A method for automatically generating broadcast videos of corporate financial news based on AIGC

Abstract

This invention discloses a method for automatically generating broadcast videos of corporate financial news based on Artificial Intelligence Generated Content (AIGC). This method extracts key information from corporate financial news using AIGC technology and automatically generates data videos, aiming to improve novice investors' understanding of complex financial information. Furthermore, this invention develops an interactive interface to enhance the controllability of the AI-assisted design process, allowing users to flexibly control the video generation process and adjust the AI-generated video content. Compared with existing technologies, this invention effectively solves the problem of simplifying and conveying complex news narratives containing multiple layers of financial information. By extracting core information from corporate financial news and combining it with a knowledge-based Large Language Model (LLM) to generate video scripts, the generative model is used to generate news images. Simultaneously, the system automatically retrieves relevant corporate financial data and generates easy-to-understand visual charts, eliminating the need for users to manually retrieve relevant information. Finally, these visual elements are automatically laid out and organized into a coherent news broadcast video. This invention helps content creators quickly generate high-quality corporate financial news videos, providing novice investors with comprehensive and in-depth interpretations of corporate financial news, and has high practical value and broad development prospects.

Description

Technical Field

[0001] This invention relates to the field of automatic product generation methods, and in particular to a method for automatically generating and broadcasting corporate financial news videos based on AIGC. Background Technology

[0002] Corporate financial news provides investors with insights into a company's overall financial health and growth prospects, serving as a crucial channel for obtaining company dynamics and financial information. However, traditional corporate financial news is typically lengthy, information-dense, and uses complex financial concepts, increasing the difficulty for novice investors to understand. Data-driven videos significantly improve the ability of non-experts to understand financial information, enhancing the efficiency and effectiveness of information retrieval and insightful understanding.

[0003] Several tools exist for automatically generating information visualizations in financial news, reducing the workload of news creators and lowering the barrier to understanding financial news. However, most of these automated tools are limited to generating visualizations of single financial information, failing to cover the complete news narrative. For example, generating visualizations of company stock market trends. In practical applications, the financial information contained in corporate financial news involves not only company stocks but also unique attributes such as company shareholders, products, and business activities. Furthermore, creating corporate financial news broadcast videos is challenging, requiring news creators to possess both storytelling and data video production skills. Existing tools assisting designers in creating data videos primarily focus on coordinating narrative and visualization animations, guiding users to easily create universally applicable data videos. However, these methods lack support for handling financial narratives and coordinating visual element design, still requiring authors to have high levels of expertise and use independent tools. Currently, there is no method that fully considers the diversity of financial information contained in corporate financial news and automatically generates broadcast videos containing relevant financial data visualizations. The emergence of AIGC (AI-generated content) provides a new avenue for automating the creation of such videos due to their ease of use and powerful generation capabilities. Therefore, we propose a method for automatically generating broadcast videos for corporate financial news based on AIGC. Summary of the Invention

[0004] The purpose of this invention is to propose a method for automatically generating and broadcasting corporate financial news videos based on AIGC to solve the problems in the background technology.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for automatically generating broadcast videos of corporate financial news based on AIGC includes the following:

[0007] S1. Set the data time range for the broadcast video and input the style of the news images through natural language, and input one or more corporate financial news texts;

[0008] S2. Based on a large-scale language model (LLM) of enterprise financial news knowledge base, use prompt word engineering to generate prompt word instructions to reproduce key information in the news;

[0009] S3. Based on the prompt word engineering described in S2, formulate a series of instructions to guide the large language model (LLM) to generate prompt words for the text-to-image generation model;

[0010] S4. Retrieve and match relevant data facts based on the financial attributes of each chapter;

[0011] S5. Based on the visual elements that appear, divide the video clips into narrative clips and visual clips. Apply an element layout algorithm based on saliency detection to lay out the visual elements in the video. Use the BASNet model to extract the saliency map from the original chart and use it as the basis for generating the visual element layout. The laid-out visual elements will be matched with the corresponding animation templates.

[0012] S6. Generate multiple broadcast video clips, automatically splice them together, and output the corporate financial news broadcast video.

[0013] Preferably, the corporate financial news knowledge base described in S2 includes knowledge of financial attributes, the inverted pyramid news paradigm, and a chapter-based organizational structure;

[0014] The financial attributes are eight privately defined attributes that are closely related to corporate financial information; the definition of a financial attribute includes the attribute name, attribute description, financial data facts, and news examples.

[0015] The inverted pyramid news paradigm and chapter-based organizational structure refer to relevant academic papers, where long texts are segmented into semantically meaningful financial knowledge fragments, and text embedding is used to convert text data into continuous vector representations to build a local knowledge base that performs searches on large datasets.

[0016] Preferably, the prompt word instruction in S2 includes four subtasks, specifically including the following:

[0017] S2.1. Divide the text T into n paragraphs p based on the context information. i where i = 1, ..., n;

[0018] S2.2, For each paragraph p i Extract its corresponding financial attribute a i ;

[0019] S2.3, Paragraphs with the same financial attributes {pi ,p j |i,j∈[1,n],a i =a j Reorganized into the same chapter S;

[0020] S2.4 Summarize each chapter S and generate m concise subtitle phrases.

[0021] Preferably, the instructions in S3 require a large language model (LLM) to act as an illustrator, used to construct basic information about the image style, company, and industry;

[0022] The task of the Large Language Model (LLM) is to expand the caption narrative into a specific visual scene, extract appropriate prompts from the scene as additional information, generate the final prompt for each image by retaining the original basic information and modifying the additional information, and input the generated prompt text into the image generative model in sequence to produce a set of images with a consistent style.

[0023] Preferably, the metal attribute retrieval process described in S4 is as follows:

[0024] Retrieve(C,A,T)={(a,f a )|a∈A,f a ∈S(C,T)}

[0025] Where C represents the formula; A represents the set of financial attributes; T represents the time range; a represents the specific attribute; f a S represents the data fact; S represents the data source;

[0026] In the financial attribute set A, each attribute a corresponds to a set of data facts f. a The data fact f a Retrieved from data source S, the data is relevant to company C within time range T; the structured data is converted into visual icons using an encoding template, and narrative icon annotations are generated using a large language model (LLM).

[0027] Preferably, step S5 specifically includes the following:

[0028] S5.1. Use single-layer spatial pyramid pooling technology to reduce the search space, adjust the saliency map to the size of the sub-window, and perform mean pooling operation on each sub-window; select the most significant and least significant pixels from the pooled saliency map as the basis for subsequent processing.

[0029] S5.2 For narrative clips, a template matching algorithm is used to maximize the presentation of salient information in the image; after identifying salient regions, their spatial distribution characteristics are analyzed. Based on the spatial distribution characteristics, the system classifies salient regions and selects a suitable horizontal or vertical template; for salient regions that cover the image, the system automatically switches to a full-screen template and applies masking technology to ensure the clarity and visual effect of the title when displayed in full screen.

[0030] S5.3 For visual clips, the layout of their visual elements (charts and text) follows three rules: no overlap, readability, and proximity.

[0031] The term "non-overlapping" means that the text annotation boxes do not interfere with the key elements of the chart. The significance threshold is determined based on the highest significance value among the least significant pixels to ensure that the candidate area does not exceed the threshold.

[0032] The readability refers to the appropriate size of the text annotation box, making it easy to read; each least significant pixel is expanded horizontally and vertically into a candidate region until a pixel exceeding the significance threshold is encountered, and the region that meets the size criteria is retained, the coordinates are mapped back to the original scale, and non-maximum suppression is applied to remove redundant candidate regions; the selection of the optimal text annotation box is based on horizontal alignment, proximity to the golden ratio, and maximum size. After calculating the ratio index, the annotation box with the smallest ratio is selected as the optimal annotation box.

[0033] The proximity refers to placing text annotation boxes near important areas to enhance visibility and relevance, and identifying the four nearest significant pixels in each corner to fine-tune the position of the optimal annotation box.

[0034] Compared with existing technologies, this invention provides a method for automatically generating broadcast videos of corporate financial news based on AIGC, which has the following beneficial effects:

[0035] This invention proposes a method for automatically generating broadcast videos of corporate financial news based on AIGC (AI-Generated Content), aiming to reduce the workload of news creators and lower the information comprehension threshold of financial news. This invention creatively proposes eight key financial attributes in corporate financial news, cleverly utilizes the natural language processing capabilities of LLM (Limited Language Management) to extract key insights from the news, generates news images through a text-to-image generative model, automatically retrieves data facts, creates visualizations containing financial information, and organizes them into news broadcast videos. This automated method not only allows users to create high-quality corporate financial news broadcast videos more quickly, but also allows viewers to gain a more intuitive and comprehensive understanding of corporate financial conditions, filling a gap in existing research in this field. Attached Figure Description

[0036] Figure 1This is a schematic diagram of the process for automatically generating corporate financial news broadcast videos as mentioned in Embodiment 1 of the present invention;

[0037] Figure 2 These are the eight financial attributes and their corresponding financial data facts mentioned in Embodiment 3 of this invention;

[0038] Figure 3 This is a schematic diagram illustrating an example of generating a broadcast video from corporate financial news, as mentioned in Embodiment 3 of the present invention. Detailed Implementation

[0039] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0040] Example 1:

[0041] This invention proposes a framework called NewsVis (News Visualization), which utilizes Artificial Intelligence Generated Content (AIGC) technology to extract key information from corporate financial news and automatically match relevant financial data to generate news broadcast videos. The framework comprises three modules: news recapitulation, visualization recommendation, and visual element organization. The news recapitulation module aims to automatically identify financial attributes and reorganize the narrative, while simultaneously generating news images using a text-to-image generative model. The visualization recommendation module can automatically retrieve relevant industry and company earnings data based on the extracted financial attributes and generate expressive annotated visualizations. The visual element organization module focuses on a heterogeneous visual element layout algorithm based on saliency detection and uses animation templates to organize elements into a coherent data video. Furthermore, this invention also develops an interactive interface that allows users to adjust and improve the AI-generated results.

[0042] See Figure 1 The present invention automates the generation and broadcasting of corporate financial news videos according to the following steps:

[0043] Step 1: Set the time range and image style, and enter the corporate financial news text;

[0044] Step 2: Divide the task into four sub-tasks to reproduce the key information in the news;

[0045] Step 3: Apply a text-to-image generative model to key information to generate news images;

[0046] Step 4: Retrieve and match relevant data facts to generate annotated data visualizations;

[0047] Step 5: Apply layout algorithms to visual elements and match animation templates;

[0048] Step 6: Generate multiple video clips and output the spliced ​​broadcast video.

[0049] Example 2:

[0050] Based on Example 1 but with a difference, this invention proposes a method for automatically generating broadcast videos of corporate financial news based on AIGC, specifically including the following:

[0051] Step a:

[0052] Set the data time range for the broadcast video and input the style of the news images using natural language, and input one or more corporate financial news texts;

[0053] Step b:

[0054] Supported by a Large Language Model (LLM) based on a corporate financial news knowledge base, key information in news articles is reproduced using cue word engineering. The corporate financial news knowledge base is characterized by encompassing knowledge in the following areas: financial attributes, an inverted pyramid news paradigm, and a chapter-based organizational structure. The financial attributes are eight privately defined attributes closely related to corporate financial information (such as...). Figure 2 (As shown). The definition of each financial attribute includes the attribute name, attribute description, financial data facts, and news examples; the knowledge of the inverted pyramid news paradigm and chapter-based organizational structure is based on relevant academic papers. Long texts are segmented into semantically meaningful financial knowledge fragments, and text embedding is used to convert text data into continuous vector representations, establishing a local knowledge base capable of efficiently performing searches on large datasets. The prompt word instruction includes four sub-tasks, with the specific steps as follows:

[0055] First, divide the text T into n paragraphs p based on the context information. i , where i = 1, ..., n; for each paragraph p i Extract its corresponding financial attribute a i Next, paragraphs with the same financial attributes {p} will be... i ,p j |i,j∈[1,n],a i =a j Reorganize them into the same chapter S. Finally, summarize each chapter S to generate m concise subtitle phrases.

[0056] Step c:

[0057] A set of instructions is defined to guide the LLM (Limited Language Builder) in generating cue words for a text-to-image generative model. These instructions require the LLM to act as an illustrator, focusing on fundamental information that constitutes the image style, company, and industry. The LLM's task is first to expand the caption narrative into a specific visual scene, then to extract the scene to generate appropriate cue words as additional information. The final cue for each image is generated by retaining the original basic information and modifying the additional information. The generated cue words are then sequentially input into the text-to-image generative model to produce a set of images with a consistent style.

[0058] Step d:

[0059] The retrieval process, which involves searching and matching relevant data facts based on the financial attributes of each chapter, can be described as Retrieve(C,A,T)={(a,f a )|a∈A,f a ∈S(C,T)}, where each attribute a in the financial attribute set A corresponds to a set of data facts f. a These data facts were retrieved from data source S and are relevant to company (C) within a time frame (T). The structured data was transformed into visual charts using a carefully designed coding template, and narrative chart annotations were generated using LLM.

[0060] e-step:

[0061] The video clips are divided into narrative clips and visual clips based on the visual elements present. An element placement algorithm based on saliency detection is applied to place the visual elements in the video. A BASNet model is used to extract a saliency map from the original graph, which is then used as the basis for generating the visual element placement. Specifically, this includes the following:

[0062] (1) The search space is reduced by using a single-layer spatial pyramid pooling technique. The saliency map is adjusted to the size of the sub-window, and mean pooling is performed on each sub-window. From the pooled saliency map, the most significant and least significant pixels are selected as the basis for subsequent processing.

[0063] (2) The narrative editing employs a precise template matching algorithm to maximize the presentation of salient information in the image. After identifying salient regions, their spatial distribution characteristics, including aspect ratio and area, are analyzed. Based on these characteristics, the system classifies salient regions and selects appropriate horizontal or vertical templates. For salient regions that cover most of the image, the system automatically switches to a full-screen template and applies masking technology to ensure the clarity and visual effect of the title when displayed in full screen.

[0064] (3) The layout of visual elements (charts and text) in the visualization clip follows three rules: no overlap, readability, and proximity. No overlap is crucial because text annotation boxes should not interfere with the chart. A salience threshold is determined based on the highest salience value among the least significant pixels, ensuring that candidate areas do not exceed this threshold. Readability means the size of the text annotation box should be appropriate, preferably horizontal, to ensure easy reading. Each least significant pixel is expanded horizontally and vertically into a candidate area until a pixel exceeding the salience threshold is encountered. Areas meeting the size criteria are retained, and their coordinates are mapped back to the original scale. Non-maximum suppression is applied to remove redundant candidate areas, retaining sufficiently large areas. The selection of the optimal text annotation box is based on factors such as horizontal alignment, proximity to the golden ratio, and maximum size. After calculating the ratio index, the annotation box with the smallest ratio is selected as the optimal annotation box. Proximity means the text annotation box should be placed near important areas to enhance visibility and relevance. The four nearest significant pixels in each corner are identified to fine-tune the position of the optimal annotation box, making the box closer to the significant area and improving the aesthetics of the annotation layout. The laid-out visual elements will match the corresponding animation template.

[0065] Step f:

[0066] Generate multiple broadcast video clips, automatically splice and edit multiple video clips, and output corporate financial news broadcast videos.

[0067] Example 3:

[0068] Based on Embodiments 1-2, but with a difference, the following uses a corporate financial news broadcast video as an example to illustrate the method for automatically generating broadcast videos of corporate financial news based on AIGC proposed in this invention. The specific content is as follows:

[0069] Step 1: Input a financial news article about company C (e.g., ...) as a TXT text file. Figure 3 (As shown in the left part), and set the time range and image style;

[0070] Step 2: With the support of LLM based on the enterprise financial news knowledge base, the key information in N is reproduced using prompt word engineering. The reproduction process includes: paragraph division, attribute extraction, paragraph merging, and summary captions.

[0071] Step 3: Input the generated captions and image styles into the LLM to generate prompts for image generation, and use the text-to-image generative model to generate news images;

[0072] Step 4: Based on the definition of financial attributes (e.g.) Figure 2As shown), relevant financial data facts of enterprise C within a set time range are retrieved from the data source and transformed into structured data. Visual charts are generated through coding templates, and narrative chart annotations are generated using LLM.

[0073] Step 5: Divide the video clips into narrative clips and visual clips based on the visual elements that appear. Apply an element layout algorithm based on saliency detection to lay out the visual elements in the video and match the corresponding animation templates;

[0074] Step 6: Generate multiple video clips and output the spliced ​​broadcast video generated by company C's financial news N (e.g., ...). Figure 3 (As shown in the right part).

[0075] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for automatically generating broadcast videos of corporate financial news based on AIGC, characterized in that, The method specifically comprises the following steps: S1. Set the data time range for the broadcast video and input the style of the news images through natural language, and input one or more corporate financial news texts; S2. Based on a large-scale language model of enterprise financial news knowledge base, use prompt word engineering to generate prompt word instructions to reproduce key information in the news; S3. Based on the prompt word engineering described in S2, formulate a series of instructions to guide the large language model to generate prompt words for the text-to-image generation model; S4. Retrieve and match relevant data facts based on the financial attributes of each chapter; S5. Divide the video clips into narrative clips and visual clips according to the visual elements that appear. Apply an element layout algorithm based on saliency detection to lay out the visual elements in the video. Use the BASNet model to extract the saliency map from the original chart and use it as the basis for generating the visual element layout; specifically, this includes the following: S5.

1. Use single-layer spatial pyramid pooling technology to reduce the search space, adjust the saliency map to the size of the sub-window, and perform mean pooling operation on each sub-window; select the most significant and least significant pixels from the pooled saliency map as the basis for subsequent processing. S5.2 For narrative clips, a template matching algorithm is used to maximize the presentation of salient information in the image; after identifying salient regions, their spatial distribution characteristics are analyzed. Based on the spatial distribution characteristics, the system classifies salient regions and selects a suitable horizontal or vertical template; for salient regions that cover the image, the system automatically switches to a full-screen template and applies masking technology to ensure the clarity and visual effect of the title when displayed in full screen. S5.3 For visual editing, the layout of visual elements follows three rules: no overlap, readability, and proximity. The term "non-overlapping" means that the text annotation boxes do not interfere with the key elements of the chart. The significance threshold is determined based on the highest significance value among the least significant pixels to ensure that the candidate area does not exceed the threshold. The readability refers to the appropriate size of the text annotation box, making it easy to read; each least significant pixel is expanded horizontally and vertically into a candidate region until a pixel exceeding the significance threshold is encountered, and the region that meets the size criteria is retained, the coordinates are mapped back to the original scale, and non-maximum suppression is applied to remove redundant candidate regions; the selection of the optimal text annotation box is based on horizontal alignment, proximity to the golden ratio, and maximum size. After calculating the ratio index, the annotation box with the smallest ratio is selected as the optimal annotation box. The proximity refers to placing text annotation boxes near important areas to enhance visibility and relevance, and identifying the four nearest significant pixels in each corner to fine-tune the position of the optimal annotation box; S6. Generate multiple broadcast video clips, automatically splice them together, and output the corporate financial news broadcast video.

2. The method for automatically generating broadcast videos of corporate financial news based on AIGC according to claim 1, characterized in that, The corporate financial news knowledge base described in S2 includes knowledge of financial attributes, the inverted pyramid news paradigm, and a chapter-based organizational structure. The financial attributes are eight privately defined attributes that are closely related to corporate financial information; the definition of a financial attribute includes the attribute name, attribute description, financial data facts, and news examples. The inverted pyramid news paradigm and chapter-based organizational structure refer to relevant academic papers, where long texts are segmented into semantically meaningful financial knowledge fragments, and text embedding is used to convert text data into continuous vector representations to build a local knowledge base that performs searches on large datasets.

3. The method for automatically generating broadcast videos of corporate financial news based on AIGC according to claim 2, characterized in that, The prompt word instruction described in S2 includes four subtasks, specifically the following: S2.1, Based on context information, extract the text. T Divided into n paragraph p i ,in i =1,…, n ; S2.2, For each paragraph p i Extract its corresponding financial attributes a i ; S2.3, Paragraphs with the same financial attributes Reorganized into the same chapter S ; S2.4, For each chapter S Summarize and generate m A concise subtitle phrase.

4. The method for automatically generating broadcast videos of corporate financial news based on AIGC according to claim 3, characterized in that, The instructions described in S3 require a large language model to act as an illustrator, used to construct basic information about image style, company, and industry; The task of the large language model is to expand the subtitle narrative into a specific visual scene and extract appropriate prompts from the scene as additional information. By retaining the original basic information and modifying additional information, the final prompt for each image is generated. The generated prompt text is then sequentially input into the image generative model to produce a set of images with a consistent style.

5. The method for automatically generating broadcast videos of corporate financial news based on AIGC according to claim 4, characterized in that, The process of retrieving financial attributes as described in S4 is as follows: in, C Representing the formula; A Represents a set of financial attributes; T Indicates a time range; a Indicates a specific attribute; f a Representing data facts; S Indicates the data source; Financial Attribute Set A Each attribute a Corresponding to a set of data facts f a The data facts f a From data source S The information was obtained through a search and is related to the company. C Within the time range T Internal correlation; structured data is converted into visual icons through encoding templates, and narrative icon annotations are generated using a large language model.

Images