Marketing short video generation method and device based on large language model

Abstract

The embodiment of the invention provides a marketing short video generation method and device based on a large language model, and the method comprises the steps: carrying out the matching degree analysis of a hot title and product data based on the large language model, and outputting a hot topic which is most matched with a product according to the sequence of the matching degree from high to low; searching a hot video related to the most matched hot topic in a video platform, and extracting an oral copy in the hot video based on an automatic voice recognition technology; generating a creation copywriting by using a large language model based on the oral copywriting, the hot topics and the product data; and performing similarity matching on the embedded vector corresponding to the created copywriting and the semantic vector of the video slice in the database, and sorting and generating a mixed video meeting the duration requirement according to the similarity.

Description

Technical Field

[0001] Embodiments of the present disclosure relate to the field of artificial intelligence technology. Specifically, it relates to a method and device for generating marketing short videos based on large language models. Background Art

[0002] On today's social media and video platforms, "riding on the coattails of hot topics" short videos have become a very popular and effective brand promotion method. This method uses current affairs hot topics, popular trends or social events to attract users' attention, and usually can obtain higher play counts and completion rates than ordinary short videos.

[0003] On the one hand, since the selection of hot events and topics often requires judging their fit with the brand and the audience, however, this judgment process is highly subjective. Different creators have different understandings and criteria, resulting in the selected hot topics not necessarily being effective or fitting the target audience. On the other hand, from obtaining hot videos, editing to creating a script and voiceover copy in combination with the brand, the whole process often requires a large amount of manpower and time, which easily leads to low efficiency and uneven content quality. Moreover, hot events and topics change rapidly, especially the hot topics on social media. Creators may find that the hot topic has shifted or become outdated during the creation process, resulting in the video content no longer being timely or attractive. Summary of the Invention

[0004] Embodiments described herein provide a method and device for generating marketing short videos based on large language models, and a computer-readable storage medium storing a computer program.

[0005] According to a first aspect of the present disclosure, there is provided a method for generating a marketing short video based on a large language model, including: performing a matching degree analysis on a hot topic title and product information based on the large language model, and outputting the hot topic that best matches the product in descending order of the matching degree; searching for popular videos related to the best-matched hot topic in a video platform, and extracting the voiceover copy in the popular videos based on automatic speech recognition technology; generating a creative copy based on the voiceover copy, hot topic and product information using the large language model; and performing a similarity matching between the embedding vector corresponding to the creative copy and the semantic vector of the video slices in the database, sorting according to the similarity, and generating a marketing short video that meets the duration requirements.

[0006] In some embodiments of the present disclosure, a matching degree analysis is performed on hot topic titles and product materials based on a large language model, and the hot topics most matching the product are output in descending order of matching degree, including: obtaining current hot topic titles and click-through rates from multiple news or media platforms; speculating and expanding hot content related to the hot topic titles; invoking a knowledge base to deeply understand brand or product materials, determining product advantages, selling points, target audiences, and the ways of combining with hot content; performing an association matching degree analysis on the hot topic titles and the product based on the hot content, product advantages, selling points, target audiences, and the ways of combining with hot content; and outputting the hot topic titles most relevant to the product and corresponding annotation information according to the sorting result in descending order of matching degree.

[0007] In some embodiments of the present disclosure, speculating and expanding hot content related to the hot topic titles includes: performing word segmentation on the hot topic titles, and extracting keywords in the hot topic titles from the word segmentation results; using a word vector model to calculate words with similar semantics to the keywords, and expanding the hot content according to the words with similar semantics and the context information of the keywords; and querying the background of hot events using a pre-constructed knowledge base or a database in the relevant field, and speculating potential hot content through an inference model.

[0008] In some embodiments of the present disclosure, performing an association matching degree analysis on the hot topic titles and the product based on the hot content, product advantages, selling points, target audiences, and the ways of combining with hot content includes: comparing the audience groups of the hot content with the target audiences of the brand products; determining whether the emotional atmosphere of the hot events is consistent with the core values and product selling points of the brand; and performing a content fusion degree analysis on the hot events and product selling points to determine whether the selling points of the product can form a natural connection with the content of the hot events.

[0009] In some embodiments of the present disclosure, generating creative copy using a large language model based on voiceover copy, hot topics, and product materials includes: using the product or brand name, product materials, hot topics, and voiceover copy extracted from popular videos as parameters of the LLM model; invoking a knowledge base and generating creative copy based on the formulated copy output rules, where the copy output rules include copy word count, format, emotion, and semantic rules.

[0010] In some embodiments of the present disclosure, the similarity matching between the embedding vector corresponding to the created copywriting and the semantic vector of the video slices in the database, sorting according to the similarity, and generating a marketing short video that meets the duration requirements include: converting the created copywriting and the video slices into embedding vectors respectively based on the text vector embedding technology; performing similarity matching between the embedding vector corresponding to the created copywriting and the embedding vectors of the video slices, sorting from high to low according to the similarity, and selecting the most matching multiple video slices; and converting the created copywriting into speech through a text-to-speech engine, and mixing and synthesizing the speech and the multiple video slices into a final video.

[0011] In some embodiments of the present disclosure, converting the created copywriting and the video slices into embedding vectors respectively based on the text vector embedding technology includes: performing vectorization processing on the copywriting using a pre-trained text embedding model to obtain the embedding vector corresponding to the created copywriting; dividing the videos in the database into multiple video slices, each video slice containing visual information and audio information; and using a visual model to extract the visual and audio features in the video slices, and converting the visual and audio features into the embedding vectors of the video slices.

[0012] In some embodiments of the present disclosure, converting the created copywriting into speech through a text-to-speech engine, and mixing and synthesizing the speech and the multiple video slices into a final video includes: adjusting the speech duration according to the duration of the video slices, and synthesizing the selected video slices in chronological order according to the sorting result of the similarity; and performing mixed editing and synthesis to generate a video that matches the created copywriting according to the generated speech duration and the content of the video slices.

[0013] According to the second aspect of the present disclosure, there is provided a marketing short video generation device based on a large language model. The device includes at least one processor; and at least one memory storing a computer program. When the computer program is executed by the at least one processor, the device is enabled to: perform a matching degree analysis on the hot titles and product materials based on the large language model, and output the hot topics that are most matched with the product in descending order of the matching degree; search for popular videos related to the most matched hot topics in a video platform, and extract the oral copywriting in the popular videos based on the automatic speech recognition technology; generate created copywriting using the large language model based on the oral copywriting, hot topics, and product materials; and perform similarity matching between the embedding vector corresponding to the created copywriting and the semantic vector of the video slices in the database, sort according to the similarity, and generate a marketing short video that meets the duration requirements.

[0014] In some embodiments of the present disclosure, when executed by at least one processor, a computer program causes a device to perform the following operations to perform a matching degree analysis on hot topic titles and product materials based on a large language model, and output the hot topics most matching the product in descending order of matching degree: obtain current hot topic titles and click-through rates from multiple news or media platforms; speculate and expand hot content related to the hot topic titles; call a knowledge base to deeply understand brand or product materials, determine product advantages, selling points, target audiences, and the ways to combine with the hot content; perform an association matching degree analysis on the hot topic titles and the products based on the hot content, product advantages, selling points, target audiences, and the ways to combine with the hot content; and output the hot topic titles most relevant to the product and corresponding annotation information in the sorting result in descending order of matching degree.

[0015] In some embodiments of the present disclosure, when executed by at least one processor, a computer program causes a device to perform the following operations to generate creative copy based on voiceover copy, hot topics, and product materials using a large language model: use the product or brand name, product materials, hot topics, and voiceover copy extracted from popular videos as parameters of the LLM model; call a knowledge base and generate creative copy based on the formulated copy output rules, and the copy output rules include copy word count, format, emotion, and semantic rules.

[0016] In some embodiments of the present disclosure, when executed by at least one processor, a computer program causes a device to perform the following operations to perform similarity matching between the embedding vectors corresponding to the creative copy and the semantic vectors of video slices in a database, sort according to the similarity, and generate a marketing short video meeting the duration requirements: convert the creative copy and video slices into embedding vectors respectively based on text vector embedding technology; perform similarity matching between the embedding vectors corresponding to the creative copy and the embedding vectors of the video slices, sort in descending order of similarity, and select the most matching multiple video slices; and convert the creative copy into voice through a text-to-speech engine, and mix and synthesize the voice and multiple video slices into a final video.

[0017] According to a third aspect of the present disclosure, there is provided a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the steps of the method for generating a marketing short video based on a large language model according to the first aspect of the present disclosure are implemented.

[0018] A method and apparatus for generating marketing short videos based on a large language model according to an embodiment of the present disclosure uses the large language model to perform matching analysis on product materials and hot titles, and can automatically extract the hottest topics most relevant to the product from a vast amount of content, which can not only improve the efficiency of video creation, but also ensure that the video content keeps up with the trend, has high market relevance and user attention. By using automatic speech recognition technology to extract the voice-over copy from popular videos, the process of manual recording and script writing is omitted. By analyzing and learning from the content of existing popular video copy, it is ensured that the newly generated video is more in line with the user's acceptance habits and platform trends. Using the large language model to combine the voice-over copy, hot topics and product materials to generate creative copy, and matching the creative copy with relevant video clips, automated and accurate video clip screening can be achieved, and a mixed video that meets the duration requirements and has consistent emotional expression can be generated, greatly improving the efficiency and quality of marketing short video generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] To briefly describe the technical solutions of the embodiments of the present disclosure more clearly, the accompanying drawings of the embodiments will be briefly described below. It should be understood that the following described drawings only relate to some embodiments of the present disclosure and do not limit the present disclosure, where:

[0020] Figure 1 FIG. shows an exemplary flowchart of a method 100 for generating a marketing short video based on a large language model according to an embodiment of the present disclosure;

[0021] Figure 2 is a schematic diagram of the marketing short video generation process according to an embodiment of the present disclosure;

[0022] Figure 3 is a schematic block diagram of an apparatus 300 for generating a marketing short video based on a large language model according to an embodiment of the present disclosure.

[0023] It should be noted that the elements in the drawings are schematic and not drawn to scale. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] To make the objectives, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions of the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present disclosure. All other embodiments obtained by those skilled in the art based on the described embodiments of the present disclosure without creative efforts shall also fall within the scope of protection of the present disclosure.

[0025] Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains. Further will be understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the specification and the relevant art, and will not be interpreted in an idealized or overly formal form unless expressly so defined herein.

[0026] In view of the problems of low efficiency in creating marketing short videos and the inability to accurately combine hot trends and product selling points, embodiments of the present disclosure propose a method and apparatus for generating marketing short videos based on a large language model, which have significant advantages in improving creation efficiency, reducing labor costs, and enhancing the market relevance and accuracy of the created content. Figure 1 FIG. 5 shows an exemplary flowchart of a method 100 for generating a marketing short video based on a large language model according to an embodiment of the present disclosure.

[0027] At Figure 1 block S102, a matching degree analysis is performed on the hot topic titles and product materials based on the large language model, and the hot topics most matching the product are output in descending order of the matching degree.

[0028] First, current hot events are obtained through various channels such as news platforms, search engines, social media, and trend analysis tools. The titles and click-through rates of current hot events are obtained from multiple news or media platforms (such as Toutiao, Douyin, Weibo, etc.). For example, for the hot events on Toutiao, the interface address of the hot data can be obtained by analyzing requests through a packet capture tool (such as Fiddler or Charles), the JSON data returned by the API is identified, and information such as the event title and click-through rate is parsed. The acquisition of Douyin hot data can directly crawl the list of recommended videos on its home page, and for dynamically loaded content, Selenium or Puppeteer can be used to simulate browser behavior to obtain information such as video titles, play counts, like counts, and comment counts. The hot list of Weibo can be obtained through the "Hot Search List" page, and the page data is crawled by parsing a specific URL, and the event titles and relevant click-through rates or popularity metrics of the top 50 in the ranking are parsed.

[0029] In some embodiments of the present disclosure, the crawled data is stored in a local database or a cloud database. Python data visualization libraries (such as Matplotlib, Plotly, Seaborn, etc.) can be used to display the ranking list. For example, the events are sorted according to the click-through rate, and trend charts, bar charts, etc. of the hot events are made.

[0030] Speculate and expand hot content related to the hot topic title. The hot topic title can be segmented, and the keywords that best represent the core meaning of the title can be extracted from the segmentation results. For example, for the title: "Artificial intelligence technology will completely change the world economy in 2025", the possible keywords extracted may include: artificial intelligence (a high-frequency and representative topic word), technology (a specific field keyword), economy (an impact field keyword).

[0031] Use a word vector model to calculate words with similar semantics to the keywords, and expand the hot content based on the words with similar semantics and the context information of the keywords. For example, starting from the keyword "artificial intelligence", the related hot content that can be expanded is: artificial intelligence applications, trends of artificial intelligence, the impact of artificial intelligence on fields such as economy, education, and healthcare.

[0032] To more accurately speculate on hot content, in addition to directly expanding keywords, relevant content can also be generated using semantic reasoning. Use a pre-constructed knowledge base or a database in the relevant field to query the background of hot events, and speculate on potential hot content through an inference model. Through natural language generation (NLG) or text summarization technology, based on the expanded keywords, generate hot content in a related field. For example, the title "A certain star attended a fashion event, triggering a heated online discussion" may be related to fashion events, star effects, and online public opinion. "Digital currency" may involve topics such as blockchain technology and financial regulation.

[0033] At the same time, based on a deep understanding of brand or product materials by a large language model, determine the product advantages, selling points, target audience, and the way to combine with hot content. Brand or product materials can be input, and the product materials include brand name, core function, target audience, market positioning, product advantages, selling points, usage scenarios, customer feedback, competitor analysis, etc.

[0034] Then, based on the hot content, product advantages, selling points, target audience, and the way to combine with hot content, conduct an analysis of the correlation matching degree between the hot topic title and the product, and output the hot topic titles most relevant to the product and the corresponding annotation information according to the sorting result from high to low matching degree.

[0035] When conducting a correlation analysis between the selling points and hot content, it is necessary to consider the matching degree between the content of the hot event and the selling points of the brand product from multiple dimensions.

[0036] According to some embodiments of the present disclosure, the audience of the hot content, including age, gender, occupation, hobbies, consumption habits, etc., is analyzed and compared with the target audience of the brand product to ensure that the target groups of the two are highly overlapped. For example, if the product is aimed at a young user group (such as trendy, high-tech watches), and the hot content is related to a social topic or trend preferred by young people, then the product selling point is highly consistent with the audience of the hot spot.

[0037] In other embodiments, it is determined whether the emotional atmosphere of a hot event is consistent with the brand's core values ​​and product selling points. For example, some brands use sports events to enhance their brand image and attract sports enthusiasts. If a brand advocates healthy living and the product has selling points of enhancing physical fitness or helping health management, this overlap in values ​​can effectively increase the attention paid to the product.

[0038] In some other embodiments, the content integration degree of hot events and product selling points is analyzed to determine whether the product selling points can form a natural connection with the content of the hot events. For example, if a product selling point involves technological innovation or fashion trends, whether it can be consistent with the current technological development or popular culture hot spots. Analyze whether the hot content can enhance the dissemination effect of the product selling point, that is, whether the hot spot has an expansion or amplification effect, so that the product selling point is strengthened among the target audience.

[0039] To sum up, assuming that the brand has launched an environmentally friendly smart home product with AI smart technology, it can be analyzed through the following dimensions:

[0040] Audience consistency: Assuming that the current hot topic is about the future development of the global green environmental protection and smart home industry, the audience of this hot topic is mainly young people with strong environmental awareness and consumers interested in smart home technology, and the target audience of the product happens to cover these groups, so the audience is highly consistent.

[0041] Content integration: Environmental protection and smart home technology are two mutually reinforcing trends, and the product's selling points and hot events can be effectively integrated. Consumers not only pay attention to the intelligence and convenience of home furnishings, but also have a high interest in the technical features of energy conservation, environmental protection and carbon emission reduction. Products can enhance the attractiveness of this hot topic.

[0042] Value relevance: As global environmental protection and smart home technologies mature, consumers are increasingly demanding environmental protection and smart living. While emphasizing green living and smart convenience, the product also meets consumers' emotional needs for environmental responsibility and technological innovation.

[0043] Through the analysis of three major dimensions: audience consistency, content integration, and value relevance, it is possible to determine whether the product selling points can be effectively emphasized or play a role in the context of current hot topics. If the product has a high degree of fit with the hot topic, it can utilize the hot event to enhance brand exposure and consumer recognition, thereby effectively promoting the product.

[0044] Finally, output the hot topic titles and corresponding annotation information that are most relevant to the product according to the sorting result from high to low in terms of matching degree.

[0045] For example, prioritize the hot topic titles according to the integration degree and attractiveness between the hot topic and the product selling points, and give priority to processing the hot topics that highly match the product selling points. Output a list in the format of a Json Array containing all hot topic titles. The json object has 3 fields, namely: title (the relevant hot topic title), reason (the reason for the association), and relationValue (the matching degree).

[0046] Suppose the product selling points are environmental protection, smart home technology, and health management, and the fields covered by the hot events are environmental protection trends, smart technology progress, healthy living, and sustainable development, etc. The following are the exemplary selected hot topic titles and their association analyses:

[0047] [{"title": "Global Green Environmental Protection Initiative: Sustainable Living Becomes the Mainstream Trend"; "reason": "This hot topic highly matches the product's environmental protection selling point, focuses on the sustainable lifestyle, and conforms to the product's environmental protection characteristics."; "relationValue": "95%"},

[0048] {"title": "The Future of the Smart Home Industry: AI Technology Drives the Home Revolution"; "reason": "Smart home and AI technology are one of the core selling points of the product, and the hot topic is highly consistent with the product's technical characteristics."; "relationValue": "90%"},

[0049] {"title": "Environmental Protection Technology: The Integration of Smart Devices and Green Life"; "reason": "The integration of environmental protection and technology is exactly one of the unique selling points of the product. This hot topic emphasizes the combination of environmental protection and smart technology, which is suitable for highlighting the product's advantages."; "relationValue": "88%"},

[0050] {"title": "AI and the Internet of Things: How to Change the Smart Home Usage Experience"; "reason": "The product relies on AI technology and the Internet of Things, and the hot topic is closely related, which can increase consumers' interest in the product."; "relationValue": "87%"},

[0051] It should be noted that there is a semicolon error in the "title" field value in the original text at line . It should be a comma to be a correct JSON object. The corrected English translation of that line would be: [{"title": "Global Green Environmental Protection Initiative: Sustainable Living Becomes the Mainstream Trend", "reason": "This hot topic highly matches the product's environmental protection selling point, focuses on the sustainable lifestyle, and conforms to the product's environmental protection characteristics.", "relationValue": "95%"},{"title": "Global Health Living Trends in 2025: Smart Devices Facilitate Health Management"; "reason": "The product emphasizes health management functions. This hot topic focuses on the role of smart devices in health management, which is directly related to the health-related selling points of the product."; "relationValue": "85%"},

[0052] {"title": "The Future of Smart Home: New Trends in Enhancing Home Living Quality"; "reason": "This hot topic focuses on the development of smart home and its improvement of living quality, which has a relatively high matching degree with the selling points of the product."; "relationValue": "82%"},

[0053] {"title": "Breakthroughs in Health Technology: Smart Devices Facilitate Health Management"; "reason": "The hot topic focuses on the application of smart devices in the field of health management, which is highly relevant to the health function selling points of the product."; "relationValue": "80%"}]

[0054] The above hot topic titles are arranged in descending order of the relationValue correlation degree, all of which are highly attractive and can resonate with the selling points of the product, thus playing a greater role in marketing and promotion. By providing a short and clear explanation for each selected hot topic title to explain its relevance to the product selling points, it can help with subsequent copywriting creation.

[0055] Subsequently, at Figure 1 in frame S104, search for popular videos related to the most matching hot topic in the video platform, and extract the oral scripts in the popular videos based on automatic speech recognition technology.

[0056] In an embodiment of the present disclosure, after multiple hot topics are matched, select one of the most relevant and potential hot topics for in-depth discussion. The key to this step is to select a hot topic that highly fits the product characteristics and target audience. Use this hot topic to search on short video platforms such as Douyin and YouTube. Through appropriate screening, such as sorting by popularity and view count, obtain the 10 most popular videos. Conduct automatic speech recognition (ASR) analysis on these 10 popular videos, extract the oral scripts in each popular video, and analyze the extracted oral scripts to find relevant viewpoints, comments, and keywords. This information can help understand the core content of the videos and even provide inspiration for further content creation. For example, how they set the attraction, how they conduct question guidance, and how they incorporate product characteristics.

[0057] Then, in frame S106, based on the oral scripts, hot topics, and product materials, use a large language model to generate creative copy.

[0058] In one embodiment of the present disclosure, the product name, product information, titles of popular topics, and the ASR-extracted oral scripts of the top-ranked hot videos under a hot topic are used as parameters for the LLM model. Markdown language is used to guide the LLM to understand the product efficacy, selling points, target audience, etc., and then combined with the hot topic to be exploited and the oral content of other videos under that hot topic to create oral scripts.

[0059] Call the knowledge base and generate creative scripts based on the formulated script output rules, which include script word count, format, emotion, and semantic rules. In the prompt words for creating hot video scripts using the LLM, script output rules can be formulated. For example, 1. The output content does not need to show any process of thinking and directly outputs the oral script. 2. Each sentence should not exceed 15 characters, and the word count should not be too few. There should be no double quotes. 3. It is required that there is a line break between each sentence, but no empty lines. 4. Do not start with a serial number unless the discourse contains numbers. Directly output each sentence of the script. 5. The overall script should not exceed 10 sentences. 6. Add appropriate emoji at the end of appropriate sentences. Emoji cannot appear at the beginning of each sentence. 7. It is required to have an online sense and be able to catch the audience's attention. 8. Ensure that the meaning of the script is coherent before breaking the sentence. 9. Make the connection between the front and back scripts smoother, echoing the topic and selling points. The connecting statements should not be too similar and can be created independently.

[0060] Suppose the product is an intelligent environmental protection home system, and the selling points include smart home, environmental protection technology, health management, energy-saving effect, etc., while the hot topics revolve around green environmental protection, the development of smart home, and the trend of healthy life, etc. The oral scripts extracted from the popular videos that have gone viral under the current hot topics highlight the core features such as intelligence, environmental protection, and health through the dual attraction of emotion and reason, enhancing the competitiveness of the product in the market.

[0061] The generated creative script is as follows:

[0062] (At the beginning, grab attention) "The change in lifestyle starts from home." Welcome to the home experience of the new era! Are you also looking for a more healthy, intelligent, and environmentally friendly living space? Today, we will take you into the future home and redefine the meaning of home.

[0063] (Introduce the product selling points and emphasize smart home) "Smart home, one-key control." Imagine coming home and having the lights, temperature, air quality, and even the curtains automatically adjusted to perfectly meet your needs. This dream has now become a reality! The smart home system we launched uses advanced AI technology and the Internet of Things to create a highly intelligent living environment. From smart lights to smart speakers, every item is carefully designed to make your home more intelligent and convenient.

[0064] "Green environmental protection, adding points to the earth." Moreover, our system also has the gene of green environmental protection. Every piece of equipment saves energy to the maximum extent and reduces the carbon footprint, allowing you to contribute to the earth while enjoying intelligent technology! The brand-new energy-saving design helps you reduce your electricity bill and also enables you to contribute to environmental protection.

[0065] "Intelligent technology, caring for your health." At the same time, smart home not only focuses on the environment but also on your health. Our home system is built with air quality monitoring and intelligent temperature control functions, which can monitor the air quality in real time, help you maintain a comfortable living environment, and protect you from harmful substances. Every moment, the system is guarding your health.

[0066] "Leading the trend of healthy living." In 2025, healthy living has become a global trend, and we are standing at the forefront of this trend. Our smart home system perfectly meets the global demand for healthy, environmentally friendly, and intelligent living, making your home not just a place to rest but a more intelligent and healthy living partner.

[0067] "Smart and environmentally friendly home, the future is here!" Join us and experience the perfect integration of intelligence and environmental protection together, making your home more intelligent, environmentally friendly, and healthy! For more information, click on the link below to take the first step into the future life!

[0068] In the above voiceover copy, through an attractive voiceover statement, "The change in lifestyle starts from home!", it immediately catches the audience's attention and arouses curiosity. Combining the selling points of smart home, it emphasizes how the intelligent control system improves the convenience of users' lives. At the same time, it integrates AI and Internet of Things technologies, echoing the trend of modern smart home. By highlighting the green environmental protection and energy-saving effects, it meets the high demand for environmental protection of modern consumers and increases the market attractiveness of the product. Through the introduction of health-related functions such as air quality and intelligent temperature control, it highlights that the product not only focuses on intelligent life but also pays attention to the health of the residents, enhancing the value sense of the product. Keeping up with the trend of healthy living in 2025 and connecting with current hot topics further enhances the market relevance and attractiveness of the product. Finally, with a strong call to action, it encourages the audience to click on the link to learn more about the product and promote conversion.

[0069] This copy can effectively combine voiceover, hot topics, and product selling points. Through the dual attraction of emotion and reason, it highlights the core features such as intelligence, environmental protection, and health, enhancing the competitiveness of the product in the market.

[0070] Finally, in block S108, the embedding vector corresponding to the created copywriting is matched with the semantic vectors of the video slices in the database, sorted according to the similarity, and a marketing short video meeting the duration requirements is generated.

[0071] According to an embodiment of the present disclosure, the created copywriting is transformed into an embedding vector based on text vector embedding technology (such as BERT, Sentence-BERT, OpenAI Embedding API, etc.). For example, for the copywriting: "The smart home system brings you a convenient living experience.", the corresponding Embedding vector values are: [0.235, -0.542, 0.123,...].

[0072] The videos in the database are divided into multiple video slices, and each video slice contains visual information and audio information. A visual model is used to extract the visual and audio features in the video slice, and the visual and audio features are transformed into the embedding vectors of the video slice. Each slice can range from a few seconds to dozens of seconds, depending on the content and target duration of the video. Each slice contains certain visual information (such as pictures, actions, dialogues, etc.) and audio information. The visual and audio features of each video slice will be transformed into an embedding vector, and this vector represents the semantic content of the segment. For example, for the voice content of the video slice: "Now show you how to operate the smart home.", the corresponding Embedding vector values are: [0.245, -0.499, 0.119,...].

[0073] The embedding vector corresponding to the created copywriting is matched with the embedding vectors of the video slices, sorted from high to low according to the similarity, and the most matching multiple video slices are selected. Methods such as cosine similarity and Euclidean distance can be used to measure the semantic similarity between the copywriting and the video slices. For example, the similarity between the copywriting vector and the first video slice is: 0.92, the similarity between the copywriting vector and the second video slice is: 0.85, and the similarity between the copywriting vector and the third video slice is: 0.80.

[0074] The created copywriting is converted into speech through a text-to-speech engine, and the speech and multiple video slices are mixed and synthesized into the final video. Among them, the text-to-speech engine (TTS) can recognize information such as words, sentence structures, and punctuation marks in the copywriting, convert the vocabulary into a standard pronunciation form, judge the grammatical structure of the sentence, and determine the stress, pause, etc. Adjust the speech duration according to the duration of the video slice, and synthesize the selected video slices in chronological order according to the sorting result of the similarity. According to the generated speech duration and the content of the video slice, mix and synthesize a video that matches the created copywriting.

[0075] For example, the generated voice duration is 5 seconds. Select the three video clips with the highest similarity, with durations of 2 seconds, 2 seconds, and 1 second respectively, and the total duration is 5 seconds. Mix and synthesize the voice with these three video clips to complete the final video. This video can be formatted and further optimized according to the requirements of the target platform (such as Douyin, YouTube, etc.). For example, adjust the video picture quality, audio volume balance, add subtitles, etc.

[0076] Figure 2 FIG. is a schematic diagram of a marketing short video generation process according to an embodiment of the present disclosure. Refer to Figure 2 As shown, first, obtain the hot topic title and product information, perform a matching degree analysis on the hot topic title and product information based on the large language model, and output the hot topic that is most matched with the product in descending order of the matching degree. Search for popular videos related to the most matched hot topic in the video platform, and extract the oral broadcast copywriting in the popular videos based on the automatic speech recognition technology. Subsequently, based on the oral broadcast copywriting, hot topic, and product information, use the large language model to generate creative copywriting. Finally, perform a similarity matching between the embedding vector corresponding to the creative copywriting and the semantic vector of the video clips in the database, and sort according to the similarity and generate a marketing short video that meets the duration requirements.

[0077] Figure 3 FIG. is a schematic block diagram of a marketing short video generation device 300 based on a large language model according to an embodiment of the present disclosure. As Figure 3 shown, the device 300 may include a processor 310 and a memory 320 storing a computer program. When the computer program is executed by the processor 310, the device 300 can execute the steps of the marketing short video generation method 100 based on the large language model as Figure 1 shown. In one example, the device 300 may be a computer device or a cloud computing node. The device 300 can perform a matching degree analysis on the hot topic title and product information based on the large language model, and output the hot topic that is most matched with the product in descending order of the matching degree; search for popular videos related to the most matched hot topic in the video platform, and extract the oral broadcast copywriting in the popular videos based on the automatic speech recognition technology; generate creative copywriting based on the oral broadcast copywriting, hot topic, and product information; and perform a similarity matching between the embedding vector corresponding to the creative copywriting and the semantic vector of the video clips in the database, and sort according to the similarity and generate a mixed video that meets the duration requirements.

[0078] In an embodiment of the present disclosure, the processor 310 may be, for example, a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a processor based on a multi-core processor architecture, etc. The memory 320 may be any type of memory implemented using data storage technology, including but not limited to random access memory, read-only memory, semiconductor-based memory, flash memory, disk memory, etc.

[0079] In addition, in an embodiment of the present disclosure, the device 300 may also include an input device 330, such as a keyboard, a mouse, etc., for inputting hot topic titles and brand or product information. Additionally, the device 300 may further include an output device 340, such as a display, etc., for outputting marketing short videos.

[0080] In other embodiments of the present disclosure, there is also provided a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, is capable of implementing the steps of the method 100 for generating a marketing short video based on a large language model as Figure 1 shown.

[0081] In summary, according to the method and device for generating a marketing short video based on a large language model in the embodiments of the present disclosure, by using the large language model to perform matching analysis on product information and hot topic titles, the hot topics most relevant to the product can be automatically extracted from a vast amount of content, which can not only improve the efficiency of video creation, but also ensure that the video content keeps up with the trend, has a high market relevance and user attention. By using automatic speech recognition technology to extract the oral broadcast copy from popular videos, the process of manual recording and script writing is omitted. By analyzing and learning from the content of existing popular video copies, it is ensured that the newly generated videos are more in line with the acceptance habits of users and the platform trend. By using the large language model to combine the oral broadcast copy, hot topics and product information to generate a creative copy, and matching the creative copy with relevant video clips, automated and accurate video clip screening can be achieved, and a mixed-editing video that meets the duration requirements and has consistent emotional expression can be generated, greatly improving the efficiency and quality of generating marketing short videos.

[0082] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatuses and methods according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagram may represent a module, a segment of a program, or a portion of an instruction, which contains one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions noted in the blocks may occur in a different order than noted in the accompanying drawings. For example, two consecutive blocks may actually be executed substantially in parallel, or they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system that performs the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0083] Unless the context clearly dictates otherwise herein, the singular forms of words used in this specification and the appended claims also include the plural, and vice versa. Thus, when reference is made to the singular, the corresponding plural is generally included. Similarly, the terms "comprising" and "including" are to be construed as inclusive rather than exclusive. Likewise, the term "or" should be interpreted as inclusive, unless expressly prohibited by the context herein. Where the term "exemplary" is used herein, particularly when it is followed by a list of terms, the "exemplary" is merely illustrative and explanatory and should not be considered exclusive or extensive.

[0084] Further aspects and scope of adaptability become apparent from the description provided herein. It should be understood that the various aspects of the present application may be implemented alone or in combination with one or more other aspects. It should also be understood that the description herein and the specific embodiments are for illustrative purposes only and are not intended to limit the scope of the present application.

[0085] The above has described in detail several embodiments of the present disclosure. However, it is obvious that those skilled in the art can make various modifications and variations to the embodiments of the present disclosure without departing from the spirit and scope of the present disclosure. The protection scope of the present disclosure is defined by the appended claims.

Claims

1. A marketing short video generation method based on a large language model, characterized in that: include: Based on the large language model, the matching degree of hot topics and product information is analyzed, and the hot topics that best match the product are output in descending order of matching degree; Searching for popular videos related to the most matching hot topic on the video platform, and extracting the spoken text in the popular videos based on automatic speech recognition technology; Based on the oral copy, hot topics and product information, a large language model is used to generate creative copy; as well as The embedding vector corresponding to the creative text is matched with the semantic vector of the video slice in the database for similarity, sorted according to the similarity, and a marketing short video that meets the time requirement is generated.

2. The marketing short video generation method based on a large language model according to claim 1 is characterized in that: The matching degree analysis of hot topics titles and product information based on the large language model is performed, and the hot topics that best match the product are outputted in descending order of matching degree, including: Get current hot headlines and clicks from multiple news or media platforms; Inferring and expanding hot content related to the hot title according to the hot title; Use the knowledge base to gain an in-depth understanding of brand or product information, determine product advantages, selling points, target audiences, and ways to integrate with hot content; Based on the hot content and the product advantages, selling points and target audience and the way of combining with the hot content, the hot title and the product are analyzed for correlation and matching degree; and According to the sorting results from high to low matching degree, the most relevant hot titles and corresponding annotation information for the product are output.

3. The marketing short video generation method based on a large language model according to claim 2 is characterized in that: The inferring and expanding the hot content related to the hot title according to the hot title includes: Perform word segmentation on the hot topic title, and extract keywords in the hot topic title from the word segmentation result; Using a word vector model to calculate words that are semantically similar to the keyword, and expanding hot content based on the words that are semantically similar and context information of the keyword; and Use pre-built knowledge bases or databases in related fields to query the background of hot events and infer potential hot content through inference models.

4. The marketing short video generation method based on a large language model according to claim 2 is characterized in that: The analysis of the correlation matching degree between the hot title and the product based on the hot content and the product advantages, selling points and target audience and the combination with the hot content includes: Compare the audience of the hot content with the target audience of the brand’s products; Determine whether the emotional atmosphere of the hot event is consistent with the brand's core values ​​and product selling points; and Conduct content integration analysis on hot events and product selling points to determine whether the product selling points can form a natural connection with the content of hot events.

5. The marketing short video generation method based on a large language model according to claim 1 is characterized in that: The method of using a large language model to generate creative copy based on the oral copy, hot topics and product information includes: The product or brand name, product information, hot topics and oral copy extracted from popular videos are used as parameters of the LLM model; The knowledge base is called and creative copy is generated based on the established copy output rules, wherein the copy output rules include the number of words, format, emotion and semantic rules of the copy.

6. The marketing short video generation method based on a large language model according to claim 1 is characterized in that: The similarity matching of the embedding vector corresponding to the creative copy with the semantic vector of the video slice in the database, sorting according to the similarity and generating a marketing short video that meets the time requirement includes: Based on the text vector embedding technology, the creative copy and video slices are converted into embedding vectors respectively; Performing similarity matching between the embedding vector corresponding to the creative text and the embedding vector of the video slice, sorting the similarities from high to low, and selecting the most matching multiple video slices; and The creative copy is converted into speech through a text-to-speech engine, and the speech and multiple video slices are mixed and cut into a final video.

7. The marketing short video generation method based on a large language model according to claim 6 is characterized in that: The method of converting the creative copy and the video slice into embedding vectors based on the text vector embedding technology includes: Use the pre-trained text embedding model to vectorize the text and obtain the embedding vector corresponding to the creative text; Divide the videos in the database into multiple video slices, each video slice containing visual information and audio information; and The visual model is used to extract visual and audio features in the video slice, and the visual and audio features are converted into embedding vectors of the video slice.

8. The marketing short video generation method based on a large language model according to claim 6, characterized in that: The step of converting the creative text into speech through a text-to-speech engine, and mixing the speech and multiple video slices into a final video comprises: The speech duration is adjusted according to the duration of the video slices, and the selected video slices are synthesized in time order according to the results of similarity sorting; Based on the generated voice duration and the content of the video slices, a mixed cut is performed to synthesize a video that matches the creative copy.

9. A marketing short video generation device based on a large language model, characterized in that: The device comprises: at least one processor; and at least one memory storing a computer program; Wherein, when the computer program is executed by the at least one processor, the device performs the steps of the marketing short video generation method based on a large language model according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the computer program implements the steps of the method for generating a marketing short video based on a large language model according to any one of claims 1 to 8.

Images