Vietnamese abstract guide story generation method based on multiple agents and Monte Carlo decision tree
By using a multi-agent approach and Monte Carlo decision trees, Vietnamese novel data is crawled and tasks are decomposed. Initial stories are generated using multi-agent approaches and iteratively optimized using Monte Carlo decision trees. This solves the problems of low resource consumption and limited content in Vietnamese story generation, and improves the coherence and richness of the generated stories.
Patent Information
- Application Number
- CN202511608808.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-05
- Publication Date
- 2026-02-27
AI Technical Summary
Existing large language models suffer from low resource bottlenecks, limited content, and a lack of optimization mechanisms in Vietnamese story generation, resulting in shallow semantic understanding, high repetition rate of generated stories, high missing rate of key plots, and an inability to iteratively optimize text quality.
Using a multi-agent approach and Monte Carlo decision trees, the task of crawling Vietnamese novel data and generating summaries is decomposed into multiple sub-tasks to be completed by different agents. The Monte Carlo decision tree module is used for iterative optimization to simulate the polishing process of human writers.
It significantly improves the coherence, fluency, and richness of Vietnamese story generation, solves the problems of insufficient deep semantic understanding, repetitive plots, and lack of motivation in low-resource languages, and achieves high-quality generation of Vietnamese stories.
Smart Images
Figure CN121580979A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the field of natural language processing and generative artificial intelligence, and specifically relates to a Vietnamese summary guided story generation method based on multi-agent and Monte Carlo decision tree, which is particularly suitable for Vietnamese scenarios. BACKGROUND
[0002] Current large language models have significant defects in Vietnamese story generation: first, low resource bottleneck, Vietnamese accounts for less than 0.1% in mainstream model training data, leading to shallow semantic understanding; second, content is thin, with a story generation repetition rate of up to 41% (experimental data) and a key scene missing rate of over 70%; third, optimization mechanism is missing, traditional end-to-end generation lacks closed-loop evaluation and cannot iteratively improve text quality. Existing technologies cannot balance the semantic depth of low-resource languages, content richness, and iterative optimization capabilities, and there is an urgent need to establish a task decoupling and dynamic evaluation mechanism that adapts to the characteristics of Vietnamese. SUMMARY
[0003] The present application provides a Vietnamese summary guided story generation method based on multi-agent and Monte Carlo decision tree to solve the problems of plot repetition, motivation absence, and content thinness in Vietnamese story generation, and to improve the coherence, fluency, and diversity of generated text.
[0004] The technical solution of the present application is: a Vietnamese summary guided story generation method based on multi-agent and Monte Carlo decision tree, comprising the following steps:
[0005] Step 1, 26,326 data are obtained by crawling Vietnamese novel websites, and ChatGPT is used to generate summaries for each story to build a Vietnamese summary guided story generation dataset;
[0006] Step 2, multi-agent collaboration generates an initial story, decomposes the complex story generation task into multiple sub-tasks, and assigns different agents driven by large language models to complete the sub-tasks step by step;
[0007] Step 3, the initial draft is iteratively optimized through the Monte Carlo decision tree module, in order to solve the shallow problems that may still exist in the initial story, the MCTS module is introduced to iteratively evaluate and modify the story, simulating the "polishing" process of human authors.
[0008] Further, the specific steps of Step 1 include:
[0009] Step 1.1, data crawling: using the Scrapy framework to collect 26,326 original Vietnamese story texts from Vietnamese novel websites through a web crawler program;
[0010] Step1.2, Data cleaning: Filter the crawled story text, remove non-narrative content (such as advertisements, comments) and incomplete text, and keep the effective stories with complete plots;
[0011] Specifically, filter the text length <200 words or >5,000 words (avoid fragmentation or redundancy); remove special symbols (URL, email address) and non-narrative text (such as poetry, news); manually check 10% samples to ensure plot integrity (such as turning events closed loop, clear character motivation);
[0012] Step1.3, Abstract generation: Use ChatGPT model to generate abstract for each effective story.
[0013] Specifically, input ChatGPT model (gpt-4-1106-preview) to generate Vietnamese abstract, prompt template is: "You are a Vietnamese editor, please create a 50-100 word abstract for the following story, strictly use the format 'Summary: [content]' story: {filtered_story}";
[0014] Further, the specific steps of Step2 include:
[0015] Step2.1, Cross-language conversion, input Vietnamese abstract to cross-language intelligent agent, generate English abstract, forced output format is "Retelling Summary: [content]";
[0016] Step2.2, Role setting, input role setting intelligent agent based on English abstract, generate character background, personality and motivation description;
[0017] Step2.3, Joint English abstract and role setting input scenario planning intelligent agent, output four-act structure outline including opening, development, climax and ending;
[0018] Step2.4, Integrate abstract, role setting and scenario outline input story writing intelligent agent, generate Vietnamese story draft.
[0019] Further, the specific steps of Step3 are:
[0020] Step3.1, Initialize decision tree: take the story draft generated in Step2 as the root node state s0, initialize action space A, reward function R, and access counter N(s)=0;
[0021] Step3.2, Traverse the decision tree from the root node, use the upper confidence interval algorithm to select actions until the leaf node;
[0022] Step3.3, expansion phase: if the leaf node does not reach the termination condition: generate action a by modifying the suggestion through the modification suggestion agent t , that is, the modification suggestion; generate a new state s by the story rewriting agent t+1 , that is, rewrite the story;
[0023] Step3.4, simulation phase: start executing actions randomly from the new state s t+1 , until the termination condition R(s)>0.95 or depth=5 is met, record the path reward G;
[0024] Step3.5, backtracking phase: update the node statistical information in the reverse direction along the path;
[0025] Step3.6, termination and output: repeat Step3.2-Step3.5 until the total iteration number K=4, output the story text corresponding to the node with the highest reward value.
[0026] The application also provides a Vietnamese summary guided story generation system based on multi-agent and Monte Carlo decision tree, which comprises a module for executing the Vietnamese summary guided story generation method based on multi-agent and Monte Carlo decision tree.
[0027] The application also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to realize the Vietnamese summary guided story generation method based on multi-agent and Monte Carlo decision tree.
[0028] The application also provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to realize the Vietnamese summary guided story generation method based on multi-agent and Monte Carlo decision tree.
[0029] The application also provides a computer program product comprising a computer program, wherein the computer program is executable by a processor to realize the Vietnamese summary guided story generation method based on multi-agent and Monte Carlo decision tree.
[0030] The application has the beneficial effect that the method significantly improves the coherence, fluency and content richness of the generated story, solves the problem of template caused by insufficient deep semantic understanding of low-resource languages, and solves the problems of plot repetition, motivation loss and content thinness in Vietnamese story generation. BRIEF DESCRIPTION OF DRAWINGS
[0031] Figure 1 The overall method framework flowchart in the application;
[0032] Figure 2Workflow of multi-agent module in the present application;
[0033] Figure 3 Overall workflow of story generation in the present application;
[0034] Figure 4 Dataset construction process in the present application;
[0035] Figure 5 Overall framework of Vietnamese story generation based on multi-agent and Monte Carlo decision tree in the present application (reflecting the "multi-agent generation → MCTS optimization" two-stage process;
[0036] Figure 6 Multi-agent collaboration generation process (refining the input and output and interaction logic of the four agents). DETAILED DESCRIPTION
[0037] Embodiment 1: As shown, the Vietnamese summary guided story generation method based on multi-agent and Monte Carlo decision tree includes the following steps: Figures 1-6
[0038] Step 1, through crawling Vietnamese novel websites and obtaining data, and using ChatGPT to generate summaries for each story, a Vietnamese summary guided story generation dataset is constructed;
[0039] Further, the specific steps of Step 1 include:
[0040] Step 1.1, data crawling: collecting original Vietnamese story texts from Vietnamese novel websites through web crawler programs;
[0041] Step 1.2, data cleaning: filtering the crawled story texts, removing non-narrative content and incomplete texts, and retaining effective stories with complete plots;
[0042] Step 1.3, summary generation: using ChatGPT model to generate summaries for each effective story.
[0043] Step 2, multi-agent collaboration generates an initial story, which divides the complex story generation task into multiple sub-tasks and assigns different agents driven by large language models to complete step by step;
[0044] Further, the specific steps of Step 2 include:
[0045] Step 2.1, cross-language conversion, input Vietnamese summary to cross-language agent, generate English summary, forced output format is "Retelling Summary: [content]";
[0046] Step2.2, role setting is performed, the role setting agent is input based on the English abstract, and role background, personality and motivation description are generated;
[0047] Step2.3, the scenario planning agent is input combined with the English abstract and the role setting, and a four-act structure outline including opening, development, climax and ending is output;
[0048] Step2.4, the story writing agent is input integrated with the abstract, the role setting and the scenario outline, and a Vietnamese story draft is generated.
[0049] Step3, the draft is iteratively optimized through a Monte Carlo decision tree module, and the MCTS module is introduced to iteratively self-evaluate and modify the story.
[0050] Further, the specific steps of Step3 are:
[0051] Step3.1, initialization of the decision tree: the story draft generated in Step2 is taken as the root node state s0, the action space A, the reward function R and the access counter N(s)=0 are initialized;
[0052] Step3.2, the decision tree is traversed from the root node, and the upper confidence interval algorithm is used to select actions until the leaf node;
[0053] Step3.3, expansion stage: if the leaf node does not reach the termination condition: the action a t , that is, the modification suggestion, is generated through the modification suggestion generation agent; the new state s t+1 , that is, the rewritten story, is generated through the story rewriting agent;
[0054] Step3.4, simulation stage: the action is randomly executed from the new state s t+1 , until the termination condition R(s)>0.95 or depth=5 is met, and the path reward G is recorded;
[0055] Step3.5, backtracking stage: the node statistical information is updated in the reverse direction along the path;
[0056] Step3.6, termination and output: Steps 3.2-3.5 are repeated until the total iteration number K=4, and the story text corresponding to the node with the highest reward value is output.
[0057] The application also provides a Vietnamese abstract guided story generation system based on multi-agent and Monte Carlo decision tree, which comprises a module for executing the Vietnamese abstract guided story generation method based on multi-agent and Monte Carlo decision tree.
[0058] The application further provides an electronic device, including a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the Vietnamese summary guided story generation method based on multi-agent and Monte Carlo decision tree when executing the program.
[0059] The application further provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the Vietnamese summary guided story generation method based on multi-agent and Monte Carlo decision tree.
[0060] The application further provides a computer program product comprising a computer program executable by a processor to implement the Vietnamese summary guided story generation method based on multi-agent and Monte Carlo decision tree.
[0061] Embodiment 2, the Vietnamese summary guided story generation method based on multi-agent and Monte Carlo decision tree, the specific steps of the method are as follows:
[0062] Step 1, obtain data by crawling Vietnamese novel websites and generate summaries for each story using ChatGPT to build a Vietnamese summary guided story generation dataset; specifically including:
[0063] When building the Vietnamese summary guided story generation dataset, first, the story text in the form of static HTML page is crawled from the Vietnamese novel website, the network crawler is written by using Python and the XPath technology is used to accurately locate and extract The story content within the tags was then analyzed. Data cleaning followed, including using the BeautifulSoup library to remove HTML tags, normalize text formatting, and remove duplicates to ensure data uniqueness. Next, ChatGPT was used to generate a Vietnamese summary for each story, which was then manually reviewed and corrected to ensure consistency between the summary and the story content. Finally, the dataset was divided into training, validation, and test sets in an 8:1:1 ratio. Statistics show that the dataset contains 26,326 data points, covering 16 story genres, with an average of 98 sentences per story and a single-sentence summary for each story.
[0064] Step 2: Multi-agent collaboration generates the initial story. This method decomposes the complex story generation task into multiple sub-tasks, which are completed step by step by agents driven by a large language model.
[0065] First, the cross-language understanding agent translates the Vietnamese summary into English to improve the model's comprehension ability. Next, the character-setting agent designs character backgrounds, personalities, and motivations based on the English summary. Then, the plot planning agent generates a four-part plot outline including an opening, development, climax, and ending based on the summary and character settings. Finally, the story-writing agent combines all the information to generate a complete Vietnamese story, ensuring rich emotion and linguistic coherence. The entire collaborative process is represented by a formula; for example, the process of the agent understanding the Vietnamese summary into English can be represented as follows: ,in This represents the English summary output by the intelligent agent.
[0066] Step 3: Iterative optimization of the initial draft using the Monte Carlo Decision Tree module. To address potential issues with the initial story's superficiality, the Monte Carlo Decision Tree (MCTS) module was introduced for iterative optimization, simulating the "polishing" process of a human writer. This module models story improvement as a Markov Decision Process (MDP), where the state represents the current story text, actions are modification suggestions, state transitions generate a new story based on actions, and the reward function evaluates story quality through a rating agent. The optimization process includes a four-step iterative mechanism: the selection phase uses the UCB formula to balance exploration and exploitation; the expansion phase generates suggestions from the modification suggestions, and the story rewriting agent executes the rewriting; the simulation phase continuously expands until the terminal state; and the backtracking phase updates the Q-value function. Finally, the story text with the highest reward is output, with hyperparameter settings such as temperature = 0.7 and maximum depth = 5 ensuring optimization efficiency.
[0067] As a preferred embodiment of the present invention, the specific steps of Step 1 are as follows:
[0068] Step 1.1 Data crawling;
[0069] The story texts in the form of static HTML pages are crawled from a Vietnamese novel website. Each novel contains multiple chapters, and each chapter corresponds to an independent HTML file. A web crawler is written in Python to accurately locate and extract the texts in each chapter by using XPath technology. The story title and the body content within the tag.
[0070] Step 1.2 Data cleaning and denoising;
[0071] Remove HTML tags (such as 、 Normalization: Replace consecutive line breaks with a single line break and compress multiple spaces into a single space; De-duplication: Compare each story text and remove completely duplicated texts to ensure the uniqueness of the dataset.
[0072] Step1.3 Summary Generation and Verification;
[0073] Generate a Vietnamese summary for each story using ChatGPT; manually review the generated results, correct any incorrect summaries, and ensure that the summaries are consistent with the story content; format the summaries as a single sentence that summarizes the core plot of the story.
[0074] As a preferred embodiment of the invention, the Step2 multi-agent collaboration generates an initial story, which divides the complex story generation task into multiple sub-tasks and assigns different agents driven by large language models to complete them step by step. The specific steps are as follows:
[0075] Step2.1 Cross-lingual Agent
[0076] Role: Vietnamese multi-language understanding expert; Task: Convert the input Vietnamese summary into English to improve model understanding; Prompt template: includes role description, task instructions, and output format requirements (e.g. "Retelling Summary: [content]").
[0077] Table 1 Cross-lingual Agent Prompt Template
[0078] Step2.2 Character Design Agent
[0079] Role: Story character designer; Task: Design character background, personality, motivation, and relationship based on English summary; Input: English summary + character design prompt template; Output: Structured character description text.
[0080] Table 2 Character Design Agent Prompt Template
[0081] After the cross-lingual understanding agent converts the input Vietnamese summary into an English summary, the invention uses a character design agent that is responsible for designing the characters that will appear in the story based on the story summary. The prompt template for the character design agent is shown in Table 2, where {summary} represents the English summary generated by the previous agent. The invention uses and
[0082] Step2.3 Plot Planning Agent: Jointly input the English summary and character setting to the plot planning agent, output a four-act structure outline including opening, development, climax, and ending.
[0083] Role: Plot planner; Task: Design a four-act story outline including opening, development, climax, and ending based on the summary and character setting; Output: A one-sentence description for each part, ensuring logical coherence and conflict setup.
[0084] Table 3 Plot Planning Agent Prompt Template
[0085] After obtaining the character setting information of the story, the plot planning agent is responsible for generating the plot outline of the story based on the story summary and the story character setting. The prompt template of this agent is shown in Table 3, where {character_info} represents the character setting generated by the previous agent. The invention uses and to represent the role description and task input of the plot planning agent, respectively. The process of the plot planning agent generating the plot outline of the story can be represented as
[0086] Step2.4 Story Writing Agent
[0087] Role: Vietnamese story writer; Task: Generate a complete Vietnamese story by combining the summary, character setting, and plot outline; Output: A story text with rich emotions and consistent language style.
[0088] Table 4 Story Writing Agent Prompt Template
[0089] Finally, the invention uses a story writing agent to generate a complete Vietnamese story text based on the previously generated story character setting and plot outline. The invention uses and to represent the role description and task input of the plot planning agent, respectively. The process of the plot planning agent generating the plot outline of the story can be represented as Vietnamese story text generated by the agent. The prompt template for this agent is shown in Table 4, where {plot_info} represents the plot outline generated by the previous agent.
[0090] As a preferred scheme of the present application, Step 3 iteratively optimizes the first draft through a Monte Carlo decision tree module, introducing a MCTS module to iteratively self-evaluate and modify the story. The specific steps are as follows:
[0091] Step 3.1 Model the problem as an MDP state (C) with tuple Express the story improvement process as a Markov decision process, where represents the state space, represents the action space, represents a state transition function , and represents a reward function . For any given current state , the present application uses a modification suggestion generation agent to generate an operation based on , where , are the role description and task input prompt used by the modification suggestion generation agent to prompt the model to generate operations. Then the present application uses a story rewriting agent to generate a new state according to the function , where , are the role description and task input prompt for the story rewriting agent. Specifically, the agent is responsible for rewriting the current story according to the current story modification suggestions , thereby generating a new story, similar to the process of a human writer modifying the story content according to the modification suggestions. The quality of each new state produced by the action is determined by the reward function .
[0092] Step 3.2 Four-step iterative optimization mechanism, Selection: use UCB formula to balance exploration and utilization, select the optimal child node; Expansion: modification suggestion generation agent proposes modification suggestions, story rewriting agent performs rewriting; Simulation: continuously expand until the terminal state is reached (reward is high enough or maximum depth is reached); Back-propagation: update Q value function along the path, aggregate future rewards.
[0093] In the selection phase, starting from the root node, each layer of the tree is traversed, selecting subsequent child nodes in each layer and stopping at the leaf nodes. The present invention uses an upper confidence bound algorithm to select nodes, which balances exploration (selecting nodes that have not been sufficiently explored) and exploitation (selecting the node that is currently most promising). The formula for the algorithm is as follows:
[0094]
[0095] where, represents the action set of node , V(n) is the visit count of node , Q(n, a) is the number of child nodes of node after performing action , and c is a search constant.
[0096] In the expansion phase, the present invention adds new child nodes to the leaf nodes selected in the selection phase. The present invention generates the actions of the current story node by modifying the suggestion generation agent, and then generates the next state of the story by using the story rewriting agent. This process can be represented as
[0097]
[0098]
[0099] Step 3.3 Agent collaboration optimization, suggestion generation agent: analyze the consistency of the story with the abstract, characters, and outline, and make improvement suggestions; story rewriting agent: rewrite the story according to the suggestions to ensure consistency with the abstract and logic; scoring agent: evaluate the quality of the story in the form of multiple choices (the probability of option A as the score).
[0100] Table 5 Suggestion generation agent prompt template
[0101] Table 6 Story rewriting agent prompt template
[0102] The prompt templates for the suggestion generation agent and the story rewriting agent are shown in Tables 5 and 6. {story_text} represents the initial story generated by the story writing agent, and {gradient_text} represents the modification suggestions generated by the suggestion generation agent.
[0103] After generating a new node, the present application scores the story corresponding to the new node by inputting a scoring hint to the model as a reward for the new node. The scoring hint is designed in the form of a multiple-choice question , and the probability of selecting option "A" is used to represent the score of the story:
[0104]
[0105]
[0106] The scoring hints used are shown in Table 7.
[0107] Table 7 Scoring hints for stories
[0108] Step 3.4 Termination condition and output, terminal state: the reward exceeds the highest score in history or reaches a preset depth (e.g. L = 5); output: the story text corresponding to the node with the highest reward.
[0109] In the simulation phase, the present application simulates the future trajectory of the selected node in the expansion step, and continues to expand the node until it reaches the terminal state of the node, in order to calculate the future reward of the node. The present application defines the terminal state as the reward obtained by the new node being greater than the maximum reward obtained by the previous node or the depth of the tree reaching a specified value. The present application iteratively operates on the node using the same expansion method as the previous expansion phase, and selects the node with the highest reward value to continue expanding the next level of the tree.
[0110] In the backtracking phase, the present application feeds back the results of the simulation to each parent node in the tree and updates their statistical information. The present application updates the value function by propagating the future reward in the reverse direction along the path from the root to the terminal node. Specifically, for each state-action pair in the path, the value function is updated by aggregating the rewards of all future trajectories starting from , as follows:
[0111]
[0112] where is the number of future trajectories starting from , and is the th state in the future trajectory starting from .
[0113] The Monte Carlo decision tree module explores multiple story rewriting results by iteratively performing the above-mentioned selection, expansion, simulation, and backtracking operations. Finally, the present application selects the story corresponding to the node with the highest reward value from the tree as the output.
[0114] Step3.5 In all experiments, the temperature of the model was set to 0.7, the top-p parameter was set to 0.9, the top-k parameter was set to 40, the maximum depth of the Monte Carlo decision tree was set to 5, the number of generated actions was set to 3, the number of iterations was set to 4, and the search constant was set to 2.5.
[0115] Table 8 Coherence indicator prompts
[0116] Table 9 Fluency indicator prompts
[0117] In recent years, research on using large language models as automatic evaluation tools has made progress. These studies show that large models can accurately understand and execute user-set evaluation instructions and exhibit evaluation capabilities comparable to human experts in text quality evaluation tasks, and using large models as evaluators is becoming a reliable evaluation mode. Referring to the work of Wang et al., the present invention designs a prompt template for evaluating the generation results of a model, and uses ChatGPT to evaluate the story generated by the model from the aspects of coherence, fluency, relevance, and diversity. Coherence mainly measures the relevance, causal relationship, and dependency in time sequence between sentences; fluency mainly measures the language quality and grammatical correctness within a sentence; relevance mainly measures the relevance between the story and the abstract; and diversity mainly measures whether similar plots or content are repeated multiple times in the story. The prompts input into ChatGPT are shown in Tables 8, 9, 10, and 11.
[0118] In addition, the present invention also uses the Distinction-n (abbreviated as D-n) indicator to measure the diversity of vocabulary in the generated story. Distinction-n evaluates the richness of vocabulary in the story by calculating the proportion of different n-grams in the generated text among all n-grams.
[0119] Table 10 Relevance indicator prompts
[0120] Table 11 Diversity indicator prompts
[0121] The specific embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above-described embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present invention.
Claims
1. A Vietnamese summary-guided story generation method based on multi-agent and Monte Carlo decision trees, characterized in that, Includes the following steps: Step 1: Obtain data by crawling Vietnamese novel websites, and use ChatGPT to generate summaries for each story. The constructed Vietnamese summaries guide the generation of the story dataset. Step 2: Multi-agent collaboration generates the initial story. The complex story generation task is broken down into multiple sub-tasks and assigned to different agents driven by a large language model to complete them step by step. Step 3: Iterate and optimize the initial draft using the Monte Carlo decision tree module, and introduce the MCTS module to conduct iterative self-evaluation and modification of the story.
2. The Vietnamese summary-guided story generation method based on multi-agent and Monte Carlo decision trees according to claim 1, wherein the specific steps of Step 1 include: Step 1.1, Data Crawling: Collect original Vietnamese story texts from Vietnamese novel websites using a web crawler program; Step 1.2, Data Cleaning: Filter the crawled story texts, remove non-narrative content and incomplete texts, and retain valid stories with complete plots; Step 1.3, Summary Generation: Use the ChatGPT model to generate a summary for each valid story.
3. The Vietnamese summary-guided story generation method based on multi-agent and Monte Carlo decision trees according to claim 1, characterized in that: The specific steps of Step 2 include: Step 2.1: Cross-language conversion. Input the Vietnamese summary into the cross-language agent to generate an English summary. Force the output format to "Retelling Summary: [content]". Step 2.2: Character setting. Based on the English summary input, the character setting agent is generated to describe the character's background, personality, and motivation. Step 2.3: Input the plot planning agent with the combined English summary and character settings, and output a four-act structure outline including the opening, development, climax and ending; Step 2.4: Input the summary, character settings, and plot outline into the story writing AI to generate a first draft of the story in Vietnamese.
4. The Vietnamese summary-guided story generation method based on multi-agent and Monte Carlo decision trees according to claim 1, characterized in that: The specific steps of Step 3 are as follows: Step 3.1 Initialize the decision tree: Using the initial story draft generated in Step 2 as the root node state s0, initialize the action space A, reward function R, and access counter N(s) = 0; Step 3.2: Traverse the decision tree starting from the root node, using the upper confidence interval algorithm to select actions until the leaf node is reached; Step 3.3, Expansion Phase: If the leaf node does not meet the termination condition: Generate action a by modifying the suggestions to generate an agent. t In other words, it's about making modification suggestions; rewriting the story to generate new states for the agent. t+1 That is, rewriting the story; Step 3.4, Simulation Phase: From the new state s t+1 Start executing random actions until the termination condition R(s)>0.95 or depth=5 is met, and record the path reward G; Step 3.5, Backtracking Phase: Update node statistics information in reverse along the path; Step 3.6, Termination and Output: Repeat Step 3.2-Step 3.5 until the total number of iterations K=4, and output the story text corresponding to the node with the highest reward value.
5. A Vietnamese summary-guided story generation system based on multi-agent and Monte Carlo decision trees, characterized in that, The system includes a module for performing the Vietnamese summary-guided story generation method based on multi-agent and Monte Carlo decision trees as described in any one of claims 1 to 4.
6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the Vietnamese summary-guided story generation method based on multi-agent and Monte Carlo decision trees as described in any one of claims 1 to 4.
7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the Vietnamese summary-guided story generation method based on multi-agent and Monte Carlo decision trees as described in any one of claims 1 to 4.
8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the Vietnamese summary-guided story generation method based on multi-agent and Monte Carlo decision trees as described in any one of claims 1 to 4.