Automatic generation method and device of plot abstract and electronic equipment

Abstract

The invention relates to an automatic generation method and device of a plot abstract and electronic equipment. The method comprises the following steps: performing scene segmentation and key plot anchor point identification on each single-set original script of a target video, and generating a corresponding single-set draft abstract by using a large language model; performing multi-dimensional quality evaluation and screening processing on each single-set draft abstract, and converting the single-set draft abstract into a structured data object; sequentially arranging all the data objects to obtain a single-set abstract sequence of the target video; inputting the single-set abstract sequence into a large language model to enable the large language model to generate a whole draft abstract of the target video in combination with a preset narrative structure guide prompt; and performing multi-dimensional quality evaluation on the whole drama draft abstract, and optimizing the whole drama draft abstract based on an evaluation result to obtain a drama abstract of the target video. According to the method and the device, the technical problem that the effect and the reliability of an existing method in automatic abstracting of the long series are poor is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of video processing, and in particular to a method and device for automatically generating a plot summary and an electronic device. BACKGROUND

[0002] With the rapid development of online video platforms, users face the problem of screening and discovering a large number of episode contents. A high-quality episode summary can help users quickly grasp the core plot, main characters and style characteristics, thereby improving the viewing experience. For content providers, automatic summary generation can support user recommendation, content management, creative assistance, marketing promotion and improvement of content accessibility, etc. Existing episode summary techniques mainly include two types: extractive and generative. The extractive type realizes the summary by selecting important fragments in the original text, ensuring the authenticity of the information but lacking coherence and generalization. The generative type is based on natural language generation technology and reorganizes the text to generate a summary, but early models such as RNN (Recurrent Neural Network) or LSTM (Long Short-Term Memory) have limited ability to handle long texts and complex narrative structures. In recent years, LLMs (Large Language Models) have made breakthroughs in text understanding and generation, and can better capture context and semantic information. However, there are still challenges in directly applying existing LLMs to long episode summaries: first, long-range context dependency is difficult to handle, resulting in missing important plot or character relationships; second, it is difficult to guarantee factual consistency, and "hallucinations" of text are prone to occur; third, it is difficult to identify and select core plots, and it is difficult to distinguish between primary and secondary information; fourth, the automaticity of summary quality evaluation is low and the dimension is single; fifth, invalid or non-plot content interferes with summary generation, such as advertisements, opening and closing credits, and repetitive scene descriptions. The above problems limit the effectiveness and reliability of existing methods in automatic summaries of long episodic dramas. SUMMARY

[0003] The present application provides a method and device for automatically generating a plot summary and an electronic device to solve the technical problem of poor effectiveness and reliability of existing methods in automatic summaries of long episodic dramas.

[0004] In a first aspect, the application provides a method for automatically generating a plot summary, comprising: performing scene segmentation and key plot anchor point identification on each single-episode original script of a target video, and generating a corresponding single-episode draft summary using a large language model; performing multi-dimensional quality assessment and screening processing on each single-episode draft summary, and converting the single-episode draft summary into a structured data object; sequentially arranging all data objects to obtain a single-episode summary sequence of the target video; inputting the single-episode summary sequence into the large language model to enable the large language model to generate an entire-drama draft summary of the target video in combination with a preset narrative structure guide prompt; performing multi-dimensional quality assessment on the entire-drama draft summary, and optimizing the entire-drama draft summary based on the assessment result to obtain a plot summary of the target video.

[0005] In a second aspect, the application provides an apparatus for automatically generating a plot summary, comprising: a first generation module configured to perform scene segmentation and key plot anchor point identification on each single-episode original script of a target video, and generate a corresponding single-episode draft summary using a large language model; a conversion module configured to perform multi-dimensional quality assessment and screening processing on each single-episode draft summary, and convert the single-episode draft summary into a structured data object; a sorting module configured to sequentially arrange all data objects to obtain a single-episode summary sequence of the target video; a second generation module configured to input the single-episode summary sequence into the large language model to enable the large language model to generate an entire-drama draft summary of the target video in combination with a preset narrative structure guide prompt; and an optimization module configured to perform multi-dimensional quality assessment on the entire-drama draft summary, and optimize the entire-drama draft summary based on the assessment result to obtain a plot summary of the target video.

[0006] As an optional example, the first generation module comprises: a first processing unit configured to determine each single-episode original script that has not been processed as a current single-episode original script, and perform the following processing on the current single-episode original script: performing scene segmentation on the current single-episode original script based on timestamps, blank lines, and scene description words in the current single-episode original script to obtain a plurality of scene segments; inputting all scene segments into the large language model to enable the large language model to perform key plot anchor point identification on all scene segments to obtain a key plot anchor point list of the current single-episode original script; and attaching a type label to each key plot anchor point in the key plot anchor point list.

[0007] As an optional example, the first generation module comprises a second processing unit configured to determine each unprocessed episodic original script as a current episodic original script, and perform the following processing on the current episodic original script: dividing the current episodic original script into a plurality of text blocks; selecting a preceding text block of each text block to form a corresponding short-term context; dynamically weighting and combining each text block, its short-term context, a preceding summary, and a key plot anchor related thereto to obtain a corresponding dynamic context; inputting the dynamic context of each text block into the large language model to enable the large language model to generate a text summary of each text block according to a prompt template containing a thought chain reasoning; and sequentially splicing the text summaries of all text blocks to obtain an episodic draft summary of the current episodic original script, and identifying and marking invalid content of the episodic draft summary.

[0008] As an optional example, the conversion module comprises a third processing unit configured to determine each unprocessed episodic draft summary as a current episodic draft summary, and perform the following processing on the current episodic draft summary: calculating an information coverage of the current episodic draft summary based on an episodic original script of the current episodic draft summary; inputting the current episodic draft summary into the large language model to enable the large language model to score the fluency of the current episodic draft summary to obtain a fluency of the current episodic draft summary; extracting key entities and relationships in the current episodic draft summary and comparing them with the episodic original script to calculate a fact consistency of the current episodic draft summary; weighting and summing the information coverage, the fluency, and the fact consistency of the current episodic draft summary to obtain a comprehensive quality score of the current episodic draft summary; marking or deleting the current episodic draft summary if the comprehensive quality score is less than a preset threshold; and converting the current episodic draft summary into a structured data object if the comprehensive quality score is greater than or equal to the preset threshold.

[0009] As an optional example, the third processing unit is further configured to extract core feature labels from the current episodic draft summary using the large language model, wherein the core feature labels comprise a core character list, core events, and emotional tones; and organize the current episodic draft summary, its episode number, and the core feature labels into a structured data object.

[0010] As an optional example, the second generation module comprises: a formulation unit configured to formulate the preset narrative structure guide prompt for the target video; and a first generation unit configured to input the single-set abstract sequence and the preset narrative structure guide prompt into the large language model, so that the large language model comprehensively understands the overall plot of the target video and generates an overall draft abstract of the target video.

[0011] As an optional example, the optimization module comprises: a first calculation unit configured to calculate a semantic relevance of the overall draft abstract by a text vector method based on the overall original script of the target video and the overall draft abstract; a second calculation unit configured to compare the overall draft abstract with the overall original script for factual consistency by a natural language reasoning model to calculate a factual consistency degree of the overall draft abstract; a third calculation unit configured to evaluate a key scene coverage of the overall draft abstract based on a key scene anchor list to calculate a key scene coverage degree of the overall draft abstract; a fourth calculation unit configured to evaluate a length and content of the overall draft abstract to calculate a conciseness degree of the overall draft abstract; a second generation unit configured to generate an optimization signal based on the semantic relevance, the factual consistency degree, the key scene coverage degree, and the conciseness degree of the overall draft abstract; and an optimization unit configured to apply the optimization signal to the overall draft abstract and optimize the overall draft abstract by an iterative loop to obtain the plot abstract of the target video.

[0012] In a third aspect, the present application provides a storage medium having a computer program stored therein, wherein the computer program is run by a processor to execute the automatic generation method of the plot abstract.

[0013] In a fourth aspect, the present application further provides an electronic device comprising a memory and a processor, wherein the memory has a computer program stored therein, and the processor is configured to execute the automatic generation method of the plot abstract by the computer program.

[0014] Compared with the prior art, the technical solution provided by the embodiments of the present application has the following advantages: The application adopts scene segmentation and key plot anchor point identification on each single-set original script of a target video, and generates a corresponding single-set draft summary by using a large language model; each single-set draft summary is subjected to multi-dimensional quality evaluation and screening processing, and is converted into a structured data object; all data objects are sequentially arranged to obtain a single-set summary sequence of the target video; the single-set summary sequence is input into the large language model to enable the large language model to generate an entire-drama draft summary of the target video in combination with a preset narrative structure guide prompt; the entire-drama draft summary is subjected to multi-dimensional quality evaluation, and is optimized based on the evaluation result to obtain a plot summary of the target video. In the above method, each single-set original script of a target video is subjected to scene segmentation and key plot anchor point identification, a single-set draft summary is generated by using a large language model, and is converted into a structured data by multi-dimensional quality evaluation and screening to form a single-set summary sequence arranged in playing order; then the sequence is input into a large language model to generate an entire-drama draft summary in combination with a preset narrative structure prompt, and is iteratively optimized by multi-dimensional evaluation results such as semantic relevance, factual consistency and key plot coverage to realize a final high-quality plot summary. Thus, the purpose of significantly improving the accuracy, integrity and coherence of the summary is achieved, and the technical problem of poor effect and reliability of the existing method in automatic summary of long continuous dramas is solved. BRIEF DESCRIPTION OF DRAWINGS

[0015] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and serve to explain the principles of the present application together with the specification.

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without any creative effort.

[0017] One or more embodiments are exemplarily illustrated by pictures in the drawings corresponding thereto, and these exemplary illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings represent similar elements, unless otherwise specified. The drawings in the drawings do not constitute a proportional limitation.

[0018] Figure 1 is a flowchart of an optional automatic plot summary generation method according to an embodiment of the present application; Figure 2 is a specific implementation flowchart of an optional automatic plot summary generation method according to an embodiment of the present application; Figure 3is a structural schematic diagram of an optional automatic generation device of a plot summary according to an embodiment of the present application; Figure 4 is a schematic diagram of an optional electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in connection with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0020] The following disclosure provides many different embodiments, or examples, for implementing different structures of the present application. For the purpose of simplicity and clarity, the description in the following text describes the components and settings of specific examples. Of course, they are only examples, and the purpose is not to limit the present application. In addition, reference numerals and / or letters can be repeated in different examples in the present application. Such repetition is for the purpose of simplification and clarity, and does not in itself indicate a relationship between the various embodiments and / or settings being discussed.

[0021] According to a first aspect of the embodiments of the present application, an automatic generation method of a plot summary is provided, which is optional, as shown in Figure 1 The above method comprises: S102, scene segmentation and key plot anchor point identification are performed on each single-episode original script of a target video, and a corresponding single-episode draft summary is generated by using a large language model; S104, multi-dimensional quality assessment and screening processing are performed on each single-episode draft summary, and the single-episode draft summary is converted into a structured data object; S106, all data objects are sequentially arranged to obtain a single-episode summary sequence of the target video; S108, the single-episode summary sequence is input into the large language model to make the large language model generate an entire-drama draft summary of the target video in combination with a preset narrative structure guide prompt; S110, multi-dimensional quality assessment is performed on the entire-drama draft summary, and the entire-drama draft summary is optimized based on the assessment result to obtain a plot summary of the target video.

[0022] Optionally, in the present embodiment, an automatic generation method of a plot summary is provided, which realizes high-quality summary generation from a single episode to an entire drama through multi-level processing of an original script of a target video, and a specific implementation process is as shown in Figure 2 The above method comprises four core stages: Phase One: Single-episode Abstract Generation Stage.

[0023] The original scripts of each episode of the target video are preprocessed and structurally analyzed. First, the single-episode original script is divided into scene segments by identifying timestamps, scene descriptions, or specific markers in the subtitles. Then, a large language model (LLM) is used to identify key plot anchors, including but not limited to plot twists, core conflicts, or key character interactions, to provide guidance information for abstract generation. On this basis, each scene segment is input into the large language model, and through dynamic context construction, the current scene segment is considered in combination with the previous scene, the generated abstract, and related key plot anchors to generate a single-episode draft abstract. Make full use of the Chain-of-Thought (COT) prompt to guide the large language model to perform intermediate reasoning before generating the abstract to ensure that the abstract content is accurate, complete, and consistent with the plot context.

[0024] Phase Two: Single-episode Abstract Filtering and Structuring Stage.

[0025] The goal of this stage is to improve the quality of single-episode abstracts and convert them into structured data. First, multi-dimensional quality assessment is performed on the single-episode draft abstract, including but not limited to information coverage, text fluency, logical coherence, and factual consistency. Abstracts that fall below the preset threshold will be marked or removed. At the same time, by detecting invalid or non-plot content (such as advertisements, opening credits, or redundant interlude descriptions), the abstract content is further optimized. Then, the filtered and corrected single-episode abstracts are supplemented with key metadata, including but not limited to main characters, core event keywords, and emotional tone, to generate structured data objects. All single-episode data objects are arranged in the order of episode playback to form a complete sequence of single-episode abstracts, providing basic information for the generation of the entire episode abstract.

[0026] Phase Three: Entire Episode Abstract Generation Stage.

[0027] The single-episode abstract sequence generated in the previous stage is used as input, combined with preset narrative structure guidance prompts, and a large language model is used to generate an entire episode draft abstract. The narrative structure prompts include but are not limited to the growth arc of the main characters, the development of the core conflict, and the final outcome, guiding the model to consider the coherence, logic, and appeal of the story when generating the entire episode abstract. Based on a comprehensive understanding of the plot development, character relationships, and theme of the entire episode, the large language model generates a coherent and complete entire episode draft abstract.

[0028] Phase Four: Mixed Evaluation and Iterative Optimization Stage.

[0029] The whole play draft summary is optimized through a multi-dimensional evaluation and feedback mechanism. First, the semantic correlation of the whole play summary with the original script or reference summary is calculated, and the natural language reasoning model is used to check the factual consistency, while the key plot coverage, the conciseness and redundancy of the summary are evaluated. According to the evaluation results, optimization signals are generated, including but not limited to prompt optimization, model fine-tuning or summary self-correction instructions. Then, the optimization signals are applied to the whole play summary, and the quality of the generated summary is iteratively optimized until the preset standard is reached, so as to obtain the final high-quality plot summary.

[0030] Through the above method, efficient and accurate automatic summary generation of long plays can be realized. First, dynamic context perception and key plot anchor point recognition can ensure that the single-set summary accurately reflects the core information of the plot; second, multi-dimensional quality evaluation and structured processing ensure that the summary content is of high quality, the structure is clear and easy for subsequent processing; the whole play summary generation stage is guided by the narrative structure, improving the coherence and logicality of the summary; the mixed evaluation and iterative optimization mechanism further improves the factual consistency and key plot coverage. Under the comprehensive action, the ability of users to quickly understand the play can be significantly improved, which assists content discovery and recommendation, improves platform user experience, and provides efficient tools for content providers in content management, creation assistance, marketing and content accessibility, with significant application value.

[0031] As an optional example, scene segmentation and key plot anchor point recognition are performed on each single-set original script of the target video, including: Each single-set original script that has not been processed is determined as a current single-set original script, and the following processing is performed on the current single-set original script: The current single-set original script is segmented based on the timestamps, blank lines and scene description words in the current single-set original script to obtain a plurality of scene segments; All scene segments are input into a large language model to enable the large language model to perform key plot anchor point recognition on all scene segments to obtain a key plot anchor point list of the current single-set original script; Each key plot anchor point of the key plot anchor point list is attached with a type label.

[0032] Optionally, in the embodiment, scene segmentation and key plot anchor recognition are performed on each single-episode original script of the target video, specifically including the following steps: first, each single-episode original script that has not been processed is determined as a current single-episode original script, and a preprocessing operation is performed on it. Based on the timestamps, blank lines and scene description words (such as “[scene]”, “[indoor]”, “[night]”, etc.) in the current single-episode original script, the original script is scene segmented, and the continuous text is divided into multiple scene segments, each of which is an independent processing unit. Subsequently, all scene segments are input into the large language model, which analyzes each scene segment and identifies key plot anchors that have important guiding significance for plot development, including turning points, core conflicts, major character interactions, etc. For each identified key plot anchor, a type label is attached, such as “turning point”, “key character interaction”, etc., to facilitate information association and priority judgment in the subsequent abstract generation and structured processing stage. Through the above processing, each single-episode original script is converted into a set of scene-organized text segments and a list of corresponding key plot anchors.

[0033] Through the above steps, the core plot information in the episode can be effectively identified, and the original script can be structured and processed to provide a clear context and plot framework for single-episode abstract generation. This processing method based on scene segmentation and key plot anchor recognition can significantly improve the accuracy and completeness of single-episode abstracts, while laying the foundation for coherent generation of full-episode abstracts, avoiding omission of important plots, and improving plot understanding efficiency and subsequent application value.

[0034] As an optional example, generating a corresponding single-episode draft abstract using a large language model includes: Each single-episode original script that has not been processed is determined as a current single-episode original script, and the following processing is performed on the current single-episode original script: The current single-episode original script is divided into multiple text blocks; The prequel text block of each text block is selected to form a corresponding short-term context; Each text block, its short-term context, prequel abstract, and related key plot anchors are dynamically weighted and combined to obtain a corresponding dynamic context; The dynamic context of each text block is input into the large language model to generate a text abstract for each text block according to a prompt template containing thought chain reasoning; The text abstracts of all text blocks are sequentially spliced to obtain a single-episode draft abstract of the current single-episode original script, and the invalid content of the single-episode draft abstract is identified and marked.

[0035] Optionally, in the embodiment, a corresponding single-episode draft summary is generated using a large language model, specifically including the following steps: first, each single-episode original script that has not been processed is determined as a current single-episode original script, and the current single-episode original script is processed. The current single-episode original script is divided into multiple text blocks, and each text block is taken as an independent summary generation unit. For each text block, several text blocks in the prologue thereof are selected as short-term context to retain semantic continuity of continuous plots. At the same time, the generated prologue summary and key plot anchors related to the content of the text block are combined by a dynamic weighting method to form dynamic context information of the text block. Subsequently, each text block and the dynamic context thereof are input into the large language model, and a prompt template containing Chain-of-Thought (COT) reasoning is used to guide the model to perform step-by-step reasoning and analysis before generating the summary, thereby generating a text summary of each text block. The text summaries of all text blocks are sequentially spliced in the order of the original script to obtain a complete single-episode draft summary of the current single-episode original script. Finally, the generated single-episode draft summary is subjected to invalid content identification and marking, including but not limited to advertisements, prologue and epilogue, or scene descriptions unrelated to the plot, to ensure that the subsequent summary processing is not disturbed by low-value content.

[0036] Through the above steps, the understanding and generation capabilities of the large language model can be fully utilized to process the single-episode original script in blocks and combine dynamic context and key plot anchors to effectively capture long-range plot dependencies and important events. The introduction of Chain-of-Thought (COT) reasoning prompts can improve the logical coherence and information coverage of the summary, and the intelligent identification and marking of invalid content further ensure the refinement and high quality of the summary. Overall, an accurate, coherent, and clearly structured single-episode draft summary can be generated, providing a solid foundation for subsequent full-episode summary generation and multi-dimensional evaluation, while improving user's quick understanding of the plot and platform content management efficiency.

[0037] As an optional example, each single-episode draft summary is subjected to multi-dimensional quality evaluation and screening processing, and is converted into a structured data object, including: Each single-episode draft summary that has not been processed is determined as a current single-episode draft summary, and the current single-episode draft summary is processed as follows: Based on the single-episode original script of the current single-episode draft summary, the information coverage of the current single-episode draft summary is calculated; The current single-episode draft summary is input into the large language model to enable the large language model to score the fluency of the current single-episode draft summary, thereby obtaining the fluency of the current single-episode draft summary; Key entities and relationships in the current single-episode draft summary are extracted and compared with the single-episode original script thereof to calculate the factual consistency of the current single-episode draft summary; The information coverage, fluency, and fact consistency of the current single-set draft summary are weighted and summed to obtain a comprehensive quality score of the current single-set draft summary. In the case where the comprehensive quality score is less than the preset threshold, the current single-set draft summary is marked or deleted. In the case where the comprehensive quality score is greater than or equal to the preset threshold, the current single-set draft summary is converted into a structured data object.

[0038] Optionally, in the embodiment, each single-set draft summary is subjected to multi-dimensional quality evaluation and screening processing, and is converted into a structured data object, specifically including the following steps: first, each single-set draft summary that has not been processed is determined as a current single-set draft summary, and information coverage is calculated based on the corresponding single-set original script to evaluate the degree of complete reflection of the original plot by the summary. Subsequently, the current single-set draft summary is input into a large language model to perform fluency scoring, evaluating the summary quality from aspects of syntax, logic, and text coherence. At the same time, key entities and relationships in the summary are extracted and compared with the original script to calculate the fact consistency, so as to ensure the accuracy of the summary content. The information coverage, fluency, and fact consistency are weighted and summed according to predetermined weights to obtain a comprehensive quality score of the current single-set draft summary. For summaries with a comprehensive quality score lower than a preset threshold, marking or deletion is performed to exclude low-quality content; for summaries with a score reaching or exceeding the threshold, the summaries are converted into structured data objects, including summary text, corresponding set number, and key feature labels (such as main characters, core event keywords, and emotional tone). By performing the above processing on all single-set draft summaries, a sequence of high-quality, structured single-set summaries can be formed.

[0039] Through the above steps, multi-dimensional automatic quality evaluation of single-set draft summaries is realized, information coverage, text fluency, and fact consistency are comprehensively considered, low-quality or redundant summaries are effectively excluded, and the accuracy and readability of the summaries are improved. At the same time, high-quality summaries are converted into structured data objects, providing a unified standardized information for episode summary generation, content retrieval, recommendation, and management, thereby improving summary generation efficiency, ensuring content integrity, and enhancing user experience in understanding the plot and utilizing platform content.

[0040] As an optional example, in the case where the comprehensive quality score is greater than or equal to the preset threshold, converting the current single-set draft summary into a structured data object includes: extracting core feature labels from the current single-set draft summary using a large language model, wherein the core feature labels include a list of core characters, core events, and emotional tone; organizing the current single-set draft summary, the set number to which it belongs, and the core feature labels into a structured data object.

[0041] Optionally, in the present embodiment, the current single-episode draft summary is converted into a structured data object when the comprehensive quality score reaches or exceeds the preset threshold, specifically including the following steps: first, input the high-quality single-episode draft summary into a large language model, understand and analyze the summary content through the model, and extract core feature labels, including but not limited to a list of core characters involved in the summary, key events, and main emotional tones, in order to accurately describe the core information and narrative color of the plot. Subsequently, the current single-episode draft summary, its set number, and the extracted core feature labels are organically integrated to form a complete structured data object. This structured data object not only contains text summary information, but also includes key metadata and content feature labels, providing a directly usable data basis for subsequent full-episode summary generation. At the same time, all single-episode data objects can be arranged in the original broadcast order of the episodes to form a complete sequence of single-episode summaries.

[0042] Through the above steps, single-episode plot information can be structured and represented under the premise of ensuring high-quality summaries, achieving unified management of text content and key feature labels. Using a large language model to extract core characters, events, and emotional tones allows the summary to not only retain core plot information but also have searchable, analyzable, and integrable characteristics. Structured data objects facilitate full-episode summary generation, multi-dimensional evaluation, content recommendation, and platform management.

[0043] As an optional example, the single-episode summary sequence is input into a large language model to generate a full-episode draft summary of the target video, including: formulating a preset narrative structure guide prompt for the target video; inputting the single-episode summary sequence and the preset narrative structure guide prompt into the large language model to enable the large language model to comprehensively understand the overall plot of the target video and generate a full-episode draft summary of the target video.

[0044] Optionally, in this embodiment, the single-episode summary sequence is input into the large language model to generate a full-draft summary of the target video in combination with the preset narrative structure guide prompts. Specifically, the following steps are included: first, the preset narrative structure guide prompts are formulated according to the type and plot characteristics of the target video. The prompt content includes a description of the typical narrative structure, such as the main character's growth arc, the generation and resolution of the core conflict, and the arrangement of the climax and ending of the plot, thereby guiding the large language model to maintain the logicality and coherence of the story when generating the summary. Subsequently, the single-episode summary sequence that has undergone quality assessment and structured processing is input into the large language model together with the preset narrative structure guide prompts. After receiving the input, the large language model comprehensively understands and analyzes all single-episode summary content, identifies the plot development context, character relationships, and thematic ideas, and generates a coherent, complete, and narrative-logical full-draft summary based on the guide prompts. The generated full-draft summary covers the overall plot context of the target video while retaining the core information and emotional tone of each episode.

[0045] Through the above steps, high-quality summary generation of the overall plot of a long series can be achieved, effectively solving the problem of coherence of the entire series caused by the fragmentation of single-episode information. By using the large language model in combination with the preset narrative structure guide prompts, not only is the logicality of the summary clear and the character relationships clear, but also the plot development and thematic ideas are fully presented, improving the readability and accuracy of the summary of the entire series.

[0046] As an optional example, the full-draft summary is subjected to multi-dimensional quality assessment, and the full-draft summary is optimized based on the evaluation results to obtain the plot summary of the target video, which includes: Through the text vector method, the semantic relevance of the full-draft summary is calculated based on the original script of the entire series and the full-draft summary of the target video; Using a natural language reasoning model, the fact consistency of the full-draft summary and the original script of the entire series is compared, and the fact consistency of the full-draft summary is calculated; According to the key scene anchor list, the key scene coverage of the full-draft summary is evaluated, and the key scene coverage of the full-draft summary is calculated; The length and content of the full-draft summary are evaluated, and the conciseness of the full-draft summary is calculated; According to the semantic relevance, fact consistency, key scene coverage, and conciseness of the full-draft summary, an optimization signal is generated; The optimization signal is applied to the full-draft summary, and the full-draft summary is optimized through an iterative loop to obtain the plot summary of the target video.

[0047] Optionally, in this embodiment, the multi-dimensional quality evaluation is performed on the whole drama draft summary, and optimization is performed based on the evaluation results, so as to obtain a high-quality target video plot summary. The specific implementation steps are as follows: Firstly, the semantic correlation of the generated whole drama draft summary is evaluated. The text vectorization technology is used to convert the original script of the whole drama and the whole drama draft summary into high-dimensional vector representations respectively, and the cosine similarity or other similarity indicators are calculated to quantify the degree of fit between the whole drama draft summary and the original script in the semantic level. It can be judged whether the summary sufficiently reflects the core content and theme of the original script.

[0048] Secondly, the fact consistency evaluation is performed. The Natural Language Inference (NLI) model is used to take the key statements in the original script of the whole drama as the premise and the corresponding statements in the whole drama draft summary as the hypothesis. The model judges the implication relationship between the two, and outputs the classification results of “implication”, “contradiction” or “neutral”. The proportion of contradictions or inconsistencies in the summary is calculated to obtain the fact consistency of the whole drama draft summary. The consistency of the generated summary and the original script information is effectively ensured, and the description of the fictional or incorrect plot is avoided.

[0049] Then, the key scene coverage evaluation is performed. According to the key scene anchor list identified in the single set summary generation stage, the key scenes contained in the whole drama draft summary are matched and counted, and the proportion of the whole drama draft summary covering the key scenes is calculated. This indicator can measure whether the summary presents the plot development context and important events completely, ensuring that users will not miss the core plot information when reading the summary.

[0050] Then, the summary brevity evaluation is performed. By analyzing the length, repeated information, redundant description and other contents of the whole drama draft summary, the brevity index is obtained. The purpose is to control the length of the summary, make it as concise as possible while maintaining the integrity of the information, and improve the readability and user experience.

[0051] Finally, after completing the above multi-dimensional evaluation, the evaluation indicators are integrated to generate optimization signals. The optimization signals specifically include but are not limited to semantic supplement suggestions, fact correction prompts, key scene completion prompts and redundant content reduction suggestions. Then, these optimization signals are applied to the whole drama draft summary, and the summary content is repeatedly adjusted and improved through an iterative cycle mechanism. In each iteration, the large language model modifies the summary according to the optimization signals, regenerates the text and performs multi-dimensional evaluation again until the indicators meet the preset standards, thereby obtaining the final target video plot summary.

[0052] Through the above steps, high-quality and controllable generation of episode plot summaries is achieved. The multi-dimensional automatic evaluation and iterative optimization mechanism ensures that the summaries meet high standards in terms of semantic relevance, factual consistency, key plot coverage, and conciseness, significantly improving the accuracy and readability of the summaries. This effectively reduces the cost of manual editing, enhances user experience and platform operation efficiency, while ensuring the complete presentation of core information in long-form series and meeting diverse application needs.

[0053] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the order of the described actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0054] According to another aspect of the embodiments of the present application, an automatic generation device for plot summaries is also provided, as shown in Figure 3 The device includes: A first generation module 302 is configured to perform scene segmentation and key plot anchor point identification on each single-episode original script of a target video, and generate a corresponding single-episode draft summary using a large language model. A conversion module 304 is configured to perform multi-dimensional quality evaluation and filtering processing on each single-episode draft summary, and convert it into a structured data object. A sorting module 306 is configured to sequentially arrange all data objects to obtain a single-episode summary sequence of the target video. A second generation module 308 is configured to input the single-episode summary sequence into a large language model to generate an entire-drama draft summary of the target video by combining a preset narrative structure guide prompt. An optimization module 310 is configured to perform multi-dimensional quality evaluation on the entire-drama draft summary, and optimize the entire-drama draft summary based on the evaluation results to obtain a plot summary of the target video.

[0055] It should be noted that the first generation module 302 in this embodiment can be used to perform step S102 in the embodiments of the present application, the conversion module 304 in this embodiment can be used to perform step S104 in the embodiments of the present application, the sorting module 306 in this embodiment can be used to perform step S106 in the embodiments of the present application, the second generation module 308 in this embodiment can be used to perform step S108 in the embodiments of the present application, and the optimization module 310 in this embodiment can be used to perform step S110 in the embodiments of the present application.

[0056] As an optional example, the first generation module includes: The first processing unit is configured to determine each unprocessed episodic original script as a current episodic original script, and perform the following processing on the current episodic original script: Scene segmentation is performed on the current episodic original script based on the time stamp, blank line and scene description word in the current episodic original script, to obtain a plurality of scene segments; All scene segments are input into the large language model, so that the large language model performs key plot anchor point recognition on all scene segments, to obtain a key plot anchor point list of the current episodic original script; Each key plot anchor point in the key plot anchor point list is attached with a type label.

[0057] As an optional example, the first generation module includes: The second processing unit is configured to determine each unprocessed episodic original script as a current episodic original script, and perform the following processing on the current episodic original script: The current episodic original script is divided into a plurality of text blocks; The antecedent text block of each text block is selected to form a corresponding short-term context; Each text block, its short-term context, the antecedent abstract and the key plot anchor point related thereto are dynamically weighted and combined to obtain a corresponding dynamic context; The dynamic context of each text block is input into the large language model, so that the large language model generates a text abstract of each text block according to a prompt template containing thought chain reasoning; The text abstracts of all text blocks are sequentially spliced to obtain an episodic draft abstract of the current episodic original script, and invalid content of the episodic draft abstract is identified and marked.

[0058] As an optional example, the conversion module includes: The third processing unit is configured to determine each unprocessed episodic draft abstract as a current episodic draft abstract, and perform the following processing on the current episodic draft abstract: The information coverage of the current episodic draft abstract is calculated based on the episodic original script of the current episodic draft abstract; The current episodic draft abstract is input into the large language model, so that the large language model performs fluency scoring on the current episodic draft abstract to obtain the fluency of the current episodic draft abstract; Key entities and relationships in the current episodic draft abstract are extracted and compared with the episodic original script thereof, to calculate the factual consistency of the current episodic draft abstract; The information coverage, fluency and factual consistency of the current episodic draft abstract are weighted and summed to obtain a comprehensive quality score of the current episodic draft abstract; In a case where the comprehensive quality score is less than a preset threshold, marking or deleting the current single-set draft summary; In a case where the comprehensive quality score is greater than or equal to the preset threshold, converting the current single-set draft summary into a structured data object.

[0059] As an optional example, the third processing unit is further configured to extract core feature labels from the current single-set draft summary by using the large language model, wherein the core feature labels include a core character list, a core event, and a sentiment tone; Organizing the current single-set draft summary, the set number to which the current single-set draft summary belongs, and the core feature labels into a structured data object.

[0060] As an optional example, the second generation module includes: A formulation unit configured to formulate a preset narrative structure guide prompt for the target video; A first generation unit configured to input the single-set summary sequence and the preset narrative structure guide prompt into the large language model, so that the large language model comprehensively understands the overall plot of the target video and generates an overall draft summary of the target video.

[0061] As an optional example, the optimization module includes: A first calculation unit configured to calculate a semantic relevance degree of the overall draft summary by using a text vector method according to the overall original script of the target video and the overall draft summary; A second calculation unit configured to calculate a fact consistency degree of the overall draft summary by using a natural language reasoning model to compare the fact consistency between the overall draft summary and the overall original script; A third calculation unit configured to calculate a key scene coverage degree of the overall draft summary by using a key scene anchor list to evaluate the key scene coverage of the overall draft summary; A fourth calculation unit configured to calculate a conciseness degree of the overall draft summary by evaluating the length and content of the overall draft summary; A second generation unit configured to generate an optimization signal according to the semantic relevance degree, the fact consistency degree, the key scene coverage degree, and the conciseness degree of the overall draft summary; An optimization unit configured to apply the optimization signal to the overall draft summary and optimize the overall draft summary by using an iterative loop to obtain a plot summary of the target video.

[0062] Other examples of the embodiment are described in the above examples, which will not be repeated here.

[0063] Figure 4 is a schematic diagram of an optional electronic device according to an embodiment of the present application, as Figure 4As shown, the electronic device includes a processor 402, a communication interface 404, a memory 406, and a communication bus 408, wherein the processor 402, the communication interface 404, and the memory 406 communicate with each other through the communication bus 408, and the memory 406 is configured to store a computer program; the processor 402 is configured to execute the computer program stored in the memory 406 to implement the following steps: scene segmentation and key plot anchor point identification are performed on each single-episode original script of the target video, and a corresponding single-episode draft summary is generated by using a large language model; Each single-episode draft summary is subjected to multi-dimensional quality evaluation and screening processing, and is converted into a structured data object; All data objects are sequentially arranged to obtain a single-episode summary sequence of the target video; The single-episode summary sequence is input into the large language model to generate an entire-episode draft summary of the target video by combining a preset narrative structure guide prompt; The entire-episode draft summary is subjected to multi-dimensional quality evaluation, and the entire-episode draft summary is optimized based on the evaluation result to obtain a plot summary of the target video.

[0064] Optionally, in the embodiment, the communication bus described above can be a PCI (Peripheral Component Interconnect) bus, an EISA (Extended Industry Standard Architecture) bus, or the like. The communication bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, Figure 4 only one thick line is used in the figure, but it does not mean that there is only one bus or only one type of bus. The communication interface is used for communication between the electronic device and other devices.

[0065] The memory can include a RAM and can also include a non-volatile memory, for example, at least one disk memory. Optionally, the memory can also be at least one storage device located away from the aforementioned processor.

[0066] As an example, the memory 406 described above can include, but is not limited to, the first generation module 302, the conversion module 304, the sorting module 306, the second generation module 308, and the optimization module 310 in the automatic plot summary generation device described above. In addition, other module units in the automatic plot summary generation device described above can also be included, but are not limited to, and will not be described in detail in this example.

[0067] The processor can be a general processor, which can include but is not limited to: a CPU (Central Processing Unit), a NP (Network Processor), etc.; or can be a DSP (Digital Signal Processing), an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0068] Optionally, the specific examples in the embodiments can refer to the examples described in the above embodiments, and the embodiments will not be described here.

[0069] Those skilled in the art can understand that, Figure 4 The structure shown is only schematic, and the device implementing the automatic generation method of the plot summary can be a terminal device, which can be a smartphone (such as an Android phone, an iOS phone, etc.), a tablet computer, a palm computer, a Mobile Internet Device (MID), a PAD, etc. Figure 4 The structure of the electronic device is not limited. For example, the electronic device can further include more or less components (such as a network interface, a display device, etc.) than Figure 4 or have a different configuration from Figure 4 The structure shown.

[0070] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the hardware related to the terminal device by a program, which can be stored in a computer readable storage medium, and the storage medium can include: a flash disk, a ROM, a RAM, a magnetic disk or an optical disk, etc.

[0071] According to another aspect of the embodiments of the present application, a computer readable storage medium is also provided, which stores a computer program, wherein the computer program is run by a processor to perform the steps in the automatic generation method of the plot summary.

[0072] Optionally, in the embodiment, a person skilled in the art can understand that all or part of the steps of the various methods in the above embodiment can be instructed by a program to complete the hardware related to the terminal device, and the program can be stored in a computer readable storage medium, and the storage medium can include a flash disk, a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a magnetic disk or an optical disk, etc.

[0073] The serial numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0074] The integrated units in the above embodiments, if realized in the form of software function units and sold or used as independent products, can be stored in the above computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that make contributions to the prior art or the whole or part of the technical solutions can be embodied in the form of software products, and the computer software products are stored in the storage medium, including a plurality of instructions for causing one or more computer devices (which can be personal computers, servers or network devices, etc.) to execute all or part of the steps of the methods described in the embodiments of the present application.

[0075] In the above embodiments of the present application, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0076] In the several embodiments provided by the present application, it should be understood that the disclosed client can be implemented in other ways. Among them, the above-mentioned device embodiment is only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed mutual ones can be indirect coupling or communication connection through some interface, unit or module, and can be electrical or other forms.

[0077] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0078] In addition, each of the functional units in the various embodiments of the present application can be integrated in one processing unit, or each of the units can exist physically, or two or more units can be integrated in one unit. The integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0079] The above is only the preferred embodiment of the present application, and it should be pointed out that, for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should be considered as the protection scope of the present application.

Claims

1. A method for automatically generating plot summaries, characterized in that, include: Scene segmentation and key plot anchor point identification are performed on the original script of each episode of the target video, and the corresponding episode draft summary is generated using a large language model; Each episode's draft summary undergoes multi-dimensional quality assessment and screening, and is then converted into structured data objects; All data objects are arranged sequentially to obtain a single-set summary sequence of the target video; The single-episode summary sequence is input into the large language model, so that the large language model, combined with the preset narrative structure guidance prompts, generates a full-episode draft summary of the target video; A multi-dimensional quality assessment is performed on the entire drama draft summary, and the entire drama draft summary is optimized based on the assessment results to obtain the plot summary of the target video.

2. The method according to claim 1, characterized in that, Scene segmentation and key plot anchor point identification for each episode of the original script in the target video include: Each unprocessed episode's original script is identified as the current episode's original script, and the following processing is performed on the current episode's original script: Based on the timestamps, blank lines, and scene description words in the original script of the current episode, the original script of the current episode is segmented into multiple scene fragments. All scene fragments are input into the large language model so that the large language model can identify key plot anchors in all scene fragments and obtain a list of key plot anchors in the original script of the current episode. Attach a type label to each key plot anchor in the list of key plot anchors.

3. The method according to claim 1, characterized in that, The generation of corresponding single-episode draft summaries using a large language model includes: Each unprocessed episode's original script is identified as the current episode's original script, and the following processing is performed on the current episode's original script: Divide the current episode's original script into multiple text blocks; Select the preceding text block of each text block to form the corresponding short-term context; Each text block, its short-term context, preceding summary, and related key plot anchors are dynamically weighted and combined to obtain the corresponding dynamic context. The dynamic context of each text block is input into the large language model so that the large language model generates a text summary of each text block based on a prompt template containing thought chain reasoning. The text summaries of all text blocks are concatenated sequentially to obtain the episode draft summary of the original script for the current episode, and invalid content in the episode draft summary is identified and marked.

4. The method according to claim 1, characterized in that, Each single-episode draft abstract undergoes multi-dimensional quality assessment and screening, and is then converted into structured data objects, including: Each unprocessed episode draft summary is identified as the current episode draft summary, and the following processing is performed on the current episode draft summary: Based on the original script of the current episode draft summary, the information coverage of the current episode draft summary is calculated. The current episode draft summary is input into the large language model so that the large language model scores the fluency of the current episode draft summary and obtains the fluency of the current episode draft summary. Extract the key entities and relationships from the current episode draft summary and compare them with its original episode script to calculate the factual consistency of the current episode draft summary; The information coverage, fluency, and factual consistency of the current episode draft summary are weighted and summed to obtain the overall quality score of the current episode draft summary; If the overall quality score is less than a preset threshold, the current single-episode draft summary is marked or deleted; If the overall quality score is greater than or equal to the preset threshold, the current single-episode draft summary is converted into a structured data object.

5. The method according to claim 4, characterized in that, When the overall quality score is greater than or equal to the preset threshold, converting the current single-episode draft summary into a structured data object includes: Using the large language model, core feature labels are extracted from the current single-episode draft summary, wherein the core feature labels include a list of core figures, core events, and emotional tone; The current episode draft summary, its episode number, and the core feature tags are organized into a structured data object.

6. The method according to claim 1, characterized in that, The single-episode summary sequence is input into the large language model, so that the large language model, combined with a preset narrative structure guidance prompt, generates a full-episode draft summary of the target video, including: The preset narrative structure guidance prompts are formulated for the target video; The single-episode summary sequence and the preset narrative structure guidance prompts are input into the large language model so that the large language model can comprehensively understand the overall plot of the target video and generate a full-episode draft summary of the target video.

7. The method according to claim 1, characterized in that, A multi-dimensional quality assessment is performed on the entire drama draft summary, and the summary is optimized based on the assessment results to obtain the plot summary of the target video, including: Using the text vector method, the semantic relevance of the full-length original script of the target video and the full-length draft summary are calculated. Using a natural language reasoning model, the factual consistency of the entire drama draft summary and the original script of the entire drama is compared, and the factual consistency degree of the entire drama draft summary is calculated. Based on the list of key plot anchors, the key plot coverage of the entire drama draft summary is evaluated, and the key plot coverage of the entire drama draft summary is calculated. The length and content of the entire script summary are evaluated, and the conciseness of the entire script summary is calculated. An optimization signal is generated based on the semantic relevance, factual consistency, key plot coverage, and conciseness of the entire drama draft summary; The optimized signal is applied to the entire drama draft summary, and the entire drama draft summary is optimized through iterative loops to obtain the plot summary of the target video.

8. An automated plot summary generation device, characterized in that, include: The first generation module is used to segment scenes and identify key plot anchors for each episode of the original script of the target video, and to generate corresponding episode draft summaries using a large language model. The conversion module is used to perform multi-dimensional quality assessment and screening of each single-episode draft summary, and convert it into structured data objects; The sorting module is used to sequentially arrange all data objects to obtain a single-episode summary sequence of the target video; The second generation module is used to input the single episode summary sequence into the large language model, so that the large language model, combined with the preset narrative structure guidance prompts, generates a full-episode draft summary of the target video; The optimization module is used to perform multi-dimensional quality assessment on the entire drama draft summary, and optimize the entire drama draft summary based on the assessment results to obtain the plot summary of the target video.

9. A computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to perform the method described in any one of claims 1 to 7.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the method described in any one of claims 1 to 7 through the computer program.

Images