing preprocessing, generation, copyright verification, and linking element extraction. The preprocessing module extracts core elements such as themes and characters, replacing initial manual brainstorming and lowering the professional barrier. The script generation module trains an LLM (Limited Language Model) with copyright-free mixed corpora to generate a complete script containing scenes, dialogue, and actions, overcoming fragmentation and lack of interaction. The copyright verification module accurately calculates similarity from three dimensions: text, plot structure, and character relationships, comprehensively mitigating the risk of substantial similarity. The linking element extraction stage outputs information needed for character design, storyboarding, and video generation, achieving seamless integration between script and video production. This comprehensive approach forms an integrated solution from "prompt input" to "useful script and production element output," shortening the creation cycle, reducing costs, ensuring copyright security, and giving the script practical value for directly empowering video production, thus comprehensively improving the efficiency and quality of the entire scriptwriting process.

[0064] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An AI-based scriptwriting method, characterized in that, Including the following steps: S1. Receive the initial prompt words and identify and extract the core elements of the script from the initial prompt words; S2. Input the core elements of the script into the pre-trained script generation model, and the script generation model generates an initial script; The script generation model employs a large language model, trained on a mixture of publicly available scripts without copyright validity and non-script domain text. Obtain publicly available scripts without copyright protection, as well as texts outside the script field; Extract pre-annotated transferable elements from non-script domain texts and generate corresponding script-like expressions based on NLP models; A hybrid corpus is constructed based on the script-like expressions of the publicly available scripts and the transferable elements, and the large language model is pre-trained in unsupervised manner. The script generation model is obtained by supervising the large language model using the publicly available scripts. S3. Calculate the text similarity, plot structure similarity, and character relationship similarity between the initial script and the copyrighted scripts in the script database; S4. Output the initial script only when all similarities are less than the preset percentage; otherwise, return to step S2.

2. The AI-based scriptwriting method according to claim 1, characterized in that, Step S3 includes: S31. Calculate the text overlap between the initial script and each copyrighted script using cosine similarity to obtain the text similarity. S32. Construct the plot graph of the initial script, and calculate the similarity between the plot graph of the initial script and the plot graph of each copyrighted script through a pre-trained graph neural network to obtain the plot structure similarity. S33. Construct the character interaction matrix of the initial script, and perform matrix similarity calculation with the character interaction matrix of each copyrighted script to obtain the character relationship similarity.

3. The AI-based scriptwriting method according to claim 2, characterized in that, Step S31 includes the following steps: S311. Based on NLP-based text structure extraction technology, the initial script and the copyrighted script are decomposed into three types of sub-texts: dialogue fragments, scene descriptions, and action prompts. S312. Using the Sentence-BERT model, each type of sub-text is encoded to obtain a text vector; S313. For the text vectors of the same type corresponding to the initial script and the copyright script, calculate the cosine similarity respectively, and perform a weighted calculation on the cosine similarity of the three types of sub-texts to obtain the text similarity.

4. The AI-based scriptwriting method according to claim 2, characterized in that, Step S32 includes the following steps: S321. Using the ACE event extraction model, time is extracted from the initial script, event nodes in the plot graph are established, and time logic relationships are marked to obtain the connections between each event node in the plot graph, thereby generating the plot graph of the initial script. S322. Based on a pre-trained graph attention network, calculate the node similarity and edge similarity of the plot graph of the initial script and the copyrighted script, and calculate the plot structure similarity by weighting the node similarity and edge similarity.

5. The AI-based scriptwriting method according to claim 2, characterized in that, Step S33 includes the following steps: S331. Use NLP models to identify characters and extract interactions, and quantify and statistically analyze interaction types to construct a character interaction matrix; S332. For the initial script and the copyrighted script, calculate the element differences of the character interaction matrix using the Manhattan distance formula, and convert them into similarity to obtain the character relationship similarity.

6. The AI-based scriptwriting method according to claim 1, characterized in that, Step S4 also includes the following steps: S41. When all similarities are less than a preset percentage, a semantic parsing model is used to extract the connecting elements information, including character design, storyboard production, and video generation, from the initial script and output them along with the initial script.

7. The AI-based scriptwriting method according to claim 6, characterized in that, Extracting character design coherence information from the initial script includes the following steps: Based on the visual description extraction model, visual features are analyzed from the initial script to obtain information on the connecting elements of the character design.

8. The AI-based scriptwriting method according to claim 6, characterized in that, Extracting the transitional elements for storyboard production from the initial script includes the following steps: The semantic segmentation model identifies objects and their positional relationships in the scene, and matches them with a preset camera angle library based on the script plot to obtain the connection element information for the storyboard production.

9. The AI-based scriptwriting method according to claim 6, characterized in that, Extracting video-generated transition element information from the initial script includes the following steps: Based on the time series annotation model, the duration is calculated by combining the number of words in the dialogue and the complexity of the actions. The visual parameter library is then used to match the emotion tags to obtain the connection element information for video generation.

10. An AI-based scriptwriting system, characterized in that, include: Preprocessing module: Receives initial prompts and identifies and extracts the core elements of the script from the initial prompts; Script generation module: Input the core elements of the script into a pre-trained script generation model, and the script generation model generates an initial script; The script generation model employs a large language model, trained on a mixture of publicly available scripts without copyright validity and non-script domain text. Obtain publicly available scripts without copyright protection, as well as texts outside the script field; Extract pre-annotated transferable elements from non-script domain texts and generate corresponding script-like expressions based on NLP models; A hybrid corpus is constructed based on the script-like expressions of the publicly available scripts and the transferable elements, and the large language model is pre-trained in unsupervised manner. The script generation model is obtained by supervising the large language model using the publicly available scripts. Copyright verification module: Calculates the text similarity, plot structure similarity, and character relationship similarity between the initial script and the copyrighted scripts in the script database; The initial script is output only when all similarities are less than a preset percentage; otherwise, the script generation module is returned to regenerate the script.

Images