Short video optimization method and device based on knowledge graph, computer equipment and readable storage medium

Abstract

The invention discloses a short video optimization method and device based on a knowledge graph, computer equipment and a readable storage medium, and the method comprises the steps: firstly obtaining a target short video link and basic information, and calling an internet red big data service for implementation; then, based on the basic information and a web red knowledge graph, depth information is extracted through a web red big language model; and finally inputting the basic and depth information into the model to obtain an optimization suggestion. According to the design, multiple technical resources are utilized for cooperation, and the short video value is effectively mined to improve the propagation effect.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a short video optimization method, device, computer equipment and readable storage medium based on a knowledge graph. Background Art

[0002] With the rapid development of the short video industry, the number of short videos has exploded. How to optimize short video content to improve its dissemination effect has become an important issue. Existing optimization methods often lack systematicity and comprehensiveness, making it difficult to fully tap the potential of short videos. Summary of the invention

[0003] The purpose of the present invention is to provide a short video optimization method, device, computer equipment and readable storage medium based on knowledge graph.

[0004] In a first aspect, an embodiment of the present invention provides a short video optimization method based on a knowledge graph, comprising: Obtain a target short video link, where the target short video link points to the target short video; Calling a preset Internet celebrity big data service to obtain basic information of the target short video; Based on the basic information and the preset Internet celebrity knowledge graph, calling the preset Internet celebrity large language model to extract information from the target short video to obtain depth information; The basic information and the depth information are input into the internet celebrity language model to obtain optimization suggestions for the target short video.

[0005] In a possible implementation, calling a preset Internet celebrity big data service to obtain basic information of the target short video includes: Calling a preset Internet celebrity big data service to obtain the effect data, basic content and background information of the target short video; the effect data includes the release time, playback volume, number of likes, number of comments and number of favorites of the target short video; the basic content includes the title, tags and audio of the target short video; the background information includes the creator of the target short video, the top competing videos and the creator of the top competing videos; Based on the effect data, the basic content and the background information as the basic information of the target short video.

[0006] In a possible implementation, based on the basic information, calling a preset internet celebrity language model to extract information from the target short video to obtain depth information includes: The target short video is transcribed by calling a preset tool set through the Internet celebrity language model, and the audio is transcribed into lines; Analyze the title, the tag, and the lines through the Internet celebrity language model to obtain the theme of the target short video; The deep information is obtained from the preset internet celebrity knowledge graph through the internet celebrity language model based on the effect data, the basic content, the background information, the lines and the theme.

[0007] In a possible implementation, the preset Internet celebrity knowledge graph is constructed in the following manner, including: Obtaining video information and account information of a plurality of stored videos stored by the preset internet celebrity big data service through the internet celebrity big language model; Structuring the video information and the account information of each stored video to obtain a text block corresponding to each stored video, wherein the text block includes the video information and the account information; Inputting the multiple text blocks into the Internet celebrity language model for multiple rounds of extraction to obtain entities, relations, and covariates corresponding to each of the multiple text blocks; Multi-layer cluster extraction is performed based on the entities, the relationships, and the covariates to obtain the preset Internet celebrity knowledge graph, which includes a four-layer graph structure constructed from top to bottom by major category clusters, subcategory clusters, attitude clusters, and theme clusters.

[0008] In a possible implementation, the acquiring the depth information from the preset internet celebrity knowledge graph based on the effect data, the basic content, the background information, the lines, and the theme through the internet celebrity language model includes: Based on the effect data, the basic content, the background information, the lines and the theme, the Internet celebrity language model performs large category cluster matching, subcategory cluster matching, attitude cluster matching and topic cluster matching in the preset Internet celebrity knowledge graph, and obtains the target entity and target text block according to the preset relevance threshold; The target entity and the target text are divided into blocks as the depth information.

[0009] In a possible implementation, the preset internet celebrity language model is obtained in the following manner, including: Acquire a basic large language model and an advanced large language model, wherein the model parameter amount of the advanced large language model is greater than that of the basic large language model; The advanced large language model performs analysis, self-verification, and self-selection extraction strategies on sample short videos based on a self-discovery mechanism, and generates knowledge extraction data through prompt engineering; Based on the knowledge extraction data, the basic large language model is fine-tuned in a LoRA manner to obtain the preset Internet celebrity large language model.

[0010] In a possible implementation, inputting the basic information and the depth information into the internet celebrity language model to obtain optimization suggestions for the target short video includes: The basic information and the depth information are input into the internet celebrity language model to obtain optimization suggestions including new titles, new tags, and new lines for the target short video.

[0011] In a second aspect, an embodiment of the present invention provides a short video optimization device based on a knowledge graph, comprising: An acquisition module is used to acquire a target short video link, wherein the target short video link points to a target short video; call a preset Internet celebrity big data service to acquire basic information of the target short video; based on the basic information and a preset Internet celebrity knowledge graph, call a preset Internet celebrity big language model to extract information from the target short video to obtain depth information; The optimization module is used to input the basic information and the depth information into the Internet celebrity language model to obtain optimization suggestions for the target short video.

[0012] In a third aspect, an embodiment of the present invention provides a computer device, comprising a processor and a non-volatile memory storing computer instructions, wherein when the computer instructions are executed by the processor, the computer device executes the method described in the first aspect.

[0013] In a fourth aspect, an embodiment of the present invention provides a readable storage medium, wherein the readable storage medium includes a computer program, and when the computer program is executed, the computer device where the readable storage medium is located is controlled to execute the method described in the first aspect.

[0014] Compared with the prior art, the beneficial effects provided by the present invention include: adopting a short video optimization method, device, computer equipment and readable storage medium based on knowledge graph disclosed by the present invention, by obtaining the target short video link and basic information, calling the Internet celebrity big data service to achieve it. Then, based on the basic information and the Internet celebrity knowledge graph, the deep information is extracted through the Internet celebrity language model. Finally, the basic and deep information are input into the model to obtain optimization suggestions. With this design, multiple technical resources are used to collaborate and effectively tap the value of short videos to improve the communication effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative work.

[0016] Figure 1 A schematic diagram of the steps of a short video optimization method based on a knowledge graph provided in an embodiment of the present invention; Figure 2 A schematic diagram of a short video optimization system framework based on a knowledge graph provided in an embodiment of the present invention; Figure 3 A schematic diagram of the Prompt framework flow provided by an embodiment of the present invention; Figure 4 A schematic diagram of multiple rounds of entity extraction provided by an embodiment of the present invention; Figure 5 A schematic diagram of multi-layer cluster extraction provided by an embodiment of the present invention; Figure 6 A schematic block diagram of the structure of a short video optimization device based on a knowledge graph provided in an embodiment of the present invention; Figure 7 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0018] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.

[0019] In order to solve the technical problems in the aforementioned background technology, Figure 1 A schematic flow chart of a short video optimization method based on a knowledge graph provided in an embodiment of the present disclosure is provided below, and the short video optimization method based on a knowledge graph is introduced in detail.

[0020] Step S201, obtaining a target short video link, where the target short video link points to a target short video; Step S202, calling a preset Internet celebrity big data service to obtain basic information of the target short video; Step S203, based on the basic information and the preset internet celebrity knowledge graph, calling the preset internet celebrity language model to extract information from the target short video to obtain depth information; Step S204: input the basic information and the depth information into the internet celebrity language model to obtain optimization suggestions for the target short video.

[0021] In the embodiment of the present invention, for example, it is assumed that the server receives a request from a client (such as a content optimization platform used by a short video operator), and the request contains a short video link on the TikTok platform, such as "https: / / www.tiktok.com / video / 123456789". This link is the target short video link, which clearly points to a specific short video. The server identifies and stores this link as a basis for subsequent operations.

[0022] After receiving the target short video link, the server starts to call the preset Internet celebrity big data service (KOL_BigData). To obtain the effect data, the server sends a query request to KOL_BigData, and KOL_BigData searches for the record corresponding to the target short video link from its huge video library (which stores billions of video data). For example, it is found that this short video was launched on May 1, 2023, and the current number of views is 100,000, the number of likes is 5,000, the number of comments is 1,000, and the number of favorites is 2,000.

[0023] For the basic content, KOL_BigData returns the target short video with the title “A great place for a summer trip”, the tags are “travel”, “summer”, and “tourist attractions”, and the audio is a cheerful background music plus some natural environment sound effects.

[0024] As for the background information, it was found that the creator was a user named "Travel Little Expert" and the top competing videos were several other popular short videos with the same travel theme. The creators of these top competing videos were "Travel Enthusiast" and "World Traveler" respectively.

[0025] The server will obtain the performance data such as the launch time (May 1, 2023), playback volume (100,000 times), likes (5,000 times), comments (1,000 times), and collections (2,000 times), the basic content such as the title ("a great place for summer travel"), tags ("travel", "summer", "tourist attractions"), audio (cheerful background music + natural sound effects), and background information such as the creator ("little travel expert"), top competitive videos (competitive video links and related information), and the creator of the top competitive videos ("travel enthusiast" and "world traveler"), to form a basic information package for the target short video. This basic information package will be used for further analysis and processing in subsequent steps.

[0026] The server calls the preset Internet celebrity language model (KOL_LLM), which is fine-tuned based on the version of Llama3-70B 1M context window. After receiving the basic information, KOL_LLM calls the transcription function in its toolset to transcribe the audio of the target short video. For example, the natural environment sound effects in the audio are ignored, the background music is identified as atmosphere, and the commentary voice of the characters is transcribed as lines such as "Welcome to this charming summer travel destination, where there are blue sea and sky, golden beaches and rich food."

[0027] KOL_LLM began to analyze the previously obtained title "A great place for summer travel", the tags "travel", "summer", "tourist attractions", and the transcribed lines "Welcome to this charming summer travel destination, with blue sea and sky, golden beaches and rich food". Through semantic analysis of these contents, KOL_LLM determined that the theme of this short video is "Summer beach travel recommendation".

[0028] First, KOL_LLM operates in the preset KOL_Graph based on all the above information (performance data, basic content, background information, lines and themes). KOL_Graph is built based on tens of millions of short videos.

[0029] For the large-category cluster matching, since the theme is "Summer beach travel recommendation", KOL_LLM matches it with the large-category clusters in KOL_Graph, and may match the large-category cluster of "travel". Then, sub-category cluster matching is performed under the large-category cluster of "travel", and the sub-category cluster of "beach travel" may be matched. Then, attitude cluster matching is performed, for example, matching to an attitude cluster such as "enthusiastic recommendation for beach travel". Finally, topic cluster matching is performed to find the topic cluster most relevant to "Summer beach travel recommendation" from the attitude cluster, such as a topic cluster containing related entities such as specific beach locations and local food.

[0030] Through these matches, according to the preset relevance threshold, KOL_LLM obtains target entities from the topic cluster, such as the specific beach name "Bali Beach", the local specialty food "Seafood BBQ", and related text blocks, such as the text block containing other user reviews of this beach. These target entities and target text blocks constitute deep information.

[0031] The server inputs the previously integrated basic information (including various information such as listing time and playback volume) and in-depth information (such as entities such as "Bali Beach" and "Seafood BBQ" and related text blocks) into KOL_LLM. KOL_LLM analyzes and processes the information.

[0032] For example, KOL_LLM may give optimization suggestions such as: new title "Bali Beach: A dream choice for summer travel, enjoy seafood barbecue", new tags add more specific tags such as "Bali" and "Seafood", and new lines can add some descriptions about the uniqueness of Bali Beach, such as "The sunset at Bali Beach is one of the most beautiful sunsets in the world, and paired with local fresh seafood barbecue, it is simply the ultimate experience of summer travel." These optimization suggestions will help increase the attractiveness of short videos, which may increase various indicators such as playback volume and likes.

[0033] In the embodiment of the present invention, calling the preset Internet celebrity big data service to obtain the basic information of the target short video can be implemented through the following example.

[0034] Calling a preset Internet celebrity big data service to obtain the effect data, basic content and background information of the target short video; the effect data includes the release time, playback volume, number of likes, number of comments and number of favorites of the target short video; the basic content includes the title, tags and audio of the target short video; the background information includes the creator of the target short video, the top competing videos and the creator of the top competing videos; Based on the effect data, the basic content and the background information as the basic information of the target short video.

[0035] In the embodiment of the present invention, illustratively, after receiving the link of the target short video, the server starts to call the preset Internet celebrity big data service (KOL_BigData). The server sends a query request containing the target short video identifier (such as a unique identifier or link of the video) to KOL_BigData, specifically querying the effect data part.

[0036] KOL_BigData's video library stores a massive amount of video information. After receiving the query request from the server, it quickly locates the record of the target short video in its data storage structure. For example, for a short video about food making, KOL_BigData finds that the video was released on March 15, 2023. This date information is very important for analyzing the timeliness of the video and its dissemination in different time periods.

[0037] Next, we searched the playback data and found that the video has been played 80,000 times. The playback volume is a key indicator to measure the popularity of a video, which can reflect the scope of the video's dissemination among the audience. The likes data showed 3,000 times, which is a direct expression of the audience's recognition and love for the video content. There are 500 comments, and the number of comments reflects the degree of interaction between the audience and the video content, which may include questions about the steps of making food, comments on the taste of dishes, or discussions about the creator's creativity. The number of collections is 1,500, which shows that a considerable number of viewers think that this video is worth watching again or referring to, perhaps because they want to try to make the food in the video.

[0038] The server continues to obtain the basic content of the target short video through KOL_BigData.

[0039] For the title part, KOL_BigData returned the title "Make delicious pasta in ten minutes". This title concisely and clearly conveys the core content of the video, which is the time efficiency of making pasta and the name of the dish.

[0040] Looking at the tag part, the tags found are "food making", "pasta", "10-minute food", and "easy recipes". These tags classify and describe the video content from different dimensions, which helps to accurately locate the video in the search and recommendation system. For example, when users search for "pasta" or "food that can be made in 10 minutes" on the platform, videos with these tags are more likely to be recommended.

[0041] Regarding the audio part, KOL_BigData provided relevant information about the audio. The audio of this food production video includes soft background music with a brisk rhythm, which matches the relaxed atmosphere of food production; there are also sounds during the production process, such as the sound of cutting vegetables, the gurgling sound of boiling noodles, and the sound of stirring sauces. These sound elements together constitute the audio content of the video, bringing a richer viewing experience to the audience.

[0042] In terms of obtaining background information, the server retrieves from KOL_BigData that the creator of the target short video is a user named "Gourmet Little Chef". This creator identity information is of great significance for understanding the video's style, quality, and audience. For example, "Gourmet Little Chef" may be well-known on the platform for making simple and delicious home-cooked meals, and her subscribed users may have high expectations and preferences for this type of food-making video.

[0043] Meanwhile, KOL_BigData also returns relevant information about the top competitor videos. Top competitor videos refer to popular videos that compete for attention and traffic with the target short video under the same theme (Food Making - Pasta). For example, the title of one of the top competitor videos is "Super Authentic Pasta Making Tutorial", and its creator is "Italian Gourmet". This competitor video may have unique features in terms of production techniques, ingredient selection, or video shooting methods, attracting a large number of viewers. By analyzing the competitor videos, the differences and advantages between the target short video and the competitors can be identified, providing a reference for optimizing the target short video.

[0044] The server integrates all the information retrieved from KOL_BigData to form the basic information of the target short video.

[0045] First is the part of the effect data. The server stores the release time (March 15, 2023), the number of views (80,000 times), the number of likes (3,000 times), the number of comments (500), and the number of collections (1,500 times) according to a specific data structure. For example, it is represented by a JSON-formatted data object as follows: { "release time": "2023-03-15", "views": 80000, "likes": 3000, "comments": 500, "collections": 1500 } For the basic content part, the title ("Make Delicious Pasta in Ten Minutes"), the tags ("Food Making", "Pasta", "Ten-Minute Food", "Simple Recipe"), and the audio information (description of the background music and production sounds) are integrated. The representation in a similar data structure may be as follows: { "title": "Make Delicious Pasta in Ten Minutes", "tags": ["Food Making", "Pasta", "Ten-Minute Food", "Simple Recipe"], "audio": { "Background music": "Soft and light music", "Make sounds": ["Chopping vegetables", "Cooking noodles", "Stirring sauce"] } } Finally, the creator in the background information ("Little Gourmet Cook"), the top competitor video (titled "Super Authentic Pasta Making Tutorial", created by "Italian Gourmet") and their creator information are integrated: { "Creator": "Gourmet Little Cook", "Top Competitive Video": [ { "Title": "Super authentic pasta making tutorial", "Creator": "Italian Foodie" } ] } Then, the server combines these three parts (effect data, basic content, and background information) into a complete basic information package, which will serve as the basis for subsequent analysis and processing of the target short video. For example, a larger JSON object is used to represent the entire basic information package: { "Effect Data": { "Availability time": "2023-03-15", "Play Count": 80000, "Likes": 3000, "Number of comments": 500, "Collections": 1500 }, "Basic content": { "Title": "Make delicious pasta in 10 minutes", "Tags": ["food preparation", "pasta", "10-minute meals", "easy recipes"], "audio": { "Background music": "Soft and light music", "Make sounds": ["Chopping vegetables", "Cooking noodles", "Stirring sauce"] } }, "Background Information": { "Creator": "Gourmet Little Cook", "Top Competitive Video": [ { "Title": "Super authentic pasta making tutorial", "Creator": "Italian Foodie" } ] } } This basic information package completely describes the various features and relevant background of the target short video, providing comprehensive data support for the subsequent call of the Internet celebrity language model to extract deep information and ultimately obtain optimization suggestions.

[0046] In the embodiment of the present invention, based on the basic information, calling a preset internet celebrity language model to extract information from the target short video to obtain depth information can be implemented through the following examples.

[0047] The target short video is transcribed by calling a preset tool set through the Internet celebrity language model, and the audio is transcribed into lines; Analyze the title, the tag, and the lines through the Internet celebrity language model to obtain the theme of the target short video; The deep information is obtained from the preset internet celebrity knowledge graph through the internet celebrity language model based on the effect data, the basic content, the background information, the lines and the theme.

[0048] In the embodiment of the present invention, exemplarily, after acquiring the basic information of the target short video, the server starts to call the preset Internet celebrity language model (KOL_LLM). KOL_LLM is fine-tuned based on the version of Llama3-70B 1M context window, and it has rich functions and pre-trained knowledge.

[0049] After KOL_LLM receives the audio part of the basic information sent by the server, it calls its preset tool set to start transcribing the audio. Taking the food production short video mentioned earlier as an example, the audio contains soft background music and various sounds during the production process. KOL_LLM's transcription tool will first pre-process the audio to separate the background music and production sounds. Since the background music is mainly used to create an atmosphere and has no direct contribution to the expression of the content, the tool will focus on identifying the vocal part of the production sound.

[0050] For example, in the process of making pasta, the creator may explain while doing it: "First, we prepare fresh pasta, boil water, and add a small spoonful of salt, so that the noodles will be more chewy." The transcription tool will accurately convert these sounds into text lines: "First, prepare fresh pasta, boil water, and add a small spoonful of salt to make the noodles chewy." In the process of cooking noodles, the creator said: "Now we can see the noodles rolling in the pot, let's turn down the heat and let it cook slowly." The transcription tool will also transcribe it as: "Now see the noodles rolling in the pot, turn down the heat and cook slowly." As the production process progresses, the creator will also introduce the making of the sauce: "Next, we will make a delicious tomato sauce. Put the chopped tomatoes into the pot, add some olive oil, minced garlic and basil leaves, and simmer slowly to fully release the tomato flavor." The transcription tool will also accurately transcribe it as: "Next, make tomato sauce, chopped tomatoes into the pot, add olive oil, minced garlic, basil leaves, and simmer slowly to release the tomato flavor." During the entire transcription process, KOL_LLM's transcription tool will identify the vocal content in the audio as accurately as possible and convert it into complete and fluent lines. These lines will provide an important basis for subsequent analysis of the theme of the video and obtaining in-depth information from the knowledge graph.

[0051] After receiving the transcribed lines, KOL_LLM began to conduct a comprehensive analysis of the target short video’s title, tags, and lines to determine the theme.

[0052] For the title “Delicious pasta in 10 minutes”, the time limit for making pasta (10 minutes) and the dish (pasta) are clearly mentioned. The tags “food making”, “pasta”, “10-minute food”, “easy recipe” further emphasize that this is a video about food making, especially pasta making, and emphasizes the simplicity and time efficiency of making it.

[0053] Let's look at the transcribed lines, such as "First, prepare fresh pasta, boil water, and add a teaspoon of salt to make the noodles chewy. Now you see the noodles tumbling in the pot, turn down the heat and cook slowly. Next, make the tomato sauce, chop the tomatoes into the pot, add olive oil, minced garlic, and basil leaves, and simmer to release the tomato flavor." These lines describe the process of making pasta in detail.

[0054] KOL_LLM conducts in-depth analysis of the semantic information in the title, tags, and lines, and identifies key elements, such as "pasta" is the core dish, "production process" is the main behavior, "ten minutes" is the time feature, and "simple" is the difficulty feature of production. Combining this information, KOL_LLM determines the theme of this short video as "Simple ten-minute pasta production process". This theme accurately summarizes the core content of the video, covering the dishes, the ease of production, and the time elements, providing a clearer direction for obtaining in-depth information from the knowledge graph in the future.

[0055] After determining the topic, KOL_LLM starts to search for information in the preset Internet celebrity knowledge graph (KOL_Graph) based on the previously obtained effect data, basic content, background information, lines and topics. KOL_Graph is a four-layer structure graph built based on tens of millions of short videos, including major category clusters, subcategory clusters, attitude clusters and theme clusters.

[0056] First, we match the large clusters. Since the theme of the target short video is "Easy ten-minute pasta making process", which belongs to the category of food making, KOL_LLM matches this theme with the large clusters in KOL_Graph. In KOL_Graph, there may be a large cluster called "food", which contains various sub-category content related to food. This "food" large cluster covers knowledge entities in different types of food making, food culture, food recommendations, etc. Through semantic analysis, KOL_LLM found that "Easy ten-minute pasta making process" is highly correlated with the "food" large cluster, and thus determined to conduct further searches under this large cluster.

[0057] After determining the "food" cluster, KOL_LLM continues to match sub-cluster categories. Under the "food" cluster, there may be multiple sub-cluster categories, such as "Western food making", "Chinese food making", "dessert making", etc. Since the target short video is about the making of pasta, which belongs to the category of Western food, KOL_LLM will match the topic with these sub-cluster categories. Through semantic analysis, it will find the "Western food making" sub-cluster, which contains more specific knowledge entities, relationships, and text blocks related to the making of Western food dishes. For example, this sub-cluster may contain information about the making of different types of pasta (such as macaroni, spiral noodles, etc.), as well as related information such as ingredients, seasonings, and kitchen utensils commonly used in Western food making.

[0058] After finding the subcategory cluster of "Western food preparation", KOL_LLM then matches the attitude cluster. An attitude cluster is a cluster that is closer to the user scenario, and it gathers together the claims and opinions on entities. In the attitude cluster of "Western food preparation", there may be different attitude clusters such as "pursuing efficient Western food preparation", "inheritance of traditional Western food preparation", and "creative Western food preparation". Since the target short video emphasizes the "ten minutes" production time and reflects an efficient production method, KOL_LLM will match the theme with the attitude cluster of "pursuing efficient Western food preparation". This attitude cluster may contain some knowledge entities, relationships, and text blocks related to tips and experience sharing on how to make delicious Western food in a short time, as well as audience feedback and evaluations on efficient Western food preparation.

[0059] After completing the attitude cluster matching, KOL_LLM performs topic cluster matching. From the attitude cluster of "Pursuing Efficient Western Food Production", KOL_LLM will look for the most suitable topic cluster. Topic cluster-the bottom layer is a cluster composed of entities that have reached a certain degree of relevance. In this case, there may be a topic cluster called "Efficient Pasta Production and Innovation". This topic cluster contains various entities related to efficient pasta production, such as specific pasta brands, suppliers of ingredients suitable for fast production, and audience comparisons of different pasta production videos. Through semantic analysis, KOL_LLM determines that this topic cluster is highly relevant to the theme of the target short video.

[0060] After completing the above matching, KOL_LLM obtains the target entity and target text segment from the topic cluster "Efficient Pasta Production and Innovation" according to the preset relevance threshold. For example, the target entity may include a specific pasta brand "XX Brand Pasta" because this brand may have unique advantages in efficient pasta production, such as easy cooking and good taste; it may also include a special seasoning "Quick Tasting Italian Seasoning" that can add rich flavor to pasta in a short time. The target text segment may include some users' comments on the quick preparation of "XX brand pasta", such as "XX brand pasta is really convenient. You can make delicious pasta in ten minutes by following the method in the video", or some tips on using "quick-tasting Italian seasoning", such as "When making ten-minute pasta, add this quick-tasting Italian seasoning. Only a small amount is needed to make the noodles full of flavor."

[0061] These target entities and target text blocks constitute the deep information of the target short video, which will provide a richer and more detailed basis for the subsequent generation of optimization suggestions.

[0062] In an embodiment of the present invention, the preset Internet celebrity knowledge graph is constructed in the following manner and can be implemented through the following examples.

[0063] Obtaining video information and account information of a plurality of stored videos stored by the preset internet celebrity big data service through the internet celebrity big language model; Structuring the video information and the account information of each stored video to obtain a text block corresponding to each stored video, wherein the text block includes the video information and the account information; Inputting the multiple text blocks into the Internet celebrity language model for multiple rounds of extraction to obtain entities, relations, and covariates corresponding to each of the multiple text blocks; Multi-layer cluster extraction is performed based on the entities, the relationships, and the covariates to obtain the preset Internet celebrity knowledge graph, which includes a four-layer graph structure constructed from top to bottom by major category clusters, subcategory clusters, attitude clusters, and theme clusters.

[0064] In the embodiment of the present invention, the server first starts the process of building a preset Internet celebrity knowledge graph (KOL_Graph). It calls the Internet celebrity big language model (KOL_LLM) to obtain relevant information of multiple stored videos stored in the preset Internet celebrity big data service (KOL_BigData).

[0065] For each stored video, KOL_BigData stores rich information. For example, a fitness short video may include the title of the video, "Efficient home fitness exercises, fast fat burning", tags, "fitness exercises", "home fitness", "fat burning exercise", audio content including brisk sports background music and coach's instructions, 50,000 views, 2,000 likes, 800 comments, 1,000 favorites, and listing time of April 10, 2023.

[0066] In terms of account information, the creator of this fitness short video is named "Fitness Expert Xiao A". The account has 100,000 subscribers. The account's focus areas are fitness, healthy lifestyle, etc. The account is highly active, posting an average of 3-4 fitness-related videos per week.

[0067] KOL_LLM obtains the video information and account information of these stored videos one by one. This process is like an information collector, systematically extracting the relevant information of each video from the huge information database of KOL_BigData to prepare for subsequent processing. This process involves a huge amount of video data, which may be hundreds of thousands or even millions of videos. KOL_LLM performs information acquisition operations on each video in sequence to ensure that no important information is missed.

[0068] After obtaining the video information and account information of each video, the server begins to structure the information.

[0069] Still taking the fitness short video as an example, the server will organize the video information into a structured text block according to a certain format. For example: “Title: Efficient home aerobics, fast fat burning; Tags: aerobics, home fitness, fat burning exercise; Audio: brisk exercise background music + coach’s commands; Effect data: playback volume - 50,000 times, likes - 2,000 times, comments - 800 times, collections - 1,000 times, listing time - 2023-04-10.” For account information, it will be organized into a similar format: “Account name: Fitness expert Xiao A; Number of subscribers: 100,000; Areas of interest: Fitness, healthy lifestyle; Account activity: 3-4 fitness-related videos posted per week.” Then combine these two pieces of information into a complete text block: “Title: Efficient home aerobics, fast fat burning; Tags: aerobics, home fitness, fat burning exercise; Audio: brisk exercise background music + coach’s commands; Effect data: 50,000 views, 2,000 likes, 800 comments, 1,000 favorites, published on April 10, 2023; Account name: Fitness expert Xiao A; Number of subscribers: 100,000; Areas of focus: fitness, healthy lifestyle; Account activity: 3-4 fitness-related videos are released every week.” This text block contains the complete information of the video itself and the account information, and is presented in a clear and orderly structure. For each video in KOL_BigData, the server will perform such structural processing and convert it into the corresponding text block. This process needs to be carried out strictly in accordance with the predetermined structural rules to ensure that the format of each text block is unified, which is convenient for subsequent processing and analysis.

[0070] The server inputs a large number of text blocks (assuming 100,000 processed text blocks) into KOL_LLM for the first round of basic extraction.

[0071] For each text segment, KOL_LLM will mine the entities, relations and covariates (entity attributes) in it. Taking the previous fitness short video text segment as an example, KOL_LLM may identify entities such as "family fitness exercises" (as the core content entity of the video), "fitness expert Xiao A" (creator entity), "fat burning exercise" (concept entity related to fitness exercises), "fitness enthusiasts" (implicit potential audience entity watching the video), etc.

[0072] In terms of relationships, it will recognize that there is a "creator-work" relationship between "fitness expert Xiao A" and "family aerobics", a "belongs to" relationship between "family aerobics" and "fat burning exercises", and an "audience-work" relationship between "fitness enthusiasts" and "family aerobics".

[0073] In terms of covariates (entity attributes), for the entity "family aerobics", the attributes may include "high efficiency" (an attribute describing the effect of aerobics) and "suitable for home environment" (an attribute describing the applicable scenarios of aerobics); for the entity "fitness expert Xiao A", the attributes include "number of subscribers - 100,000", "areas of interest - fitness, healthy lifestyle", etc.

[0074] This process is a parallel processing of a large number of text blocks. KOL_LLM will quickly scan the information in each text block and use its pre-trained knowledge and algorithms to accurately mine the entities, relations and covariates. This process may consume certain computing resources, but because KOL_LLM has been optimized, it can complete the first round of basic extraction operations on a large number of text blocks in a reasonable time.

[0075] After the first round of basic extraction is completed, the server will send a signal to KOL_LLM to extract again and start the second round of extraction.

[0076] KOL_LLM will dig deeper in the second round of extraction. For example, for the entity "home aerobics" identified in the first round, it will further analyze the more complex relationships that may exist between it and other entities. It may be found that there is a "matching use" relationship between "home aerobics" and certain specific fitness equipment (such as yoga mats, dumbbells, etc.). Although these fitness equipment are not directly mentioned in the original text segmentation, KOL_LLM has mined this potential relationship through its knowledge reserve and semantic analysis capabilities.

[0077] The covariates (attributes) of the entity will also be mined more deeply. For example, for the entity "fitness enthusiast", its deeper attributes such as "fitness goals - weight loss, body shaping" may be mined. These attributes may be inferred by analyzing the comments in the video or based on the general understanding of the fitness field.

[0078] The second round of extraction process is a deep mining based on the first round, aiming to discover more hidden entities, relationships and more detailed covariates, and further enrich the information content corresponding to each text block.

[0079] After the first two rounds of extraction, the server starts the third round of merging. In this round, KOL_LLM will remove duplicates and merge all entity names identified in the previous two rounds.

[0080] For example, the entity "aerobics" may be mentioned multiple times in different text blocks, but there may be different expressions, such as "family aerobics", "efficient aerobics", etc. KOL_LLM will merge these entities with different expressions into the entity "aerobics", and integrate its related relations and covariates.

[0081] The same is true for commutative relationships. There may be multiple similar relationships expressed, such as "the creator produced the work" and "the work was created by the creator". KOL_LLM will merge and unify these relationships.

[0082] Through the third round of merging operations, a complete, accurate, and non-redundant combination of entities, relations, and covariates corresponding to each text block is finally generated. This makes the information of each text block more refined and easy to understand, providing a high-quality data foundation for the subsequent construction of the knowledge graph.

[0083] The server starts to extract multi-layer clusters based on the entities, relations, and covariates corresponding to each text block. First, the large-category clusters are constructed.

[0084] Among all entities, entities with broad commonalities are grouped into a large cluster. For example, entities such as "aerobics", "yoga courses", and "aerobics courses" are grouped into the large cluster of "fitness courses" because they all belong to the course content in the fitness field. Entities such as "fitness expert Xiao A", "fitness coach Xiao B", and "fitness influencer Xiao C" are grouped into the large cluster of "fitness creators" because they are all people who create content in the fitness field.

[0085] The construction of these large clusters is based on the high-level classification of entities, aiming to build the top-level structure of the knowledge graph to facilitate the subsequent organization and retrieval of information under more specific classifications.

[0086] After building the large cluster, the server continues to build the sub-cluster.

[0087] Take the "fitness courses" cluster as an example, and build sub-cluster under it. "Aerobics" is a separate sub-cluster, because aerobics has its unique characteristics and audience in fitness courses; "Yoga courses" is also a sub-cluster, which has a different movement system, exercise purpose and audience from aerobics; "Aerobics courses" are further divided into smaller sub-cluster according to different forms of exercise (such as running courses, swimming courses, etc.).

[0088] For the large cluster of "fitness creators", sub-clusters are divided according to the creators' professional fields or creative styles. For example, creators who focus on strength training teaching are classified into the sub-cluster of "strength training creators", and creators who are good at yoga teaching are classified into the sub-cluster of "yoga creators", etc.

[0089] The construction of sub-category clusters enables the knowledge graph to be further refined on the basis of major categories, and can more accurately locate and organize related knowledge entities, relationships, and covariates.

[0090] Next, we construct attitude clusters. Under the "aerobics" sub-cluster, there may be different attitude clusters.

[0091] For example, there is an attitude cluster of "advocating the weight loss effect of aerobics", which contains entities, relations, and covariates related to videos that emphasize the significant effect of aerobics in weight loss. For example, in some videos, coaches emphasize that aerobics can quickly burn fat, and there are many cases of students who have successfully lost weight. This information will be classified into this attitude cluster.

[0092] There is also an attitude cluster of "focusing on the fun of aerobics", which contains video-related content that highlights the interesting and easy-to-stick characteristics of aerobics. For example, some aerobics videos add pop music and interesting movement designs, and the entities, relations, and covariates involved in these videos will be classified into this attitude cluster.

[0093] The construction of attitude clusters is closer to users' actual needs and views on things, and can directly reflect the distribution of knowledge content under different attitudes.

[0094] Finally, the theme cluster - bottom layer is constructed. The theme cluster is constructed under the attitude cluster of "advocating the weight loss effect of aerobics".

[0095] For example, there is a topic cluster called "Efficient combinations of aerobics for weight loss", which contains entities, relationships, and covariates related to videos that specifically introduce which combinations of aerobics movements can effectively lose weight. It may include specific aerobics movement names (such as squat jumps, high leg raises, etc.), the relationship between these movements and weight loss effects (such as how many calories are consumed per minute by squat jumps, etc.), and related information such as the demonstration and explanation of these movement combinations in different videos.

[0096] Through this top-down construction, we finally get a four-layer structure of pre-set Internet celebrity knowledge graph (KOL_Graph) including major category clusters, subcategory clusters, attitude clusters and theme clusters. This knowledge graph provides a rich knowledge reserve and powerful search tools for subsequent short video analysis and optimization operations.

[0097] In an embodiment of the present invention, the deep information is obtained from the preset internet celebrity knowledge graph based on the effect data, the basic content, the background information, the lines and the theme through the internet celebrity language model, which can be implemented through the following examples.

[0098] Based on the effect data, the basic content, the background information, the lines and the theme, the Internet celebrity language model performs large category cluster matching, subcategory cluster matching, attitude cluster matching and topic cluster matching in the preset Internet celebrity knowledge graph, and obtains the target entity and target text block according to the preset relevance threshold; The target entity and the target text are divided into blocks as the depth information.

[0099] In an embodiment of the present invention, for example, assuming that we are processing a travel short video, the server has obtained the effect data of the short video (such as 150,000 views, 8,000 likes, 2,000 comments, 3,000 collections, and the listing time is July 5, 2023), basic content (titled "Explore the Mysterious Bali", tags are "Bali", "Travel", "Tropical Style", audio includes the sound of waves, local folk music and the tour guide's explanation), background information (the creator is a user named "Travel Explorer", and the top competing videos are other popular videos about Bali travel, and the creators are "Bali Travel Experts", etc.), lines (obtained through transcription, such as "Now we have come to Kuta Beach in Bali, the beach here is fine and soft, and the waves are very suitable for surfing enthusiasts", etc.) and theme (determined to be "Bali Travel Experience").

[0100] The server calls the KOL_LLM, takes this information as input, and matches the large clusters in the preset KOL_Graph. The large clusters in the KOL_Graph include multiple broad categories, such as "travel", "food", "fitness", etc. KOL_LLM performs semantic analysis on the input information. Since the theme is "Bali travel experience", it clearly belongs to the "travel" cluster. Therefore, KOL_LLM determines to perform further matching operations under the "travel" cluster. This process is like finding the correct large classification area in a huge knowledge warehouse first, laying the foundation for more accurate searches later.

[0101] After determining the "tourism" cluster, KOL_LLM began to match sub-cluster categories. Under the "tourism" cluster, there are many sub-cluster categories, such as "domestic tourism", "foreign tourism", "island tourism", "mountain tourism", etc.

[0102] KOL_LLM analyzes the relevant information of the short video. Since the theme is "Bali Travel Experience" and Bali is an island, it will focus on the sub-cluster of "Island Tourism". This sub-cluster contains various entities, relationships, and text blocks related to island tourism. For example, tourist attractions on different islands, seasons suitable for island tourism, and special activities for island tourism (such as diving, sea fishing, etc.). Through this matching, KOL_LLM further narrows the search scope and locates more specific sub-cluster from the many possibilities under the large cluster in order to obtain more relevant in-depth information.

[0103] After finding the sub-cluster of "island tourism", KOL_LLM then performs attitude cluster matching. In the attitude cluster of "island tourism", there may be different attitude clusters such as "enjoying the tranquility and beauty of the island", "being keen on island adventure activities", and "pursuing a luxurious holiday experience on the island".

[0104] By analyzing the lines in the short video, such as the description of enjoying the sun and the beautiful sea view on the beach in Bali, without too many descriptions of adventurous activities, KOL_LLM will match it with the attitude cluster of "enjoying the tranquility and beauty of the island". This attitude cluster may include relevant knowledge entities, relationships, and text blocks such as tourists' love for the tranquil atmosphere of the island, praise for the beautiful sea view, and some recommended places and activities suitable for enjoying tranquility and beauty. Through this matching, KOL_LLM is closer to the emotions and experiences conveyed by the short video, thereby more accurately locating the knowledge content related to the video content.

[0105] After completing the attitude cluster matching, KOL_LLM conducts the theme cluster matching. From the attitude cluster of "enjoying the tranquility and beauty of the island", KOL_LLM searches for the most suitable theme cluster.

[0106] For example, there may be a theme cluster called "The Tranquil Beauty of Bali Beaches". This theme cluster includes information about the characteristics of various beaches in Bali (such as Kuta Beach, Nusa Dua Beach, etc.), the specific manifestations of the tranquil atmosphere on the beach (such as the sunset view in the evening, the tranquil sea in the morning, etc.), relaxing activities that can be done on the beach (such as beach yoga, seaside walks, etc.), and tourists' evaluations of the tranquil beauty of these beaches. Through a comprehensive analysis of the short video, KOL_LLM determined that this theme cluster is highly correlated with the theme of the short video, "Bali Travel Experience".

[0107] After completing all the above cluster matching, KOL_LLM obtains the target entity and target text segment from the topic cluster "The Tranquil Beauty of Bali Beaches" according to the preset relevance threshold.

[0108] Assuming the preset relevance threshold is 0.8, KOL_LLM will calculate and evaluate the relevance of each entity and text block in the topic cluster to the short video. The target entity may include the specific location entity "Kuta Beach" because Kuta Beach is mentioned in the short video and it is one of the representative beaches in Bali; it may also include the activity entity "Beach Yoga" because it is a relaxing activity that can be performed on the beach.

[0109] The target text segment may contain tourists' comments on Kuta Beach: "Kuta Beach is one of the most charming beaches in Bali. The sand here is fine and the waves gently lapping the shore, making people feel extremely relaxed, especially in the evening, watching the beautiful sunset, it's like time has stood still." This text segment is highly relevant to the theme of the short video and meets the preset relevance threshold.

[0110] In this way, KOL_LLM filters out target entities and target text blocks that are closely related to short videos from the preset Internet celebrity knowledge graph. These will serve as the in-depth information of the short video and provide detailed and targeted basis for subsequent optimization suggestions.

[0111] After receiving the target entities (such as "Kuta Beach" and "Beach Yoga") and target text blocks (such as tourists' comments on Kuta Beach) filtered out by KOL_LLM, the server stores and subsequently processes them as the depth information of the target short video.

[0112] This depth information is integrated into a dedicated data structure, for example, in JSON format: { "Depth Information": { "Target Entity": ["Kuta Beach", "Beach Yoga"], "target text chunk": ["Kuta Beach is one of the most charming beaches in Bali. The sand here is fine and the waves gently lapping the shore, making people feel extremely relaxed, especially in the evening, watching the beautiful sunset, as if time has stopped."] } } This deep information data structure will be used together with the basic information obtained previously for subsequent operations, such as input into the Internet celebrity language model to generate optimization suggestions for the target short video. By using the target entity and target text blocks as deep information, the server can use this richer and more targeted information to deeply understand the characteristics and advantages of short videos, thereby providing more valuable reference for optimizing short videos.

[0113] In an embodiment of the present invention, the preset Internet celebrity language model is obtained in the following manner and can be implemented through the following examples.

[0114] Acquire a basic large language model and an advanced large language model, wherein the model parameter amount of the advanced large language model is greater than that of the basic large language model; The advanced large language model performs analysis, self-verification, and self-selection extraction strategies on sample short videos based on a self-discovery mechanism, and generates knowledge extraction data through prompt engineering; Based on the knowledge extraction data, the basic large language model is fine-tuned in a LoRA manner to obtain the preset Internet celebrity large language model.

[0115] In the embodiment of the present invention, exemplarily, in the process of constructing the preset KOL_LLM, the server first needs to obtain a basic large language model and an advanced large language model.

[0116] The server selects a suitable basic large language model from the model repository. For example, Llama3-70B is selected as the basic large language model. This model has certain pre-trained knowledge and capabilities, and can handle various natural language processing tasks. Its structure and parameters are obtained through pre-training on a large amount of text data, and contain rich semantic information and language patterns. Llama3-70B is selected as the basic model because it performs well in text generation, multi-task support, reasoning, general performance, and has the advantage of open source and easy deployment in the server environment.

[0117] At the same time, the server obtains an advanced large language model, which has a larger number of model parameters than the basic large language model. For example, a model with a larger number of parameters is selected, such as a large language model optimized for knowledge extraction tasks, whose parameter volume reaches 100 billion parameters (this is just an example, the actual model depends on the specific situation). This advanced large language model has stronger capabilities in handling complex semantic analysis and knowledge mining tasks. Due to its larger number of parameters, it can accommodate more knowledge and language patterns, thereby providing a deeper and more comprehensive understanding when analyzing short video related content.

[0118] The server selects a batch of sample short videos from the preset short video sample library for the generation of knowledge extraction data. This sample short video library contains various types of short videos, covering multiple fields such as tourism, food, fitness, and entertainment. For example, a sample short video of tourism was selected, titled "Exploring Ancient Kyoto", with tags such as "Kyoto", "Japanese Tourism", and "History and Culture". The audio includes the noise of Kyoto streets, the bells of temples, and the explanations of tour guides. The video content shows Kyoto's ancient buildings, traditional tea ceremony performances, etc.

[0119] The server inputs the relevant information of this travel sample short video into the advanced large language model. The advanced large language model starts to analyze the short video based on the self-discovery mechanism.

[0120] For the title of the video "Exploring Ancient Kyoto", the model analyzed through the self-discovery mechanism that "Kyoto" is an important entity and is associated with the attribute "ancient", which may imply that the historical and cultural value of Kyoto is a core content of this video. For the tags "Kyoto", "Japan Travel", and "History and Culture", the model further understands that this video is mainly about traveling in Kyoto, Japan, and emphasizes its historical and cultural characteristics.

[0121] When analyzing the audio content, the model recognizes that the temple bell is a symbolic element of Kyoto culture, and the guide's explanation may contain important information about Kyoto's historical buildings and cultural traditions. For the ancient buildings and tea ceremony performances shown in the video content, the model self-discovered that these are key elements that reflect Kyoto's history and culture. Through this analysis based on the self-discovery mechanism, the model is able to automatically mine the latent semantic information in short videos without the need to predefine clear rules or patterns.

[0122] Based on the analysis, the advanced large language model begins to self-verify. For example, for the previously identified conclusion that "Kyoto's historical culture is the core content", the model will once again check whether all elements in the video support this conclusion.

[0123] The model will check whether the architectural style in the video, the details of the tea ceremony performance, and the content of the tour guide's explanation are closely related to Kyoto's history and culture. If some elements are found to be contradictory or irrelevant, the model will readjust its analysis results. For example, if the tour guide suddenly mentions Kyoto's modern commercial development in the explanation, which deviates from the previously identified core content of history and culture, the model will further analyze the relationship between this part and the whole, and may determine that this is a comparison or supplementary explanation, rather than a change in the core content, thereby verifying the previous conclusion that history and culture are the core content.

[0124] After analysis and self-validation, the advanced large language model uses a self-selected extraction strategy to determine the knowledge content to be extracted.

[0125] For this sample travel short video, the model automatically selects entities, relationships, and attributes directly related to Kyoto's history and culture as knowledge content. For example, it selects "Kinkaku-ji Temple" (as a famous historical building entity in Kyoto), "tea ceremony" (as a traditional art form entity in Kyoto), "Kinkaku-ji Temple-represents Kyoto's historical and cultural heritage" (relationship), "tea ceremony-embodies Kyoto's traditional culture" (relationship) and other knowledge content. Based on its judgment of the importance of the short video content, the model automatically selects these knowledge elements that best represent the core meaning of the video, while ignoring some minor content or content that is not closely related to the core theme, such as the occasional image of passers-by in the video.

[0126] After self-selecting knowledge content, the advanced large language model generates knowledge extraction data through prompt engineering.

[0127] The server provides the model with a series of prompts to guide the generation of knowledge extraction data. For example, for the previously selected "Kinkaku-ji" entity, the prompt may be "Please describe in detail the status of Kinkaku-ji in Kyoto's history and culture, its architectural features, and its relationship with other attractions." Based on this prompt, the model generates detailed knowledge extraction data about Kinkaku-ji, such as "Kinkaku-ji is one of Kyoto's most representative historical and cultural heritages. It was built in the 14th century and its architectural feature is its golden appearance, which incorporates traditional Japanese architectural style. Compared with other attractions in Kyoto, Kinkaku-ji is one of the must-visit places for tourists. The surrounding garden landscape also complements the architecture, together showing the charm of Kyoto's ancient culture." Similarly, for the entity "tea ceremony", the prompt may be "Please explain the significance of tea ceremony in Kyoto's traditional culture, its process, and its connection with the lives of local residents", and the model generates corresponding knowledge extraction data, such as "Tea ceremony has a very high status in Kyoto's traditional culture. It is an art form for self-cultivation. The process of tea ceremony includes the preparation of tea utensils, the making and tasting of matcha. In Kyoto, tea ceremony is closely connected with the lives of local residents. Many families still retain the traditional tea ceremony ritual. It is not only a beverage culture, but also a way to pass on family culture and social etiquette." In this way, for each self-selected knowledge content, the advanced large language model generates detailed knowledge extraction data based on the prompt engineering, which will be used to fine-tune the basic large language model.

[0128] Before fine-tuning, the server first prepares the relevant parameters and settings required by the LoRA (Low-Rank Adaptation) method. LoRA is an effective model fine-tuning method that adjusts the model parameters by adding a low-rank matrix to the original model, so that it can quickly adapt to specific tasks while maintaining the basic capabilities of the model.

[0129] The server determines the range of parameters to be adjusted. For example, some layers or a specific subset of parameters in the basic large language model (Llama3-70B) are selected for fine-tuning. At the same time, the rank parameter of LoRA is set, which determines the size of the low-rank matrix, thus affecting the effect and computational cost of fine-tuning. For example, setting the rank to 8 is a value determined based on experiments and experience that can both ensure the fine-tuning effect and control the computational cost.

[0130] The server applies the knowledge extraction data previously generated by the advanced large language model and prompt engineering to the fine-tuning process of the basic large language model.

[0131] Taking the knowledge extraction data about "Kinkaku-ji Temple" and "Tea Ceremony" generated by the previous tourism sample short video as an example, during the fine-tuning process, the basic large language model (Llama3-70B) will adjust its internal parameters according to these knowledge data.

[0132] For knowledge data related to "Kinkaku-ji Temple", the model will adjust parameters related to architecture, historical culture, tourist attractions, etc. For example, if the model's previous understanding of the entity "Kinkaku-ji Temple" was not accurate or in-depth enough, through fine-tuning, it will better understand the importance of Kinkaku-ji Temple in Kyoto's tourism and historical culture, as well as its relationship with other related entities (such as tourists, garden landscapes, etc.).

[0133] For the knowledge data of "tea ceremony", the model will adjust parameters related to traditional culture, art form, social etiquette, etc. For example, the model will more accurately understand the significance of tea ceremony in the lives of local residents in Kyoto, as well as the relationship between each link in the tea ceremony process and cultural heritage.

[0134] During the entire fine-tuning process, due to the use of the LoRA method, the basic large language model will only make small adjustments to selected parameters. For example, the amount of parameter adjustment is less than 0.005%. This small adjustment can retain the basic capabilities of the model, such as general capabilities such as text generation and multi-tasking, while increasing the ability to understand and process short video content.

[0135] After fine-tuning LoRA based on knowledge extraction data, the basic large language model is transformed into the preset Internet celebrity large language model (KOL_LLM).

[0136] This KOL_LLM has the ability to optimize short video content when processing tasks related to short videos. For example, when analyzing a new tourism short video, KOL_LLM can more accurately identify tourist attraction entities in the video, understand the historical and cultural significance behind the attractions, and analyze the relationship between tourists and attractions. It can better utilize the knowledge learned from sample short videos to conduct more in-depth analysis and processing of different types of short videos, thereby providing more effective support for tasks such as optimization suggestions and content analysis of short videos.

[0137] In an embodiment of the present invention, the inputting of the basic information and the depth information into the internet celebrity language model to obtain optimization suggestions for the target short video may be implemented through the following examples.

[0138] The basic information and the depth information are input into the internet celebrity language model to obtain optimization suggestions including new titles, new tags, and new lines for the target short video.

[0139] In the embodiment of the present invention, exemplarily, after acquiring the basic information and in-depth information of the target short video, the server inputs the information into the KOL_LLM.

[0140] Taking the previously mentioned travel short video as an example, its basic information includes effect data (150,000 views, 8,000 likes, 2,000 comments, 3,000 collections, and the listing time is July 5, 2023), basic content (the original title is "Explore the Mysterious Bali", the original tags are "Bali", "Travel", "Tropical Style", and the audio includes the sound of waves, local folk music, and the tour guide's explanation), background information (the creator is a user named "Travel Explorer", and the top competing videos are other popular videos about Bali travel, and the creators are "Bali Travel Expert" and so on).

[0141] The depth information may include target entities (such as "Kuta Beach" and "Beach Yoga") and target text chunks (such as "Kuta Beach is one of the most charming beaches in Bali. The sand here is fine and the waves gently lapping the shore, making people feel extremely relaxed, especially in the evening, watching the beautiful sunset, as if time has stopped.") The server organizes the basic information and in-depth information in a specific data format, such as constructing a JSON object containing all the information, and then sends the object to the KOL_LLM. After receiving the data, the KOL_LLM begins to analyze and process it in order to generate optimization suggestions for the target short video.

[0142] After receiving the basic and in-depth information of the target short video, KOL_LLM first analyzes how to optimize the title.

[0143] For the original title "Explore the mysterious Bali", KOL_LLM will optimize it according to the target entity and text segment in the depth information. For example, Kuta Beach in the depth information is an important attraction in Bali, and the description of Kuta Beach in the target text segment is very attractive. KOL_LLM may generate a new title: "Bali Kuta Beach: An Intoxicating Tropical World". This new title more specifically highlights the characteristic attraction of Kuta Beach in Bali, and uses an attractive expression such as "an intoxicating tropical world", which can better attract the audience's attention.

[0144] When KOL_LLM generates a new title, it also considers the performance data in the basic information. If the data such as the number of views and the number of likes show that the original title may not be attractive enough, KOL_LLM will focus on adding more attractive elements to the new title. For example, if the analysis finds that the original title has a low click-through rate, it may be because the expression "Explore the mysterious" is vague, and the new title directly points out the specific location of "Kuta Beach", which can make the audience know the content focus of the video more clearly, thereby increasing the click-through rate.

[0145] Next, KOL_LLM began to consider the generation of new tags.

[0146] The original tags include "Bali", "Travel", and "Tropical Style". KOL_LLM supplements and optimizes them based on the content in the deep information. Since the deep information mentions Kuta Beach and beach yoga, KOL_LLM may add the two new tags "Kuta Beach" and "Beach Yoga". These two tags describe the content elements in the video more specifically, which helps to more accurately locate the target audience in the search and recommendation system.

[0147] At the same time, KOL_LLM will also analyze the background information in the basic information. If it is found that the creator's style or the target audience's preferences are related to certain tags, corresponding adjustments will be made. For example, if the creator often publishes beach-related travel videos and the subscribers are more interested in beach activities, KOL_LLM may further optimize the tag and adjust "Beach Yoga" to "Bali Beach Yoga" to make it more targeted.

[0148] In addition, KOL_LLM will also evaluate the effectiveness of tags based on performance data. If the search volume corresponding to a tag is low, but the video content is related to it, KOL_LLM may modify or replace the tag. For example, if the tag "tropical style" can summarize the overall atmosphere of Bali, but is not accurate enough in the search, KOL_LLM may replace it with "Bali beach style", so that the video is more likely to be recommended when searching for content related to Bali beaches.

[0149] When generating new lines, KOL_LLM will consider multiple aspects of basic information and in-depth information.

[0150] First, KOL_LLM will analyze the original lines (obtained through transcription, such as "Now we have come to Kuta Beach in Bali, the sand here is fine and soft, and the waves are very suitable for surfing enthusiasts", etc.). If the original lines are not vivid enough or the information is incomplete, KOL_LLM will optimize them.

[0151] According to the target text segmentation in the depth information (such as "Kuta Beach is one of the most charming beaches in Bali. The sand here is fine and soft, and the waves gently hit the shore, making people feel extremely relaxed, especially in the evening, watching the beautiful sunset, as if time has stopped."), KOL_LLM may suggest adding a description of the sunset scenery of Kuta Beach in the evening to the original line. The optimized line may be: "Now we have come to Kuta Beach in Bali. The sand here is fine and soft, and the waves are very suitable for surfing enthusiasts. In the evening, Kuta Beach shows its charming side. The beautiful sunset dyes the entire beach orange-red, and the waves gently hit the shore, making people feel extremely relaxed, as if time has stopped." KOL_LLM also considers the audio content in the basic information. If there are some unique elements in the audio, such as local folk music, KOL_LLM may suggest adding a description of the fusion of this music and the beach atmosphere in the lines, for example: "Now we have come to Kuta Beach in Bali, accompanied by the melodious local folk music, the beach here is fine and soft, and the waves are very suitable for surfing enthusiasts. In the evening, Kuta Beach shows its charming side. The beautiful sunset dyes the entire beach orange-red, and the waves gently hit the shore, making people feel extremely relaxed, as if time has stopped." In addition, KOL_LLM will use the performance data to determine which parts of the original script may need to be optimized. If the analysis finds that the audience is more interested in the part about local culture in the video (for example, through the distribution of comments and likes in this part), but the original script contains less description in this regard, KOL_LLM may suggest adding descriptions of the connection between Bali's local culture (such as local religious ceremonies, traditional customs, etc.) and Kuta Beach to make the script richer and more attractive.

[0152] By generating the above optimization suggestions for new titles, new tags, and new lines, KOL_LLM provides comprehensive optimization directions for the target short videos. These optimization suggestions help improve the attractiveness, searchability, and viewing experience of short videos, thereby potentially improving various performance data of short videos, such as playback volume, likes, comments, and favorites.

[0153] In order to more clearly describe the solution provided by the embodiment of the present invention, a more detailed implementation is provided below.

[0154] Please refer to Figure 2 , Figure 2 Schematic diagram of the system framework for short video optimization based on knowledge graph provided in an embodiment of the present invention.

[0155] 1) Log in to the platform and open the chat window of the robot expert; 2) Enter the online link or design of the short video. As shown in the picture below, the input box contains a short video link on the TikTok platform; 3) After sending, the robot expert starts running; 4) Visit KOL_BigData to obtain basic information about the video: Get effect data: listing time, playback volume, likes, comments, and favorites; Get content: title, tag, audio; Get background information: creators, top competitor videos, and creators; If the video data does not exist in the storage server, call the real-time capture service to obtain and store it; 5) Visit the Graph RAG service to get more information about the video: Get the lines: Use the transcription function of the KOL_LLM tool set to convert the audio into text; Get the theme / scene: Analyze the title, tags, and lines through KOL_LLM to get the theme of the video content; Get the depth content: Through KOL_LLM, a batch of related entities and relationships are identified in KOL_Graph based on the information, lines, topics / scenes obtained in 4); Use them to repeatedly extract more related entities, relationships, and cluster reports from the knowledge graph, and also extract the original text chunks of the video content based on relevance; Filter and sort the search results by relevance, taking into account the context window size of KOL_LLM; Return the final search results to KOL_LLM, such as tags, topics, items, hot events related to the creator, people, other videos and their playback volume, likes and other effect data, hot events related people, places, etc. This kind of in-depth, detailed and rich relationship information; Output optimization suggestions and new content design: 6) Return the information of 4) and 5) to KOL_LLM; The big model understands, extracts, and summarizes optimization suggestions for the video from the information, as well as the design of complete new titles, tags, and lines; 7) Output optimization suggestions and new solutions to the chat window.

[0156] Core service introduction: KOL BigData 1) Video library: Data storage of hundreds of millions of videos, which can return relevant information of the videos in real time; 2) Crawling service: video crawler, which can crawl required video-related information in real time; KOL Large Language Model (KOL_LLM) Fine-tune the version based on the Llama3-70B 1M context window, use a model with a larger number of parameters, and use the invented Prompt framework to extract the data required for fine-tuning.

[0157] 1) Reasons for choosing Llama3-70B: It is an open source model with 70 billion parameters, and it is at the forefront in terms of text generation, multi-task support, reasoning, and general performance, and is easy to deploy; 2) Reasons for selecting 1M context window: Establishment of Internet celebrity knowledge graph: Entity and relationship extraction requires the input of as much video content as possible to ensure accurate and efficient extraction and merging of entities and relationships, avoid incomplete local extraction and inefficiency of multiple rounds of entity merging; Knowledge graph retrieval: This stage will return a large amount of knowledge data such as clusters, entities, and relationships. Sufficient data must be available in one analysis to make the analysis and optimization results more comprehensive and accurate. 3) Fine-tuning process: Objective: Improve the efficiency, completeness, and accuracy of extracting knowledge (entities, relations, covariates - attributes of entities) from short video content Fine-tuning data production - using the Gpt-4o model with larger parameters: Production method: Using prompt engineering, a knowledge extraction framework for short video content: Prompt framework idea: fully tap the self-discovery mechanism of the large model, adopt self-analysis, self-verification, and self-selection extraction strategies based on the short video content to complete accurate and comprehensive knowledge extraction; Prompt framework implementation, please refer to Figure 3 : Generate data: Select 50,000 video contents of various types and in the past three years, about 100MB of data, and generate corresponding knowledge; Data effect: Compared with the extraction effect of non-self-discovery prompt, the total amount of knowledge is 20% more, and it basically covers the knowledge extracted by it, which is equivalent to stimulating 20% ​​of the potential of the large model; Fine-tuning: Fine-tune Llama3-70B using LoRA; Fine-tune the model using 50,000 video contents and the corresponding generated dataset; Effect: The parameter change is less than 0.005%, which retains the basic capabilities of the model while increasing the ability to understand and process short video content. The fine-tuned model extracts video knowledge faster than before fine-tuning and Gpt-4o, and the effect is significantly better than before fine-tuning and on par with Gpt-4o; Cost: Since it is an open source, self-built model, the cost of use is much lower than Gpt-4o and meets the online service standards; Internet celebrity knowledge graph (KOL_Graph) 1) Knowledge graph establishment - based on tens of millions of short videos Text chunking - source data to build foundation: Get the information of each video from KOL_BigData as a basic information unit; Pass in KOL_LLM, and the model will automatically call the tool set to obtain the lines of the video; The model organizes each video information into two categories: Video and Channel, which are the video information and the account information; According to the specified structured format, it is converted into text blocks as the bottom-level data, and each block contains the information of a video; Multi-round extraction of entities-building basic knowledge structure: The first round of basic extraction: pass as many text blocks as possible into KOL_LLM, close to half of the context window, and mine entities, relationships, and covariates (entity attributes). For example, extract entities such as video, author, tag, game, book, etc., their relationships, and entity attributes from video information. Pass in and extract entities from all text blocks multiple times. Second round of re-extraction: This will give the model a clear signal to re-extract, and it will dig deeper; The third round of merging: pass the existing entity names to KOL_LLM, perform deduplication and merging, and generate all the final entity, relationship, and covariate combinations. Please refer to Figure 4 ; Multi-layer cluster extraction - building knowledge of different dimensions: There are 4 layers in total. Each layer is clustered according to the relevance of the lower layer or entity, which can improve the usability, density, and regularity of information, and greatly help the quality of search results and search efficiency. The data of each cluster includes cluster report (title, summary, score, score description, key insights) and elements. Among them, the report is to better express the core information of the cluster, making the search direction and results more accurate, and the elements provide the next search direction and results; Level3-Big Category Clusters-Top level: Clusters composed of related sub-category clusters, which facilitates locating search directions through larger categories; Level2-Sub-category clusters: clusters composed of related attitude clusters, which facilitate finding more detailed directions through small categories; Level 1-attitude cluster: The cluster closest to the user scenario will aggregate the opinions and views on the entity, making the detailed information of these opinions more comprehensive and accurate, and facilitating the direct location of the content that meets the search requirements; Level 0- Theme Cluster- Bottom Layer: Clusters composed of entities with a certain degree of relevance. For example, the game Black God x cluster is composed of entities and corresponding relationships such as Game-Black God x, R&D Team-Game Sience, Book-Journey to the West, and Video-Best Game Recommendations. This is convenient for improving knowledge density and making it easier to find related entities through common nouns and themes. Please refer to Figure 5 ; 2) Knowledge retrieval Query understanding and optimization: Optimize search queries through KOL_LLM to highlight query intent, topics, and entities involved. For example, if a user gives a link to a video review of game xxx, the big model will focus on extracting: game information, review methods, dimensions, and results, information about influencers and their attitudes, etc., and then obtain more information from the knowledge graph; Cluster positioning: Category cluster matching: Use the extracted key points to perform semantic matching on category clusters. For example, a game review video is matched to three category clusters: game, review, and game influencer. Sub-category cluster matching: Use key points to continue semantic matching in the sub-category clusters included in the hit major cluster, for example, matching to sub-category clusters such as "domestic ARPG games", "domestic gaming influencers", and "game reviews"; Attitude cluster matching: This layer of semantic matching has a high probability of directly hitting a cluster that is very relevant to the input video, which can maximize the retrieval efficiency. For example, matching attitude clusters such as "the game Dark God X is very difficult" is highly relevant to the content of the analyzed video. Topic cluster matching: semantically matching appropriate topic clusters from attitude clusters Each layer of matching is based on the extracted key points and cluster reports. The matching results are sorted in descending order of relevance and the configured number of results is returned. The lower the layer, the more results are returned. Entity and text block positioning: semantically match suitable entities and relations from the selected topic clusters, return a large amount of information, most of the relevant near and far entities and relations in the input video will be found, and the source text blocks can also be located; Results returned: A certain number of located entities and text blocks are also sorted in descending order based on relevance. For example, for the input game review video, game review cheats, the creator's subscriber preference information, the recent game reputation, etc. can all be mined through the relationship attributes of the knowledge graph to comprehensively help analyze the quality of the video and optimize it.

[0158] Please refer to Figure 6 , Figure 6 A short video optimization device 110 based on a knowledge graph provided in an embodiment of the present invention includes: The acquisition module 1101 is used to acquire a target short video link, where the target short video link points to a target short video; call a preset Internet celebrity big data service to acquire basic information of the target short video; based on the basic information and a preset Internet celebrity knowledge graph, call a preset Internet celebrity big language model to extract information from the target short video to obtain depth information; The optimization module 1102 is used to input the basic information and the depth information into the Internet celebrity language model to obtain optimization suggestions for the target short video.

[0159] It should be noted that the implementation principle of the aforementioned short video optimization device 110 based on knowledge graph can refer to the implementation principle of the aforementioned short video optimization method based on knowledge graph, which will not be repeated here. It should be understood that the division of the various modules of the above device is only a division of logical functions. In actual implementation, they can be fully or partially integrated into a physical entity, or they can be physically separated. And these modules can all be implemented in the form of software called by processing elements; they can also be all implemented in the form of hardware; some modules can also be implemented in the form of software called by processing elements, and some modules can be implemented in the form of hardware. For example, the short video optimization device 110 based on knowledge graph can be a separately established processing element, or it can be integrated in a chip of the above device. In addition, it can also be stored in the memory of the above device in the form of program code, and called and executed by a processing element of the above device. The function of the short video optimization device 110 based on knowledge graph. The implementation of other modules is similar. In addition, all or part of these modules can be integrated together, or they can be implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each module above can be completed by an integrated logic circuit of hardware in a processor element or an instruction in software form.

[0160] For example, the above modules may be one or more integrated circuits configured to implement the above methods, such as one or more application specific integrated circuits (ASIC), or one or more microprocessors (digital signal processors, DSP), or one or more field programmable gate arrays (FPGA), etc. For another example, when a module above is implemented in the form of a processing element scheduling program code, the processing element may be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules may be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0161] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned short video optimization device 110 based on knowledge graph. Figure 7 As shown, Figure 7 The computer device 100 is a block diagram of a structure of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a short video optimization device 110 based on a knowledge graph, a memory 111, a processor 112 and a communication unit 113.

[0162] In order to realize data transmission or interaction, the memory 111, the processor 112 and the communication unit 113 are electrically connected to each other directly or indirectly. For example, the electrical connection between these elements can be achieved through one or more communication buses or signal lines. The short video optimization device 110 based on the knowledge graph includes at least one software function module that can be stored in the memory 111 in the form of software or firmware or solidified in the operating system (OS) of the computer device 100. The processor 112 is used to execute the short video optimization device 110 based on the knowledge graph stored in the memory 111, such as the software function modules and computer programs included in the short video optimization device 110 based on the knowledge graph.

[0163] An embodiment of the present invention provides a readable storage medium, which includes a computer program. When the computer program is running, it controls the computer device where the readable storage medium is located to execute the aforementioned short video optimization device 110 based on the knowledge graph.

[0164] For illustrative purposes, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed. Numerous modifications and variations are possible in accordance with the above teachings. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to suit the intended specific application.

Claims

1. A short video optimization method based on knowledge graph, characterized in that: include: Obtain a target short video link, where the target short video link points to the target short video; Calling a preset Internet celebrity big data service to obtain basic information of the target short video; Based on the basic information and the preset Internet celebrity knowledge graph, calling the preset Internet celebrity large language model to extract information from the target short video to obtain depth information; The basic information and the depth information are input into the internet celebrity language model to obtain optimization suggestions for the target short video.

2. The method according to claim 1, characterized in that: The calling of the preset Internet celebrity big data service to obtain the basic information of the target short video includes: Calling a preset Internet celebrity big data service to obtain the effect data, basic content and background information of the target short video; the effect data includes the release time, playback volume, number of likes, number of comments and number of favorites of the target short video; the basic content includes the title, tags and audio of the target short video; the background information includes the creator of the target short video, the top competing videos and the creator of the top competing videos; Based on the effect data, the basic content and the background information as the basic information of the target short video.

3. The method according to claim 2, characterized in that Based on the basic information, calling a preset internet celebrity language model to extract information from the target short video to obtain depth information includes: The target short video is transcribed by calling a preset tool set through the Internet celebrity language model, and the audio is transcribed into lines; Analyze the title, the tag, and the lines through the Internet celebrity language model to obtain the theme of the target short video; The deep information is obtained from the preset internet celebrity knowledge graph through the internet celebrity language model based on the effect data, the basic content, the background information, the lines and the theme.

4. The method according to claim 3, characterized in that The preset Internet celebrity knowledge graph is constructed in the following ways, including: Obtaining video information and account information of a plurality of stored videos stored by the preset internet celebrity big data service through the internet celebrity big language model; Structuring the video information and the account information of each stored video to obtain a text block corresponding to each stored video, wherein the text block includes the video information and the account information; Inputting the multiple text blocks into the Internet celebrity language model for multiple rounds of extraction to obtain entities, relations, and covariates corresponding to each of the multiple text blocks; Multi-layer cluster extraction is performed based on the entities, the relationships, and the covariates to obtain the preset Internet celebrity knowledge graph, which includes a four-layer graph structure constructed from top to bottom by major category clusters, subcategory clusters, attitude clusters, and theme clusters.

5. The method according to claim 4, characterized in that The obtaining the depth information from the preset internet celebrity knowledge graph based on the effect data, the basic content, the background information, the lines, and the theme through the internet celebrity language model includes: Based on the effect data, the basic content, the background information, the lines and the theme, the Internet celebrity language model performs large category cluster matching, subcategory cluster matching, attitude cluster matching and topic cluster matching in the preset Internet celebrity knowledge graph, and obtains the target entity and target text block according to the preset relevance threshold; The target entity and the target text are divided into blocks as the depth information.

6. The method according to claim 1, characterized in that The preset Internet celebrity language model is obtained by the following methods, including: Acquire a basic large language model and an advanced large language model, wherein the model parameter amount of the advanced large language model is greater than that of the basic large language model; The advanced large language model performs analysis, self-verification, and self-selection extraction strategies on sample short videos based on a self-discovery mechanism, and generates knowledge extraction data through prompt engineering; Based on the knowledge extraction data, the basic large language model is fine-tuned in a LoRA manner to obtain the preset Internet celebrity large language model.

7. The method according to claim 1, characterized in that The step of inputting the basic information and the depth information into the internet celebrity language model to obtain optimization suggestions for the target short video includes: The basic information and the depth information are input into the internet celebrity language model to obtain optimization suggestions including new titles, new tags, and new lines for the target short video.

8. A short video optimization device based on knowledge graph, characterized in that: include: An acquisition module, used to acquire a target short video link, where the target short video link points to the target short video; Calling a preset Internet celebrity big data service to obtain basic information of the target short video; Based on the basic information and the preset Internet celebrity knowledge graph, calling the preset Internet celebrity large language model to extract information from the target short video to obtain depth information; The optimization module is used to input the basic information and the depth information into the Internet celebrity language model to obtain optimization suggestions for the target short video.

9. A computer device, characterized in that: The computer device comprises a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device executes the method according to any one of claims 1 to 7.

10. A readable storage medium, characterized in that: The readable storage medium includes a computer program, and when the computer program is executed, the computer device where the readable storage medium is located is controlled to execute the method according to any one of claims 1 to 7.

Images