Short video copywriting generation method and device of large language model based on network hotspots
By introducing online hot databases, popular video topics and golden three-second structure in the short video copy generation method, combined with a large language model, the problem of insufficient knowledge richness and trendiness in the existing technology is solved. The generated short video copy has Higher trends and uniqueness reduce the risk of plagiarism.
Patent Information
- Application Number
- CN202411070791.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-06
- Publication Date
- 2024-11-22
AI Technical Summary
The existing short video copy generation methods cannot guarantee the richness of knowledge and timely reflect online hotspots, resulting in the generated text lacking the trendy and hot topic style, the structure is boring, and it is easily judged as plagiarism.
Using a large language model method based on network hotspots, through the search of the video content strategy database and hotspot vector database, the golden three-second structure and hotspot matching the video type and topic selected by the user are obtained, and pre-processed with the big model to generate short videos. Copywriting.
It realizes the real-time tracking of popular online events in short video copy, which increases the trend and uniqueness of copywriting, reduces the risk of plagiarism, and increases the attractiveness and probability of generating copywriting.
Smart Images

Figure CN119003757A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of short video copywriting, and in particular to a method and device for generating short video copywriting based on a large language model of network hot spots. Background Art
[0002] The current method of generating short video copywriting is to generate it through a large language model (LLM); first, input the relevant requirements, and then retrieve and match the previously imported copywriting knowledge from the vector database. The vector database stores candidate copywriting knowledge vector features, which are used to meet the needs of copywriting generation. Next, obtain the corresponding knowledge vector features according to the requirements, and generate prompt text containing contextual information to guide the large language model to generate copywriting. Finally, this prompt text is input into the model together with the prompt words for generating copywriting to generate the target copywriting that meets the requirements.
[0003] Currently, the existing technology uses vector databases to obtain contextual information that matches the needs and then inputs it into the large model. However, the knowledge expression in the existing vector database is provided by users, which cannot guarantee the richness of knowledge and cannot reflect the latest network hot spots in a timely manner, resulting in the final generated text having no information related to network hot spots, which does not meet the theme style of short video copywriting that needs to keep up with trendy hot spots; as for the text structure, it is too boring and does not meet the rule of the golden three seconds at the beginning, and cannot arouse the audience's interest in continuing to watch; as for the text content, the current technology uses candidate copywriting as a reference, which will result in the output copywriting being very similar to the candidate copywriting, unable to show uniqueness, and easily judged as plagiarism.
[0004] In view of this, this application is filed. Summary of the invention
[0005] The present invention provides a method and device for generating short video text based on a large language model of network hot spots, which can at least partially improve the above-mentioned problems.
[0006] To achieve the above object, the present invention adopts the following technical solutions:
[0007] A method for generating short video copy based on a large language model of network hot spots, comprising:
[0008] Obtain the video type selected by the current user, and search the video content strategy database according to the video type, select a candidate topic and golden three-second structure corresponding to the video type, or obtain the topic filled in by the user;
[0009] Searching a hotspot vector database according to the candidate topic or the topic filled in by the user to find k hotspots that best match the candidate topic or the topic filled in by the user;
[0010] According to the user's selection signal, a hot spot is determined from k hot spots for creation to obtain a short video copy. The hot spot is analyzed to obtain the creation direction, and a large model is used to pre-process the hot spot, topic, creation direction, and golden three-second structure to obtain the short video copy.
[0011] Preferably, the video content strategy database stores the golden three-second structures and candidate topics of various short video types, wherein multiple current popular short videos are obtained through a popular short video acquisition interface, and classified short video translation results are obtained through a translation interface. A large model is used to analyze the topics of multiple popular short videos in the classified short video translation results, and then m3e is used for vectorization and clustering to obtain n candidate topics.
[0012] Preferably, multiple popular short videos in the classified short video translation results are classified, the first sentence of each classified short video is extracted, the first sentences of all short videos are vectorized using m3e, all the first sentences are divided into m categories according to the hierarchical clustering method, and the structure of each first sentence is disassembled using a large model to obtain m golden three-second structures.
[0013] Preferably, the hotspot vector database stores multiple hotspots, hotspot description information and hotspot types, wherein multiple current hotspots and hotspot related information are obtained through a hotspot related information acquisition interface, and hotspot classification is obtained through a hotspot classification interface.
[0014] Preferably, according to the user's selection signal, a hotspot is determined from k hotspots for creation to obtain a short video copy, specifically:
[0015] The large model is called to analyze and process the determined hot spots and the hot background of the hot spots to obtain multiple creative directions;
[0016] Randomly select one from multiple creative directions, and randomly select a golden three-second structure from the video content strategy database;
[0017] The candidate topics or topics filled in by users, the selected golden three-second structure, the determined hot spots and the creative direction are input into the big model to generate short video copy.
[0018] The present invention also provides a short video copy generation device based on a large language model of network hot spots, which includes:
[0019] A topic search unit, used to obtain the video type selected by the current user, and search the video content strategy database according to the video type, select a candidate topic and golden three-second structure corresponding to the video type, or obtain the topic filled in by the user;
[0020] A hotspot search unit, configured to search a hotspot vector database according to the candidate topic or the topic filled in by the user, and find k hotspots that best match the candidate topic or the topic filled in by the user;
[0021] The short video copy generation unit is used to determine a hot spot from k hot spots for creation according to the user's selection signal to obtain a short video copy. The hot spot is analyzed to obtain the creation direction, and a large model is used to pre-process the hot spot, topic, creation direction, and golden three-second structure to obtain the short video copy.
[0022] Preferably, the video content strategy database stores the golden three-second structures and candidate topics of various short video types, wherein multiple current popular short videos are obtained through a popular short video acquisition interface, and classified short video translation results are obtained through a translation interface. A large model is used to analyze the topics of multiple popular short videos in the classified short video translation results, and then m3e is used for vectorization and clustering to obtain n candidate topics.
[0023] Preferably, multiple popular short videos in the classified short video translation results are classified, the first sentence of each classified short video is extracted, the first sentences of all short videos are vectorized using m3e, all the first sentences are divided into m categories according to the hierarchical clustering method, and the structure of each first sentence is disassembled using a large model to obtain m golden three-second structures.
[0024] Preferably, the hotspot vector database stores multiple hotspots, hotspot description information and hotspot types, wherein multiple current hotspots and hotspot related information are obtained through a hotspot related information acquisition interface, and hotspot classification is obtained through a hotspot classification interface.
[0025] Preferably, according to the user's selection signal, a hotspot is determined from k hotspots for creation to obtain a short video copy, specifically:
[0026] The large model is called to analyze and process the determined hot spots and the hot background of the hot spots to obtain multiple creative directions;
[0027] Randomly select one from multiple creative directions, and randomly select a golden three-second structure from the video content strategy database;
[0028] The candidate topics or topics filled in by users, the selected golden three-second structure, the determined hot spots and the creative direction are input into the big model to generate short video copy.
[0029] In summary, the short video copywriting generation method based on the large language model of network hot spots first introduces the network hot spot database, and continuously updates and maintains the network hot spot database, so that the generated short video copy can keep up with the popular network events or buzzwords in real time. Secondly, the short video copywriting generation method based on the large language model of network hot spots also introduces the topics and golden three-second structure of popular videos. By classifying the types of popular short videos, analyzing the content and disassembling the structure, the materials that can be used for reference are continuously accumulated in the database. When the user selects the corresponding type, he can get the corresponding type of material. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 It is a flow chart of the method for generating short video text based on the large language model of network hot spots provided by the first aspect of the present invention;
[0031] Figure 2 It is a flow chart of a method for generating short video texts based on a large language model of network hot spots provided by the second aspect of the present invention;
[0032] Figure 3 It is a module diagram of a short video copy generation device based on a large language model of network hot spots provided by the second embodiment of the present invention. DETAILED DESCRIPTION
[0033] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0034] refer to Figure 1 , Figure 2 As shown, the first embodiment of the present invention discloses a method for generating short video text based on a large language model of network hot spots, which can be executed by a short video text generating device based on a large language model of network hot spots (hereinafter referred to as a generating device), and in particular, executed by one or more processors in the generating device to implement the following method:
[0035] S101, obtaining the video type selected by the current user, searching the video content strategy database according to the video type, selecting a candidate topic and golden three-second structure corresponding to the video type, or obtaining the topic filled in by the user;
[0036] Preferably, the video content strategy database stores the golden three-second structures and candidate topics of various short video types, wherein multiple current popular short videos are obtained through a popular short video acquisition interface, and classified short video translation results are obtained through a translation interface. A large model is used to analyze the topics of multiple popular short videos in the classified short video translation results, and then m3e is used for vectorization and clustering to obtain n candidate topics.
[0037] Preferably, multiple popular short videos in the classified short video translation results are classified, the first sentence of each classified short video is extracted, the first sentences of all short videos are vectorized using m3e, all the first sentences are divided into m categories according to the hierarchical clustering method, and the structure of each first sentence is disassembled using a large model to obtain m golden three-second structures.
[0038] In this embodiment, the video content strategy database mainly contains two contents: golden three-second structure and candidate topics. Golden three-second is based on the popular short videos that have been crawled. After the short videos are classified, the first sentence of each category of short videos is extracted, and all the first sentences are vectorized using m3e (Moka Massive Mixed Embedding). Then, the hierarchical clustering method is used to divide all the first sentences into m categories. Finally, the first sentence structure of each category is disassembled using a large model to obtain m golden three-second structures. Candidate topics are the topics of each video analyzed using a large model. They are also vectorized and clustered using m3e to obtain n candidate topics.
[0039] Among them, decomposing the first sentence structure of each category and obtaining the golden three-second structure can be understood as: inputting the first sentence and the prompt word into the large model to obtain the golden three-second structure. The prompt word is probably "help me parse the structure of this sentence."
[0040] S102, searching a hotspot vector database according to the candidate topic or the topic filled in by the user, and finding k hotspots that best match the candidate topic or the topic filled in by the user;
[0041] Preferably, the hotspot vector database stores multiple hotspots, hotspot description information and hotspot types, wherein multiple current hotspots and hotspot related information are obtained through a hotspot related information acquisition interface, and hotspot classification is obtained through a hotspot classification interface.
[0042] In this embodiment, the hotspot vector database stores network hotspots that are updated by regular crawlers, and uses a large model to search the Internet to obtain background introductions related to the hotspots.
[0043] S103, according to the user's selection signal, determine a hot spot from k hot spots for creation, and obtain a short video copy, wherein the hot spot is analyzed to obtain the creation direction, and a large model is used to pre-process the hot spot, topic, creation direction, and golden three-second structure to obtain the short video copy.
[0044] Specifically, step S103 includes: calling a large model to analyze and process the determined hotspot and the hotspot background of the hotspot to obtain multiple creation directions;
[0045] Randomly select one from multiple creative directions, and randomly select a golden three-second structure from the video content strategy database;
[0046] The candidate topics or topics filled in by users, the selected golden three-second structure, the determined hot spots and the creative direction are input into the big model to generate short video copy.
[0047] In this embodiment, the type of video selected by the user to generate is first obtained, and the corresponding video content strategy database is searched to provide candidate topics under this type. The user can fill in the topic by himself or select a candidate topic, and find the k hot spots that best match the topic by searching the hot spot vector database. The user can arbitrarily select a hot spot for creation. When creating, the big model is called to analyze the hot spot and the hot spot background to obtain several creation directions, and a creation direction is randomly selected. The topic, a randomly selected golden three-second structure, the hot spot, and the creation direction are input into the big model to generate a short video copy. If the user is not satisfied, he can continue to click to generate and iterate continuously until a satisfactory result is generated.
[0048] For example, if the user selects the type of "female emotions", the corresponding topics are all related to "female emotions", such as "Characteristics of Bad Men" and "Middle-aged Love Choices", etc.; hot topics must also be relevant to female emotions, such as "What does it mean when a girl says yes" and "The essence of love is to be seen", etc.; the creative direction is generated based on the hot topics. For example, for the hot topic "The essence of love is to be seen", there can be the following creative directions:
[0049] (1) "Love's Blind Spot": Comparing relationships that focus only on superficial companionship and ignore inner needs, this film explores how invisible love can lead to misunderstanding and alienation, and guides the audience to think about what true love is.
[0050] (2) "Negative Example": Through a skit, it demonstrates the consequences of ignoring the other person's inner feelings and emphasizes the importance of "being seen".
[0051] (3) "Seeing in Art": By analyzing the relationships between characters in films and literary works, the theme of "being seen" is expounded, allowing the audience to understand this concept from an artistic perspective.
[0052] The large model used may be a general large model. When used, it is emphasized in the prompt words that the opening structure must be a golden three-second structure.
[0053] In summary, the short video copywriting generation method based on the large language model of network hot spots first introduces the network hot spot database. By adding network hot spots and background information that have a high degree of match with user topics, the trendiness of short videos is increased, and hot spots can be "taken advantage of" to attract users to watch. Secondly, the topics of popular videos and the golden three-second structure are introduced. Users not only do not need to think about the topic of the video themselves, but can also refer to the golden three seconds of popular videos. This not only increases the online sense and the probability of a hit of the short video, but also will not be judged as plagiarism due to comprehensive imitation.
[0054] Among them, using the same golden three seconds will not be judged as plagiarism. This is because it is not the original golden three seconds that are used, but the golden three seconds structure. For example: [Have you found that some xx-year-old women / men will aa when zz...], the above sentence is the extracted structure. Different topics and hot spots will be filled in with different xx zz and aa, so the final results will be different.
[0055] See also Figure 3 The second embodiment of the present invention further provides a short video copy generation device based on a large language model of network hot spots, which includes:
[0056] The topic search unit 201 is used to obtain the video type selected by the current user, and search the video content strategy database according to the video type, select a candidate topic and golden three-second structure corresponding to the video type, or obtain the topic filled in by the user;
[0057] A hotspot search unit 202, configured to search a hotspot vector database according to the candidate topic or the topic filled in by the user, and find k hotspots that best match the candidate topic or the topic filled in by the user;
[0058] The short video copy generation unit 203 is used to determine a hot spot from k hot spots for creation according to the user's selection signal to obtain a short video copy, wherein the hot spot is analyzed to obtain the creation direction, and a large model is used to pre-process the hot spot, topic, creation direction, and golden three-second structure to obtain the short video copy.
[0059] Preferably, the video content strategy database stores the golden three-second structures and candidate topics of various short video types, wherein multiple current popular short videos are obtained through a popular short video acquisition interface, and classified short video translation results are obtained through a translation interface. A large model is used to analyze the topics of multiple popular short videos in the classified short video translation results, and then m3e is used for vectorization and clustering to obtain n candidate topics.
[0060] Preferably, multiple popular short videos in the classified short video translation results are classified, the first sentence of each classified short video is extracted, the first sentences of all short videos are vectorized using m3e, all the first sentences are divided into m categories according to the hierarchical clustering method, and the structure of each first sentence is disassembled using a large model to obtain m golden three-second structures.
[0061] Preferably, the hotspot vector database stores multiple hotspots, hotspot description information and hotspot types, wherein multiple current hotspots and hotspot related information are obtained through a hotspot related information acquisition interface, and hotspot classification is obtained through a hotspot classification interface.
[0062] Preferably, according to the user's selection signal, a hotspot is determined from k hotspots for creation to obtain a short video copy, specifically:
[0063] The large model is called to analyze and process the determined hot spots and the hot background of the hot spots to obtain multiple creative directions;
[0064] Randomly select one from multiple creative directions, and randomly select a golden three-second structure from the video content strategy database;
[0065] The candidate topics or topics filled in by users, the selected golden three-second structure, the determined hot spots and the creative direction are input into the big model to generate short video copy.
[0066] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. A short video copywriting generation method based on a large language model of network hot spots, It is characterized in that include: Obtain the video type selected by the current user, and search the video content strategy database according to the video type, select a candidate topic and golden three-second structure corresponding to the video type, or obtain the topic filled in by the user; Searching a hotspot vector database according to the candidate topic or the topic filled in by the user to find k hotspots that best match the candidate topic or the topic filled in by the user; According to the user's selection signal, a hot spot is determined from k hot spots for creation to obtain a short video copy. The hot spot is analyzed to obtain the creation direction, and a large model is used to pre-process the hot spot, topic, creation direction, and golden three-second structure to obtain the short video copy.
2. According to claim 1, the short video copywriting generation method based on the large language model of network hot spots, It is characterized in that The video content strategy database stores the golden three-second structures and candidate topics of various short video types, wherein multiple current popular short videos are obtained through a popular short video acquisition interface, and classified short video translation results are obtained through a translation interface. A large model is used to analyze the topics of multiple popular short videos in the classified short video translation results, and then m3e is used for vectorization and clustering to obtain n candidate topics.
3. The method for generating short video texts based on a large language model of network hot spots according to claim 2, It is characterized in that Classify multiple popular short videos in the classified short video translation results, extract the first sentence of each classified short video, use m3e to vectorize the first sentences of all short videos, divide all the first sentences into m categories according to the hierarchical clustering method, and use a large model to disassemble the first sentence structure of each category to obtain m golden three-second structures.
4. The method for generating short video texts based on a large language model of network hot spots according to claim 1, It is characterized in that The hotspot vector database stores multiple hotspots, hotspot description information and hotspot types, wherein multiple current hotspots and hotspot related information are obtained through a hotspot related information acquisition interface, and hotspot classification is obtained through a hotspot classification interface.
5. The method for generating short video texts based on a large language model of network hot spots according to claim 1, It is characterized in that According to the user's selection signal, a hotspot is determined from k hotspots for creation, and a short video copy is obtained, which is specifically: The large model is called to analyze and process the determined hot spots and the hot background of the hot spots to obtain multiple creative directions; Randomly select one from multiple creative directions, and randomly select a golden three-second structure from the video content strategy database; The candidate topics or topics filled in by users, the selected golden three-second structure, the determined hot spots and the creative direction are input into the big model to generate short video copy.
6. A short video copywriting generation device based on a large language model of network hot spots, It is characterized in that include: A topic search unit, used to obtain the video type selected by the current user, and search the video content strategy database according to the video type, select a candidate topic and golden three-second structure corresponding to the video type, or obtain the topic filled in by the user; A hotspot search unit, configured to search a hotspot vector database according to the candidate topic or the topic filled in by the user, and find k hotspots that best match the candidate topic or the topic filled in by the user; The short video copy generation unit is used to determine a hot spot from k hot spots for creation according to the user's selection signal to obtain a short video copy. The hot spot is analyzed to obtain the creation direction, and a large model is used to pre-process the hot spot, topic, creation direction, and golden three-second structure to obtain the short video copy.
7. The short video copy generation device based on the large language model of network hot spots according to claim 6, It is characterized in that The video content strategy database stores the golden three-second structures and candidate topics of various short video types, wherein multiple current popular short videos are obtained through a popular short video acquisition interface, and classified short video translation results are obtained through a translation interface. A large model is used to analyze the topics of multiple popular short videos in the classified short video translation results, and then m3e is used for vectorization and clustering to obtain n candidate topics.
8. The short video copy generation device based on the large language model of network hot spots according to claim 7, It is characterized in that Classify multiple popular short videos in the classified short video translation results, extract the first sentence of each classified short video, use m3e to vectorize the first sentences of all short videos, divide all the first sentences into m categories according to the hierarchical clustering method, and use a large model to disassemble the first sentence structure of each category to obtain m golden three-second structures.
9. The short video copy generation device based on the large language model of network hot spots according to claim 6, It is characterized in that The hotspot vector database stores multiple hotspots, hotspot description information and hotspot types, wherein multiple current hotspots and hotspot related information are obtained through a hotspot related information acquisition interface, and hotspot classification is obtained through a hotspot classification interface.
10. The short video copy generation device based on the large language model of network hot spots according to claim 6, It is characterized in that According to the user's selection signal, a hotspot is determined from k hotspots for creation, and a short video copy is obtained, which is specifically: The large model is called to analyze and process the determined hot spots and the hot background of the hot spots to obtain multiple creative directions; Randomly select one from multiple creative directions, and randomly select a golden three-second structure from the video content strategy database; The candidate topics or topics filled in by users, the selected golden three-second structure, the determined hot spots and the creative direction are input into the big model to generate short video copy.
Citation Information
Patent Citations
Multimedia content analysis method and device, equipment and storage medium
CN116702749A
Big language model-based copywriting generation method, apparatus and device, and storage medium
CN117611254A
Video generation method and device, electronic equipment and storage medium
CN117651156A
Video copywriting generation method and device based on large language model, equipment and medium
CN118095228A
