Presentation generation method

Abstract

The present invention relates to the technical field of automated presentation generation, and specifically to a presentation generation method. A reference presentation and an input document are obtained, wherein the reference presentation includes a structure slide and a content slide; functional clustering is performed on the reference presentation to obtain one to multiple structure slide groups and one to multiple content slide groups; the structure slide group and the content slide group are analyzed to obtain a content pattern framework, wherein the content pattern framework includes content areas, area descriptions, and filling requirements; the input document is extracted to obtain an outline of the presentation to be generated; code instructions are constructed based on the outline and the content pattern framework to generate a new presentation. The present invention can overcome the defects of the prior art in terms of insufficient consistency between visual style and content, greatly improve the overall aesthetics and professionalism of the presentation, and make the generated presentation more coherent in terms of logical hierarchy and visual communication.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic presentation generation, and in particular to a presentation generation method Background Art

[0002] Existing automated presentation generation methods can be divided into two categories: text-based generation methods and layout generation methods based on preset templates.

[0003] 1. Presentation generation based on text content

[0004] Text-based presentation generation methods typically utilize large language models (such as GPT-3 or GPT-4) to extract key information from input documents and generate logically coherent text content. These methods develop dedicated datasets and training models to ensure that the generated presentations have good content logic and information delivery, while also performing well in increasing the diversity of text content. To further improve the layout of content, some methods also introduce additional modules, such as automatically arranging the text and image elements of the presentation through object locators, or using similarity-based strategies to improve the accuracy of text and images, thereby improving the overall effect.

[0005] 2. Presentation generation based on preset templates

[0006] This technical solution automatically layouts content using pre-set layout templates. Typically, these templates include common layout elements such as titles, images, and text boxes, with the goal of quickly meeting users' basic presentation needs. Because pre-set templates are visually designed, they ensure aesthetics to a certain extent and are easy to use, significantly reducing manual design time. This method provides a fast and efficient solution for presentation generation by selecting the appropriate template for content filling.

[0007] In the field of automated presentation generation, existing technologies primarily employ text-based and template-based approaches. While these technologies offer advantages in content logic and generation efficiency, they suffer from inherent flaws in their principles and methods, limiting their practical application.

[0008] 1. Lack of visual consistency and diversity

[0009] Presentation generation methods based on text generation primarily rely on large language models to extract key information from input documents and generate corresponding text content. While this approach ensures the logicality of the content, it neglects visual design requirements and lacks effective control over complex visual layouts. This design mechanism inherently lacks fine-grained visual control during presentation generation, resulting in presentations that lack appeal and consistency. For example, in complex presentation scenarios, single-page text generation cannot maintain the visual harmony and aesthetics of a multi-page presentation, compromising the effectiveness of information communication.

[0010] 2. Insufficient technical efficiency and adaptability

[0011] Existing technologies often struggle to adapt efficiently to changes when processing complex content (such as those containing multimedia elements and charts). For example, template-based approaches require multiple manual interventions and adjustments to adapt to different content and layouts, increasing operational complexity and costs while failing to meet users' demands for rapid generation. Summary of the Invention

[0012] To solve the above problems, the present invention proposes a presentation manuscript generating method, the steps of which include:

[0013] Obtaining a reference presentation and an input document, wherein the reference presentation includes a structure slide and a content slide;

[0014] Performing functional clustering on the reference presentation to obtain one-to-many structure slide groups and one-to-many content slide groups;

[0015] Analyzing the structure slide group and the content slide group to obtain a content model framework, wherein the content model framework includes content areas, area descriptions, and filling requirements;

[0016] Extracting the input document to obtain an outline of a presentation to be generated;

[0017] Code instructions are constructed based on the outline and the content model framework to generate a new presentation.

[0018] Furthermore, functional clustering is performed based on the text features of the reference presentation to obtain a structured slide group.

[0019] Furthermore, functional clustering is performed on the reference presentations by image similarity to obtain a content slide group.

[0020] Furthermore, the content area includes a title part and / or a subtitle part and / or a logo part and / or a footer link and / or license information; and the area description includes text and / or an image.

[0021] Furthermore, the outline includes one or more entries, and each entry corresponds to a slide to be generated.

[0022] Furthermore, the entry includes a reference presentation, an index of an input document section, and a title and description of the slide to be generated.

[0023] Furthermore, the new presentation is generated by a large language model.

[0024] The present invention also provides a presentation generation system, comprising:

[0025] A data acquisition module is used to acquire a reference presentation and an input document, wherein the reference presentation includes a structure slide and a content slide;

[0026] A functional clustering module, configured to perform functional clustering on the reference presentation to obtain one to multiple structure slide groups and one to multiple content slide groups;

[0027] A framework construction module is used to analyze the structure slide group and the content slide group to obtain a content model framework, wherein the content model framework includes content areas, area descriptions and filling requirements;

[0028] An outline generation module, configured to extract the input document and obtain an outline of a presentation to be generated;

[0029] The presentation generating module is used to construct code instructions based on the outline and the content model framework, thereby generating a new presentation.

[0030] The present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing the above method.

[0031] The present invention also provides a storage medium storing a computer program, wherein the computer program implements the above method when executed by a computer.

[0032] The beneficial effects of the present invention are as follows:

[0033] The present invention extracts the structure and layout information of reference presentations and organizes them into functional clusters accordingly. This ensures that when generating a new presentation, appropriate existing presentations can be selected as editing targets. Finally, the content model of each presentation is extracted to ensure that the generated content is consistent with the requirements. By introducing the content model of the presentation, the present invention overcomes the shortcomings of the existing technology in lacking consistency between visual style and content, significantly improving the overall aesthetics and professionalism of the presentation, and making the generated presentation more coherent in terms of logical hierarchy and visual communication.

[0034] The present invention enables a large language model to generate code instructions to modify the content in a reference presentation, greatly reducing the need for manual intervention. This approach can automatically perform content editing operations, improve generation efficiency, and support the automated processing of various media content. Finally, when the large language model generates erroneous code, external feedback is provided to help the large language model achieve self-correction. Through this innovative technical means, the present invention solves the problems of low operating efficiency and high labor costs in the existing technology when generating complex presentations, and is suitable for a variety of application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Flowchart of an embodiment of the present invention.

[0036] Figure 2 This is an example diagram of a reference presentation for an embodiment of the present invention.

[0037] Figure 3 This is an example diagram of a new presentation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present invention are described clearly and completely below. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention. Figure 1 shown.

[0039] S1 Get reference presentation and input documents

[0040] In this step, a reference presentation is obtained, and the selection of the reference presentation is based on the needs of the user.

[0041] In this step, an input document is obtained and used as the input for the presentation to be generated.

[0042] S2 Reference Presentation Analysis

[0043] This step functionally clusters the reference presentations, classifying them into two categories based on their functional attributes: one supporting the presentation structure, such as the opening, transition, and closing pages; and the other conveying specific content, such as the bulleted title page. To achieve effective presentation clustering, this step employs different clustering algorithms based on textual or visual features of the slides.

[0044] For the structural slides, the large language model (LLM) is used to infer the functional role of each slide, and they are grouped according to their unique text features to obtain the structural slide group.

[0045] For content slides, this step uses a hierarchical clustering method based on image similarity, and infers the layout pattern of each cluster using a multimodal large language model (MLLM). This algorithm uses a similarity matrix between slides and a set similarity threshold to effectively group slides using a hierarchical clustering method. First, the algorithm receives a slide similarity matrix consisting of content slides and a similarity threshold of 0.65 as input and initializes an empty set of clusters. During the iteration process, the algorithm continuously selects the two slides with the greatest similarity from the similarity matrix. If either of these two slides already belongs to an existing cluster, they are merged into that cluster. If they do not yet belong to any cluster, a new cluster is created and the two slides are added to it. This process continues until the maximum similarity in the similarity matrix is ​​less than the threshold, and finally returns the set of all clusters to obtain the content slide group.

[0046] Through this algorithm, the slides in the reference presentation can be effectively clustered to assist in the generation of subsequent presentations.

[0047] S3 content pattern analysis

[0048] This step identifies and summarizes common content and structure found in slides and generates corresponding content schemas, providing detailed content layout guidance for new presentations. By analyzing the content types and structure of reference presentations, the system automatically generates a content schema framework that breaks down each slide's content area into specific content types and their descriptions.

[0049] This content model structure includes multiple, well-defined content areas, such as a title, subtitles, logo, footer links, and license information. Each area has a corresponding description, data type (e.g., text or image), and content requirements. The content model framework standardizes the functionality and presentation requirements of each content area, allowing for automatic matching of appropriate content types and data when generating new presentations.

[0050] This step ensures consistency in the presentation's structure, enabling the system to efficiently and accurately populate content. It also improves the presentation's logical hierarchy and the coherence of its information. Through this structured content model, the system can adapt to varying content needs, enabling intelligent presentation content generation and optimized layout.

[0051] S4 outline generation

[0052] In the outline generation step, the outline and main content of the presentation are first constructed by extracting key information from the input document. This process plays a vital role in ensuring that the presentation structure is clear and the information is smoothly transmitted. The outline generation is divided into multiple entries, each entry corresponds to a slide, and the entry content includes the reference presentation, the index of the input document chapter, and the title and description of the slide to be generated. Specifically, the system uses Qwen2 to streamline the input document, splitting the long content into multiple chapters, each chapter contains a title and content paragraphs, and then generates entries. This process can help the large language model (LLM) process only the necessary information, thereby reducing context complexity and improving generation efficiency.

[0053] S5 presentation generation

[0054] Guided by the outline, the slide generation process iteratively edits the reference presentation to generate a new one. To enable efficient interaction between the large language model and the presentation, this paper provides a set of Python APIs for interacting with specific elements in the presentation, including adding, deleting, and editing presentation elements, as shown in Table 1. Furthermore, to facilitate model understanding, this paper parses the presentation and presents it in HTML format, allowing the LLM to more intuitively analyze the presentation structure. For each slide, the content of the new slide is generated in JSON format based on the content of the corresponding entry, guided by the content model framework.

[0055] Next, the LLM receives the slide content to be generated, an HTML representation of the reference presentation, and detailed API documentation to generate executable editing operations, with each operation step represented by a line of executable code. These code instructions are then executed in an interactive execution environment (REPL). During execution, the system detects errors and provides feedback in real time, such as syntax errors or incorrect references to elements in the reference presentation, to help the model self-correct. Through this self-correction mechanism, the system can continuously optimize editing operations using intermediate results, ensuring the robustness of the generation process and the high quality of the presentation content.

[0056] Table 1

[0057] Function Name describe del_span Delete specific passages of text. del_image Deletes an image element. clone_paragraph Create a copy of an existing paragraph. replace_span Replaces the contents of a specific paragraph of text. replace_image Replace the path of the image element

[0058] The reference presentation used in the embodiment of the present invention and the presentation finally generated are as follows: Figure 2 and Figure 3 shown.

[0059] Through the innovative design of the above steps, this invention demonstrates significant technical results in automated presentation generation. To further verify the effectiveness of each module and the causes of errors under different conditions, this invention conducted ablation experiments to compare performance under different configurations. The specific experimental setup and technical results are analyzed as follows:

[0060] Ablation experiment setup

[0061] To further explore the independent contributions of each step in the present invention and its impact on the overall system performance, the present invention designs three ablation configurations, each removing one of the key steps, as follows:

[0062] 1. w / o code rendering

[0063] In this configuration, the system's HTML code rendering is replaced with PPTC-style formatting, removing the HTML code rendering mechanism and directly using a simple PPTC-style format for layout display. This setting is used to verify the role of the code rendering mechanism in content generation.

[0064] 2. No error feedback

[0065] In this configuration, the feedback-driven self-correction mechanism is removed, preventing the system from automatically adjusting and optimizing the generated presentation based on API feedback. This setting is used to evaluate the contribution of the error correction mechanism to generation accuracy and content coherence.

[0066] 3. w / o layout summary

[0067] In this configuration, a slide is randomly selected as a blueprint, rather than selecting a suitable presentation template for layout inference based on the method of the present invention. This setting is used to verify the effectiveness of the layout induction module in ensuring visual consistency.

[0068] These ablation experiment configurations were tested on a combination of Qwen2 processing text input and GPT-4o processing visual input to compare the contribution of each component to the overall system performance. The method of the present invention is PPTAgent.

[0069] Qwen2 is an advanced large language model that focuses on natural language processing tasks and is capable of generating high-quality text content. Qwen2 excels at extracting key information from input documents, generating coherent text paragraphs, and performing well in multi-round dialogue and content generation tasks.

[0070] GPT-4o is an optimized version of GPT-4 that can accept and process visual information, thus performing well in multimodal tasks. In addition to text generation and comprehension capabilities, GPT-4o can also combine visual information (such as images, charts, etc.) with text information to achieve cross-modal content generation and analysis. In the task of automated presentation generation, GPT-4o is responsible for generating text with clear logic and coherent content, and is able to explain or supplement the input visual content. This gives GPT-4o higher accuracy and adaptability in structured information extraction and content generation, ensuring that the generated presentation meets both logical requirements and has a style and information accuracy consistent with the visual content.

[0071] The ablation experiment results are shown in Table 2.

[0072] Table 2

[0073]

[0074] FID is a metric that measures the similarity between generated content and a reference sample in feature space. It evaluates the visual quality of the generated content by comparing the statistical features of the generated and reference content. Lower FID values ​​indicate that the generated content is more similar in visual style to the reference content, thus meeting the requirement of visual consistency. FID is commonly used in image generation tasks, but it is also applicable to assessing visual consistency in presentation generation.

[0075] Fidelity refers to whether the generated presentation is faithful to the input content or template in terms of content and structure. High fidelity means the generated presentation closely matches the input content in terms of information delivery and visual layout, avoiding content deviations or layout errors. This metric is used to assess the accuracy and consistency of generated results.

[0076] G-Eval is a comprehensive evaluation metric that rates the organization, clarity, and coherence of generated content. Scores typically range from 1 to 5, with higher scores indicating superior organization and logical coherence of the generated presentation. G-Eval is suitable for evaluating the quality of text generation, visual layout, and other aspects.

[0077] Experimental results show that our method performs better than methods without the content model framework in terms of visual consistency metric FID (Fréchet Inception Distance). Specifically, the inclusion of the content model framework reduces the FID of the generated presentation from 3.48 to 4.32, indicating that the visual style of the generated presentation is closer to that of the reference presentation.

[0078] In addition, the present invention also performs well in the success rate (SR) of the presentation generation task. For example, the success rate of the method without using the content model framework dropped to 73.0%, and there was also a decline in indicators such as fidelity and G-Eval.

[0079] Success Rate (SR) refers to the percentage of presentations generated by the system that successfully meet the expected content and structure. A higher success rate indicates fewer errors or failures during the generation process, indicating greater system stability and generation capabilities. This metric, expressed as a percentage, measures the reliability of automated generation.

[0080] The method of the present invention converts the generated presentation code into HTML format, so that LLM can accurately understand and adjust the element layout in the presentation; at the same time, the error feedback mechanism allows the method to capture and correct generation errors during the generation process, thereby reducing manual intervention and significantly improving generation efficiency and accuracy.

[0081] Experimental data shows that both HTML code rendering and the error feedback mechanism have a significant impact on the success rate and generation quality of the method. The method that removes code rendering sees its success rate drop from 84.8% to 44.4%, and its G-Eval score drop from 3.72 to 2.64, indicating that the lack of code rendering significantly affects the organization and readability of the content. The method that removes the error feedback mechanism sees its success rate drop from 84.8% to 56.4%, and its G-Eval score drop from 3.72 to 2.97, indicating that the error feedback mechanism also plays an important role in improving generation accuracy and coherence. These two features together ensure that the final generated presentation meets high-quality standards in terms of content organization, visual effects, and coherence.

[0082] The above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Those skilled in the art may modify or make equivalent substitutions for the technical solutions of the present invention without departing from the spirit and scope of the present invention. The scope of protection of the present invention shall be based on the claims.

Claims

1. A method for generating a presentation, comprising the steps of: Obtain reference presentations and input documents; Functional clustering is performed on the reference presentation to obtain one to many structure slide groups and one to many content slide groups; wherein, Functional clustering is performed on the reference presentation to obtain one to many structure slide groups and one to many content slide groups, including: dividing the reference presentation into structure slides and content slides according to functional attributes of the reference presentation; The functional role of each structural slide is inferred using a large language model, and the structural slides are grouped based on the text features of the functional role to obtain one-to-many structural slide groups. The content slides are processed by hierarchical clustering based on image similarity, and the layout pattern of each cluster is inferred by a multimodal large language model to obtain one or more content slide groups; wherein the process of obtaining the one or more content slide groups includes: Generate a similarity matrix of content slides; Based on the similarity matrix, the two content slides with the greatest similarity are selected; If either of the two content slides belongs to a cluster, the two content slides are merged into the cluster; if neither of the two content slides belongs to any cluster, a new cluster is created and the two content slides are added to the new cluster; Returning to the step of selecting two content slides with the greatest similarity based on the similarity matrix, until the maximum similarity in the similarity matrix is ​​less than a set threshold; Analyzing the structure slide group and the content slide group to obtain a content model framework, wherein the content model framework includes content areas, area descriptions, and filling requirements; Extracting the input document to obtain an outline of the presentation to be generated; wherein the outline includes one or more entries, each entry corresponding to a slide to be generated, and the entry content includes a reference presentation, an index of the input document section, and a title and description of the slide to be generated; Constructing code instructions based on the outline and the content model framework to generate a new presentation; wherein constructing code instructions based on the outline and the content model framework to generate a new presentation includes: Generate API documentation, the API documentation including the interaction between the Python API and elements in the reference presentation, the elements including: adding, deleting, and editing presentation elements; Generate an HTML representation of the reference presentation; Generate the content of the slides to be generated in JSON format according to the entry content under the guidance of the content model framework; Feeding the slide content to be generated, the HTML representation of the reference presentation, and the API documentation into the large language model to generate executable editing operations; wherein each operation step in the editing operation is represented by a line of executable code; Executing the code in an interactive execution environment, and when an execution error occurs, feeding the error back to the large language model to correct the error in the code; Generate a new presentation based on the corrected code.

2. The method according to claim 1, characterized in that The content area includes a title part and / or a subtitle part and / or a logo part and / or a footer link and / or license information; and the area description includes text and / or an image.

3. A presentation generation system, comprising: A data acquisition module is used to acquire a reference presentation and an input document, wherein the reference presentation includes a structure slide and a content slide; A functional clustering module is used to perform functional clustering on the reference presentation to obtain one to many structure slide groups and one to many content slide groups; wherein the functional clustering on the reference presentation to obtain one to many structure slide groups and one to many content slide groups includes: dividing the reference presentation into structure slides and content slides according to functional attributes of the reference presentation; The functional role of each structural slide is inferred using a large language model, and the structural slides are grouped based on the text features of the functional role to obtain one-to-many structural slide groups. The content slides are processed by hierarchical clustering based on image similarity, and the layout pattern of each cluster is inferred by a multimodal large language model to obtain one or more content slide groups; wherein the process of obtaining the one or more content slide groups includes: Generate a similarity matrix of content slides; Based on the similarity matrix, the two content slides with the greatest similarity are selected; If either of the two content slides belongs to a cluster, the two content slides are merged into the cluster; if neither of the two content slides belongs to any cluster, a new cluster is created and the two content slides are added to the new cluster; Returning to the step of selecting two content slides with the greatest similarity based on the similarity matrix, until the maximum similarity in the similarity matrix is ​​less than a set threshold; A framework construction module is used to analyze the structure slide group and the content slide group to obtain a content model framework, wherein the content model framework includes content areas, area descriptions and filling requirements; an outline generation module, configured to extract the input document and obtain an outline of the presentation to be generated; wherein the outline includes one or more entries, each entry corresponding to a slide to be generated, and the entry content includes a reference presentation, an index of the input document section, and a title and description of the slide to be generated; A presentation generation module is configured to construct code instructions based on the outline and the content model framework to generate a new presentation; wherein constructing code instructions based on the outline and the content model framework to generate a new presentation includes: Generate API documentation, the API documentation including the interaction between the Python API and elements in the reference presentation, the elements including: adding, deleting, and editing presentation elements; Generate an HTML representation of the reference presentation; Generate the content of the slides to be generated in JSON format according to the entry content under the guidance of the content model framework; Feeding the slide content to be generated, the HTML representation of the reference presentation, and the API documentation into the large language model to generate executable editing operations; wherein each operation step in the editing operation is represented by a line of executable code; Executing the code in an interactive execution environment, and when an execution error occurs, feeding the error back to the large language model to correct the error in the code; Generate a new presentation based on the corrected code. 4 . An electronic device comprising a memory and a processor, wherein the memory stores a computer program, the computer program being configured to be executed by the processor, the computer program comprising instructions for executing the method according to claim 1 .

5. A storage medium storing a computer program, wherein when the computer program is executed by a computer, the method according to any one of claims 1 to 2 is implemented.

Images