Creation intention recognition method, device and equipment and storage medium

Abstract

A creation intention recognition method, device and equipment and a storage medium are related to the technical field of artificial intelligence. The method comprises the following steps: obtaining input data of a creation content and prompt information, the input data being text information describing the creation content, and the prompt information being used for guiding the recognition of the input data; inputting the input data and the prompt information into a generative language model, and obtaining a creation intention recognition result of the creation content by recognizing the input data under the guidance of the prompt information by the generative language model; and generating a semantic feature portrait of the creation content based on the creation intention recognition result. The input data of the creation content is recognized by the generative language model, which can more accurately mine the hidden creation intention behind the creation content, which is conducive to generating a more refined semantic feature portrait.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device and storage medium for recognizing creative intent. Background Technology

[0002] In recommendation systems, material profiles refer to the descriptions of the features and attributes of materials (such as news, articles, videos, and products). Material profiles help recommendation systems better understand materials.

[0003] In related technologies, material profiling includes relatively refined ID (Identity Document) and statistical features, but the semantic feature profiling at the semantic level is relatively coarse, which is not conducive to improving the clarity of material profiling. Therefore, how to generate more refined semantic feature profiling is an urgent problem to be solved. Summary of the Invention

[0004] This application provides a method, apparatus, device, and storage medium for identifying creative intent. The technical solutions provided by these embodiments include the following:

[0005] According to one aspect of the embodiments of this application, a method for identifying creative intent is provided, the method comprising:

[0006] The system acquires input data and prompts for the created content. The input data is text information describing the created content, and the prompts are used to guide the recognition of the input data.

[0007] The input data and the prompt information are input into a generative language model, which then identifies the input data under the guidance of the prompt information to obtain the creative intent recognition result of the creative content.

[0008] Based on the creative intent recognition results, a semantic feature profile of the creative content is generated.

[0009] According to one aspect of the embodiments of this application, a creative intent recognition device is provided, the device comprising:

[0010] The input information acquisition module is used to acquire input data and prompt information of the created content. The input data is text information describing the created content, and the prompt information is used to guide the recognition of the input data.

[0011] The recognition result acquisition module is used to input the input data and the prompt information into the generative language model, and the generative language model recognizes the input data under the guidance of the prompt information to obtain the creative intent recognition result of the creative content;

[0012] The feature profile generation module is used to generate a semantic feature profile of the creative content based on the creative intent recognition result.

[0013] In some embodiments, the input data includes: basic information of the creator of the created content, basic information of the creator's historical created content, and basic information of the created content. The prompt information is used to indicate the creative intent of the created content by combining the basic information of the creator and the basic information of the historical created content.

[0014] The recognition result acquisition module is further configured to input the input data and the prompt information into the generative language model, and under the guidance of the prompt information, the generative language model combines the creator's basic information and the basic information of the historical creation content to identify the basic information of the creation content and obtain the creative intent recognition result of the creation content.

[0015] In some embodiments, the type of creative intent indicated by the prompt information includes at least one of the following: the degree of matching between the creative content and the title of the creative content, the theme type of the creative content, and the emotional type of the creative content;

[0016] The recognition result acquisition module is further configured to have the generative language model recognize the input data under the guidance of the prompt information, and obtain at least one of the following: the first intention recognition result of the creative content under the matching degree, the second intention recognition result of the creative content under the topic type, and the third intention recognition result of the creative content under the emotion type;

[0017] Wherein, the first intent recognition result is used to indicate whether the created content matches the theme of the created content, the second intent recognition result is used to indicate the predicted theme type of the created content, and the third intent recognition result is used to indicate the predicted emotion type of the created content.

[0018] In some embodiments, the first intent recognition result includes the predicted matching degree between the created content and the topic of the created content; the second intent recognition result includes at least one predicted topic type corresponding to the created content, and a prediction score corresponding to each of the at least one predicted topic type, wherein the prediction score of the predicted topic type is used to indicate the probability that the topic of the created content belongs to the predicted topic type; the third intent recognition result includes the predicted sentiment type of the created content, wherein the predicted sentiment type of the created content is used to indicate whether the created content has negative sentiment.

[0019] In some embodiments, the feature profile generation module is further configured to, when the first intent recognition result, the second intent recognition result, and the third intent recognition result are all obtained, respectively obtain the encoded data of the first intent recognition result, the encoded data of the second intent recognition result, and the encoded data of the third intent recognition result; and generate a semantic feature profile of the created content based on the encoded data of the first intent recognition result, the encoded data of the second intent recognition result, and the encoded data of the third intent recognition result.

[0020] In some embodiments, the feature profile generation module is further configured to obtain the encoded data corresponding to each of the at least one predicted topic type in the second intent recognition result; and to perform a weighted summation of the encoded data corresponding to each of the at least one predicted topic type using the prediction scores as weights to obtain the encoded data of the second intent recognition result.

[0021] In some embodiments, the content creation recall module is used to determine, based on the semantic feature profiles of at least two content creations and the content creations that the user account has historically interacted with, the target content to be recalled for the user account from the at least two content creations.

[0022] Alternatively, a statistical feature prediction module can be used to predict the click-through rate or negative feedback rate of the created content based on the semantic feature profile of the created content and the existing features of the created content.

[0023] In some embodiments, the input information acquisition module is further configured to acquire input data and prompt information for each of the N sample creation contents, where N is an integer greater than 1;

[0024] The recognition result acquisition module is further configured to, for any one of the at least two sample creation contents, input the input data and prompt information of the sample creation content into the generative language model, and under the guidance of the prompt information of the sample creation content, the generative language model predicts the creation intention recognition result of the sample creation content based on the input data of the sample creation content.

[0025] The content creation filtering module is used to filter out M sample content from the N sample content based on the recognition results of the creative intent of each of the N sample content, where M is a positive integer;

[0026] The language model adjustment module is used to adjust the generative language model based on the creative content of the M samples and the creative intent recognition results of each of the M samples, so as to obtain the adjusted generative language model.

[0027] According to one aspect of the embodiments of this application, a computer device is provided, the computer device including a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the above-described creative intent recognition method.

[0028] According to one aspect of the embodiments of this application, a computer-readable storage medium is provided, wherein a computer program is stored in the storage medium, the computer program being loaded and executed by a processor to implement the above-described creative intent recognition method.

[0029] According to one aspect of the embodiments of this application, a computer program product is provided, comprising a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the aforementioned creative intent recognition method.

[0030] The technical solutions provided in this application embodiment may include the following beneficial effects.

[0031] By using generative language models to identify input data for creative content, the powerful text semantic understanding capabilities and vast amounts of universal knowledge of generative language models enable more accurate uncovering of the creative intent hidden behind the content. This facilitates the generation of more refined semantic feature profiles, thereby contributing to the construction of clearer material profiles. Furthermore, targeting creative content based on these more refined semantic feature profiles improves the accuracy of content delivery. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a schematic diagram of a computer system provided in one embodiment of this application;

[0034] Figure 2 This is a flowchart of a creative intent recognition method provided in one embodiment of this application;

[0035] Figure 3This is a schematic diagram of a method for adjusting a large language model provided in one embodiment of this application;

[0036] Figure 4 This is a schematic diagram of a method for obtaining semantic feature profiles provided in one embodiment of this application;

[0037] Figure 5 This is a schematic diagram illustrating the use of a semantic feature profiling method provided in one embodiment of this application;

[0038] Figure 6 This is a block diagram of a creative intent recognition device provided in one embodiment of this application;

[0039] Figure 7 This is a block diagram of a creative intent recognition device provided in another embodiment of this application;

[0040] Figure 8 This is a schematic diagram of the structure of a computer device provided in one embodiment of this application. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0042] Please refer to Figure 1 This diagram illustrates a computer system provided in one embodiment of this application. The computer system can implement the technical solutions provided in the embodiments of this application within a specific environment. The computer system may include: a terminal device 10 and a server 20.

[0043] Terminal device 10 can be an electronic device such as a mobile phone, tablet computer, multimedia playback device, PC (Personal Computer), wearable device, in-vehicle terminal device, VR (Virtual Reality) device, AR (Augmented Reality) device, MR (Mixed Reality) device, etc. A client application for the target application can be installed and run on terminal device 10. The target application can be any application capable of providing creative intent recognition services for content creation. For example, the target application can include at least one of the following: creative intent recognition application, material profile building application, material recommendation application, material recall application, material sorting application, shopping application, social entertainment application, browser application, or information flow platform. This application embodiment does not limit this.

[0044] The embodiments of this application do not limit the implementation form of the target application described above. For example, it can be an application that needs to be downloaded and installed, a mini-program that does not need to be installed, a web application, etc.

[0045] Server 20 can be used to provide background services for the clients of the aforementioned target applications (such as author intent recognition applications). For example, server 20 can be a background server for the aforementioned target applications. Server 20 can be a single server, a server cluster consisting of multiple servers, or a cloud computing service center.

[0046] Terminal device 10 can communicate with server 20 via network 30, such as a wireless or wired network.

[0047] The technical solutions provided in this application can be applied to any scenario that requires the identification of creative intent, such as the construction of semantic feature profiles of creative content, the recommendation of creative content, the recall of creative content, the identification of creative intent in creative content, and the ranking of creative content.

[0048] For example, in the scenario of constructing a semantic feature profile of created content, after identifying the creative intent of the created content, a semantic feature profile of the created content can be constructed based on the creative intent; in the scenario of recommending created content, after constructing a semantic feature profile of the created content, the recommended objects of the created content can be determined based on the semantic feature profile of the created content; in the scenario of recalling created content, after constructing a semantic feature profile of the created content, the created content to be recalled can be determined based on the semantic feature profile of the created content.

[0049] For example, taking the construction of a semantic feature profile of creative content as an example, after the client of the target application obtains the creative content, it first constructs the input data and prompt information of the creative content. The input data is the text information describing the creative content, and the prompt information is used to guide the recognition of the input data. Then, the input data and prompt information of the creative content are input into the generative language model. The generative language model recognizes the input data under the guidance of the prompt information and obtains the creative intent recognition result of the creative content. Finally, based on the creative intent recognition result, a semantic feature profile of the creative content is generated.

[0050] Optionally, the rendering process of the semantic feature profile can also be performed by the server of the target application. For example, after the client of the target application obtains the created content, it sends the created content to the server, and the server performs the step of generating the semantic feature profile of the created content to obtain the semantic feature profile of the created content. This application embodiment does not limit this.

[0051] Please refer to Figure 2The diagram illustrates a flowchart of a creative intent recognition method provided in one embodiment of this application. The execution entity of each step of the method can be... Figure 1 The terminal device 10 or server 20 shown. This embodiment of the application uses server 20 as an example to illustrate the execution of each step of the method. The method may include the following steps (201-203).

[0052] Step 201: Obtain the input data and prompt information for the created content. The input data is text information describing the created content, and the prompt information is used to guide the recognition of the input data.

[0053] In this application embodiment, "created content" can refer to original content produced or published by an individual or team using various media formats, such as original content generated using media formats such as text, images, audio, and video. For example, the aforementioned created content can be articles, news, messages, literary works, etc., generated by an individual in text form. The aforementioned created content can also be songs, lines, scripts, product descriptions, equipment descriptions, advertisements, notices, etc., generated by a team in text form. This application embodiment does not limit this.

[0054] In one example, the aforementioned creative content can be implemented as cold material, such as content created by a creator for the first time, or creative content that has not generated interaction data. For example, the aforementioned creative content is original content produced or published for the first time by an individual or team.

[0055] Optionally, the aforementioned individuals or teams can be collectively referred to as creators, i.e., creators of content. Creators can be identified using user accounts. For example, for an individual, the user account created by that individual in the client can be used to identify that individual, such as the user account created by the first user in the client can be used to identify the first user. For a team, the user accounts created by that team in the client can be used to identify the team, such as the user accounts created by a news team in the client (e.g., xx Daily) can be used to identify the news team.

[0056] In one example, the input data for the created content is related to at least one of the following: the creator of the created content, and the created content. Exemplarily, the input data may include at least one of the following: basic information about the creator of the created content, basic information about the creator's previous created content, and basic information about the created content. Optionally, all of the above basic information is implemented as text information.

[0057] The creator's basic information describes the creator. For example, the creator's basic information includes at least one of the following: the creator's identification information, or the creator's introductory information. For instance, the creator's identification information could be the creator's user account, name, etc., and the creator's introductory information could be a brief introduction of the creator.

[0058] The creator's historical content refers to content created before the latest content. Basic information about the historical content describes it. For example, this basic information may include at least one of the following: the title, category, and negative feedback rate. Optionally, the creator's recent content may be defined as historical content, where the time difference between the recent and latest content is less than a time threshold. This application does not limit the number of pieces of basic information about historical content included in the input data; these can be set and adjusted according to actual needs, such as 3, 4, and 5. Optionally, if the creator does not have historical content, the input data may not include basic information about historical content.

[0059] Basic information about the created content is used to describe the created content. For example, the basic information about the created content includes at least one of the following: the title of the created content, or a portion of the created content. The portion of the created content can be extracted from the created content as needed. For example, the first and last sentences of a paragraph in the created content can be extracted as a portion of the content, or bold text in the created content can be extracted as a portion of the content, or text content including the theme of the created content can be used as a portion of the content; this embodiment of the application does not limit this. Optionally, the number of characters in the above-mentioned portion of the content is less than or equal to a number threshold, which can be set and adjusted according to actual usage needs, such as 400, 500, and 600.

[0060] For example, the input data for content creation includes: basic information about the creator of the content, basic information about the creator's past content creation, and basic information about the content itself. For instance, the input data for content creation can be represented as follows:

[0061] Input data:

[0062] 1. Creator Account: Name, Description

[0063] 2. Recent Content Creation: (Content 1: Title, Category, Negative Feedback Rate), ...

[0064] 3. Content of this creation: Title, and a portion of the content.

[0065] ".

[0066] The "content created this time" can refer to the aforementioned content, which may be the creator's latest creation. The creator's account is the creator's user account. This application embodiment constructs input data for the creation content based on the creator's basic information, basic information of historical creation content, and basic information of the creation content itself. This provides rich background information for the identification of the creation content, thereby improving the accuracy of identifying the creator's intent.

[0067] Content prompts, also known as hints, guide generative language models in identifying creative intent. Generative language models (GLMs) are a type of natural language processing model whose primary task is to generate natural language text. As a large-scale model, GLMs possess powerful text semantic understanding capabilities and a vast amount of universal knowledge. Applying them to creative intent identification scenarios can more accurately uncover the hidden creative intent behind the content, facilitating the generation of more refined semantic feature profiles and thus contributing to the construction of clearer material profiles.

[0068] Here, "large model" refers to a deep learning model with a large number of parameters (hundreds to trillions), which is trained on massive amounts of data to perform various complex tasks. Optionally, the above-mentioned generative language model can be implemented as at least one of the following: LLM (Large Language Model), LLaMa (Large Language Model Meta AI), Hybrid Large Model (a general large language model), BERT (Bidirectional Encoder Representations from Transformers), and UniLM (Unified Language Model). This application does not limit the architecture of the generative language model; it can be set and adjusted according to actual usage requirements.

[0069] Large language models can process natural language text, or multimodal inputs such as text-image and speech. By being trained on massive amounts of data, large language models can perform a variety of complex tasks.

[0070] In one example, the above prompt is used to instruct the creator's basic information and the basic information of the historical creations to identify the creative intent of the creations. That is, to analyze the creative intent of the creations using the creator's basic information and the basic information of the historical creations as background information. This helps to guide the generative language model to accurately identify the creative intent of the creations.

[0071] Optionally, the type of creative intent indicated by the prompt message may include at least one of the following: the degree of matching between the creative content and the title of the creative content, the theme type of the creative content, and the emotional type of the creative content.

[0072] The higher the match between the content and its title, the more obvious the creator's intention is, rather than simply aiming to increase the click-through rate (CTR). The CTR is the ratio of the number of times a piece of content is clicked to the number of times it is displayed. In this case, the content is less likely to be of low quality.

[0073] Optionally, the matching degree between the created content and its title can be divided into different levels according to actual usage needs to distinguish between high-quality and low-quality created content. Optionally, the matching degree level in the prompt message can be preset.

[0074] The more clearly defined the theme of the content, the more obvious the creator's intention. In this case, the likelihood of the content being low-quality is lower. Optionally, the theme of the content can be represented by keywords, such as "economic data update" representing the theme of "economy," and "policy interpretation" representing the theme of "policy." This application embodiment does not limit this. Optionally, the keywords in the prompt information can be pre-selected or extracted from the content.

[0075] The emotional type of creative content can reflect the creator's emotional inclination and subjective intentions. Content with negative emotions can be suppressed, while content with positive emotions can be prioritized for support, which is conducive to the dissemination of positive emotions. Optionally, the emotional type in the prompt message can be pre-selected.

[0076] For example, the types of creative intent indicated by the above-mentioned prompt information include: the degree of matching between the created content and the theme of the created content, the theme type of the created content, and the emotional type of the created content. By considering the creative intent under these three aspects—the degree of matching, theme type, and emotional type—the embodiments of this application enable the quality of the created content to be accurately reflected, which helps to improve the refinement and accuracy of the semantic feature images of the created content.

[0077] For example, the prompt message for creating content can be expressed as follows:

[0078] Prompt message:

[0079] Based on the creator's historical content (recent content), analyze the creator's latest content (this content) and output the creative intent of the latest content from the following three aspects:

[0080] 1. Title and content matching degree. Output one of three level categories: low, medium, or high;

[0081] 2. Topic Recognition. Input topic words and their predicted scores, in the format {{Topic 1: score 1, Topic 2: score 2, ...}}, sorted by predicted scores from largest to smallest. The number of topic words shall not exceed 5.

[0082] 3. Emotion Recognition. Output one of three categories: positive, neutral, or negative.

[0083] ".

[0084] The special identifier {{}} represents the input and output parts according to the rules. Specifically, {{recent creations}}, {{creator account}}, and {{current creations}} correspond to the input data of the creation content. Generative language models have a clear understanding and processing of the special identifier {{}}.

[0085] Optionally, once the prompt information for a particular creative content is determined, it can serve as a fixed template for the prompt information of all creative content, meaning that the prompt information is the same for different creative content. This facilitates the comparability of creative intentions across different creative content.

[0086] Step 202: Input the input data and prompts into the generative language model. Guided by the prompts, the generative language model identifies the input data and obtains the creative intent recognition result of the content.

[0087] After receiving the input data and prompts for the created content, the server encodes both data, resulting in encoded input data and encoded prompts. Encoding is a process of converting various types of data into computer-understandable data; for example, encoded input data is the data obtained by encoding the input data, and encoded prompts are the data obtained by encoding the prompts. The server then inputs the encoded input data and encoded prompts into a generative language model. Guided by the encoded prompts, the generative language model identifies the encoded input data, thus obtaining the creative intent recognition result for the created content.

[0088] In one example, when the input data includes basic information about the creator of the content, basic information about the creator's previous content, and basic information about the content, the prompt information is used to indicate the creative intent of the content by combining the basic information about the creator and the basic information about the previous content.

[0089] The server inputs the input data and prompts into the generative language model. Guided by the prompts, the generative language model combines the creator's basic information and the basic information of the historical creations to identify the basic information of the creation content and obtain the creative intent recognition result.

[0090] Among them, the generative language model can extract the encoded data of the creator's basic information, the encoded data of the basic information of the historical creation content, and the encoded data of the basic information of the creation content from the encoded data of the input data. Then, guided by the encoded data of the prompt information, it combines the encoded data of the creator's basic information and the encoded data of the basic information of the historical creation content to identify the encoded data of the basic information of the creation content, and obtain the creative intent identification result of the creation content.

[0091] The aforementioned creative intent recognition results are used to indicate the creative intent of the created content. These results are generated according to the rules indicated by the prompt information, such as the type of creative intent, the way the creative intent is represented, and the number of creative intents.

[0092] In one example, the generative language model identifies the input data under the guidance of prompts and obtains at least one of the following: the first intention recognition result of the creative content under the above matching degree, the second intention recognition result of the creative content under the above topic type, and the third intention recognition result of the creative content under the above emotion type.

[0093] The first intent recognition result indicates whether there is a match between the created content and its topic. Optionally, the first intent recognition result includes the predicted degree of matching between the created content and its topic. The predicted degree of matching is the degree of matching predicted by the generative language model. For example, the first intent recognition result can be one of the following: low, medium, and high. When the first intent recognition result is low, the degree of matching between the created content and its topic is low, that is, there is no match between the created content and its topic. When the first intent recognition result is high, the degree of matching between the created content and its topic is high, that is, there is a match between the created content and its topic.

[0094] The second intent recognition result is used to indicate the predicted topic type of the created content. The predicted topic type is the topic type predicted by the generative language model. Optionally, the second intent recognition result includes at least one predicted topic type corresponding to the created content, and a prediction score corresponding to each of the at least one predicted topic type.

[0095] The prediction score for a predicted topic type indicates the likelihood that the topic of the created content belongs to that predicted topic type. The higher the prediction score, the greater the likelihood that the topic of the created content belongs to that predicted topic type. Optionally, the topic of the created content may belong to one or more topic types simultaneously.

[0096] This application does not limit the number of predicted topic types included in the second intent recognition result. For example, the number of predicted topic types included in the second intent recognition result may not exceed five.

[0097] The third intent recognition result is used to indicate the predicted sentiment type of the created content. The predicted sentiment type is the sentiment type predicted by the generative language model. Optionally, the third intent recognition result includes the predicted sentiment type of the created content, which can be used to indicate whether the created content contains negative emotions.

[0098] For example, the third intent recognition result can be one of the following: positive, neutral, or negative. Here, "negative" is short for negative emotion, and "positive" is short for positive emotion. If the third intent recognition result is positive or neutral, the created content does not contain negative emotion. If the first intent recognition result is negative, the created content contains negative emotion.

[0099] For example, the creative intent recognition result of the created content includes the first intent recognition result, the second intent recognition result, and the third intent recognition result described above. For instance, the creative intent recognition result of the created content can be represented as follows:

[0100] "middle

[0101] Economic data update: 0.76, Policy interpretation: 0.55, Future economic forecast: 0.08

[0102] neutral

[0103] ".

[0104] The first intent recognition result is "neutral," the second intent recognition result includes three predicted topic types, and the third intent recognition result is "neutral." The output format of the creative intent recognition result conforms to the requirements of the prompt information.

[0105] This application's embodiments use a generative language model to accurately obtain the creative intent in three aspects: matching degree, topic type, and emotion type. This helps to filter out creative content that is clickbait, low-quality content, or negative content, thereby improving the rationality of content delivery.

[0106] Step 203: Based on the results of the creative intent recognition, generate a semantic feature profile of the creative content.

[0107] Semantic profiling refers to the semantic representation of creative content, such as a profile expressed using semantic features. Semantic features refer to the characteristics and attributes of creative content at the semantic level.

[0108] In one example, the semantic feature profile is represented using coded data of the creative intent recognition results. This coded data is the data obtained by encoding the creative intent recognition results, and can be represented using word vectors, phrase vectors, sentence vectors, etc. The coded data of the creative intent recognition results includes the semantic information of the creative intent recognition results. The embedding is a numerical vector composed of multiple floating-point numbers, which describes various attributes and properties of the creative content in a high-dimensional space. It can be obtained through a trained neural network model.

[0109] For example, the process of constructing a semantic feature profile can be as follows:

[0110] 1. When the first intent recognition result, the second intent recognition result, and the third intent recognition result are all obtained, the encoded data of the first intent recognition result, the encoded data of the second intent recognition result, and the encoded data of the third intent recognition result are obtained respectively.

[0111] For example, a Word2Vec model trained on a Chinese corpus can be used to embed the creative intent recognition results to obtain encoded data of the first intent recognition result, encoded data of the second intent recognition result, and encoded data of the third intent recognition result.

[0112] Optionally, since the first intent recognition result and the third intent recognition result belong to the preferred minority category results, they can simply be assigned an embedding.

[0113] The second intent recognition result may have at least one predicted topic type. The encoded data of the second intent recognition result can be obtained by combining the embedding of at least one predicted topic type. Therefore, embodiments of this application may further include at least one of the following:

[0114] (1) For at least one predicted topic type in the second intent recognition result, obtain the encoded data corresponding to at least one predicted topic type respectively; use the prediction scores corresponding to at least one predicted topic type respectively as weights, and sum the encoded data corresponding to at least one predicted topic type respectively to obtain the encoded data of the second intent recognition result.

[0115] For example, if the second intent recognition result includes predicted topic type 1 and predicted subject type 2, the encoded data for predicted topic type 1 and predicted subject type 2 are obtained through the Word2Vec model. The prediction score for predicted topic type 1 is used as the weight of the encoded data for predicted topic type 1, and the prediction score for predicted topic type 2 is used as the weight of the encoded data for predicted topic type 2. The weighted sum of the encoded data for predicted topic type 1 and predicted topic type 2 yields the encoded data for the second intent recognition result.

[0116] This application embodiment obtains the encoded data of the second intent recognition result by weighting and summing the encoded data of at least one predicted topic type in the second intent recognition result with the prediction score as the weight. This makes the encoded data of the second intent recognition result more accurate and comprehensive, thereby improving the accuracy of semantic feature profiling.

[0117] (2) For at least one predicted topic type in the second intent recognition result, obtain the encoded data corresponding to at least one predicted topic type respectively; input the encoded data corresponding to at least one predicted topic type respectively into the trained neural network model, and obtain the encoded data of the second intent recognition result from the trained neural network model.

[0118] For example, if the second intent recognition result includes predicted topic type 1 and predicted subject type 2, the encoded data of predicted topic type 1 and predicted subject type 2 are obtained through the Word2Vec model. The encoded data of predicted topic type 1 and predicted topic type 2 are then input into the trained neural network model, which, based on these data, obtains the encoded data of the second intent recognition result.

[0119] The neural network model can be implemented as at least one of the following: MLP (Multilayer Perceptron), CNN (Convolutional Neural Networks), RNN (Recurrent Neural Network), ResNet (Residual Networks), and DNN (Deep Neural Networks).

[0120] Optionally, the neural network model can be trained using a contrastive learning method. For example, by iteratively adjusting the parameters of the neural network and optimizing the training loss function of the neural network model, the cosine distance between the encoded data of two second intent recognition results belonging to the same creator and the cosine distance between the encoded data of two second intent recognition results not belonging to the same creator can be minimized. By combining the prior knowledge of the neural network model, the encoded data of the second intent recognition results can be accurately obtained.

[0121] 2. Based on the encoded data of the first intent recognition result, the encoded data of the second intent recognition result, and the encoded data of the third intent recognition result, a semantic feature profile of the created content is generated.

[0122] Optionally, the encoded data of the first intent recognition result, the encoded data of the second intent recognition result, and the encoded data of the third intent recognition result can be concatenated together in sequence to obtain a semantic feature profile of the created content.

[0123] Optionally, the encoded data of the first intent recognition result, the encoded data of the second intent recognition result, and the encoded data of the third intent recognition result are weighted and summed to obtain a semantic feature profile of the created content. This application does not limit this approach.

[0124] This application embodiment, by comprehensively considering the creative intent of the creative content in three aspects, can obtain a more refined and comprehensive semantic feature profile, which is conducive to building a more accurate and clear material profile for the creative content. Especially when the creative content is a cold material, it makes up for the lack of ID features and statistical features of cold materials, which is conducive to improving the accuracy of the delivery of creative content.

[0125] In one example, if the creative intent recognition result of the creative content only includes one of the first intent recognition result, the second intent recognition result, and the third intent recognition result, the encoded data of the intent recognition result included in the creative intent recognition result is directly determined as the semantic feature profile of the creative content.

[0126] In one example, if the creative intent recognition result of the creative content only includes two of the first intent recognition result, the second intent recognition result, and the third intent recognition result, the encoded data of the intent recognition results included in the creative intent recognition result can be concatenated to obtain the semantic feature profile of the creative content.

[0127] In summary, the technical solution provided in this application identifies the input data of creative content through a generative language model. Because generative language models possess powerful text semantic understanding capabilities and a vast amount of universally applicable knowledge, they can more accurately uncover the creative intent hidden behind the content. This facilitates the generation of more refined semantic feature profiles, thereby contributing to the construction of clearer material profiles. Furthermore, delivering creative content based on these more refined semantic feature profiles improves the accuracy of content delivery.

[0128] In some embodiments, the generative language model described above is a trained model. For example, since the generative language model is a type of large model, its training can be completed through fine-tuning. Therefore, embodiments of this application may further include the following.

[0129] 1. Obtain the input data and prompts for each of the N sample creation contents, where N is an integer greater than 1.

[0130] The sample content is the content created to train the generative language model. The generative language model is pre-trained on a large amount of data; fine-tuning aims to adapt the pre-trained model to a specific domain or task to improve its application performance. In other words, the fine-tuned generative language model can be widely applied to downstream tasks across various areas of Artificial Intelligence (AI). In this embodiment, the fine-tuned generative language model can be applied to the task of recognizing the creative intent of the created content.

[0131] The method for constructing the input data and prompt information for the sample creation content is the same as the method for constructing the input data and prompt information for the creation content described above. For content not described in the embodiments of this application, please refer to the above embodiments, and it will not be repeated here.

[0132] 2. For any one of the at least two sample creations, input the input data and prompts of the sample creations into the generative language model. Guided by the prompts of the sample creations, the generative language model predicts the creative intent recognition result of the sample creations based on the input data of the sample creations.

[0133] Optionally, the encoded data of the input data of the sample creation content and the encoded data of the prompt information are input into the generative language model. Guided by the encoded data of the prompt information of the sample creation content, the generative language model predicts the creative intent recognition result of the sample creation content based on the encoded data of the input data of the sample creation content.

[0134] Optionally, the creative intent recognition results of the sample creative content include the first intent recognition result, the second intent recognition result, and the third intent recognition result of the sample creative content.

[0135] 3. Based on the creative intent recognition results of each of the N sample creations, select M sample creations from the N sample creations, where M is a positive integer.

[0136] Optionally, this embodiment employs Supervised Fine Tuning (SFT) to adjust the generative language model. Supervised fine tuning relies on high-quality <input data, creative intent recognition result> data pairs. Unlike outputs that rely entirely on manual annotation, thanks to the capabilities of the generative language model, this embodiment only needs to construct the input data and prompt information to obtain a large number of creative intent recognition results as preliminary label data. Subsequently, by simply judging whether the creative intent recognition results meet expectations, thousands of <input data, creative intent recognition result> data pairs can be selected.

[0137] For example, M sample creations that meet expectations can be selected manually from N sample creations. Alternatively, M sample creations can be randomly selected from N sample creations. Furthermore, M sample creations can be selected from N sample creations according to preset rules, such as determining the sample creations that meet expectations based on a matching degree greater than a matching threshold between the input data and the creation intent recognition result. This application embodiment does not limit this approach.

[0138] 4. Based on the creative content of M samples and the recognition results of the creative intent of each of the M samples, adjust the generative language model to obtain the adjusted generative language model.

[0139] Optionally, the creative intent recognition results of each of the M sample creations are used as label data to iteratively adjust the parameters of the generative language model, resulting in the adjusted generative language model.

[0140] For example, for any sample creation content among M sample creation content, the new creation intention recognition result of the sample creation content is obtained through a generative language model; based on the difference between the new creation intention recognition result of the sample creation content and the creation intention recognition result of the sample creation content, the training loss of the generative language model is calculated; with the goal of minimizing the training loss of the generative language model, the parameters of the generative language model are iteratively adjusted to obtain the adjusted generative language model.

[0141] Optionally, the parameters of the generative language model can be directly adjusted to obtain the adjusted generative language model; alternatively, a low-rank matrix network layer can be added to the generative language model to form a modified generative language model. For example, two low-rank matrices (such as a weight matrix and a bias matrix) can be introduced into the fully connected layer of the generative language model to replace the original fully connected layer. In this case, only the parameters of the low-rank matrix network layer are adjusted, while keeping the other parameters of the generative language model fixed, to obtain the adjusted generative language model.

[0142] An adapter can be added to the generative language model to form a modified generative language model. For example, the adapter takes the output of the generative language model as input. Or, the generative language model takes the output of the adapter as input. In this case, only the parameters of the adapter are adjusted, keeping the parameters of the generative language model fixed, to obtain the modified generative language model. This application does not limit the method for adjusting the parameters of the generative language model.

[0143] An adapter is a module that can be inserted into a pre-trained neural network model (such as the generative language model mentioned above) to fine-tune the network to adapt it to new tasks or datasets. An adapter typically contains one or two layers, such as fully connected layers or convolutional layers, which perform transformations on the input data during the network's forward propagation. By training the inserted adapter, we can adapt the pre-trained neural network model to new tasks without changing the weights of the original network.

[0144] In one example, after generating a semantic feature profile of the created content, the semantic feature profile of the created content can be fed into downstream applications. In this case, the embodiments of this application may also include at least one of the following:

[0145] 1. Having obtained semantic feature profiles of at least two types of content, based on the semantic feature profiles of each type of content and the content that the user account has interacted with in the past, determine the target content to be recalled for the user account from the at least two types of content.

[0146] Optionally, for any given content, the semantic feature profile of the content is calculated, and the similarity between the semantic feature profile of the content and the content that the user's account has interacted with in the past is calculated. Then, at least two content items are sorted in descending order of similarity to obtain a content sequence. The first content item in the content sequence can be identified as the target content item.

[0147] The embodiments of this application do not limit the method of similarity calculation, such as cosine similarity, Euclidean distance, and any algorithm that can be used to calculate the distance between vectors.

[0148] 2. Based on the semantic feature profile of the created content and its existing features, predict the click-through rate or negative feedback rate of the created content.

[0149] The existing features of the created content can refer to features other than semantic features, such as the ID features and statistical features mentioned above. Optionally, the semantic feature profile and the existing features of the created content are input into the recommendation ranking model, which then predicts the click-through rate or negative feedback rate of the created content based on these two factors. The recommendation ranking model can be constructed using a neural network model.

[0150] The server then decides whether to recommend the created content to users based on the click-through rate or negative feedback rate of the content.

[0151] 3. Based on the semantic feature profile of the created content, determine the user account corresponding to the content creation from multiple user accounts.

[0152] The user account to which content is to be delivered refers to the user account that creates the content. Optionally, for any user account among multiple user accounts, the semantic feature profile of the created content is calculated, and the similarity between the semantic feature profile of the created content and the user account's historically interacted content is calculated. Then, the multiple user accounts are sorted in descending order of similarity to obtain a user account sequence. The first user account in the user account sequence can be identified as the user account to be delivered.

[0153] This application does not limit the downstream applications that can be applied to the semantic feature profile of the created content in the embodiments of this application.

[0154] In summary, the technical solution provided in this application, through fine-tuning of the generative language model, allows it to be adapted for the task of identifying the creative intent of content creation. This enables the application of the generative language model's powerful textual semantic understanding capabilities and vast universal knowledge to the task of identifying creative intent, thereby more accurately uncovering the hidden creative intent behind the content. This facilitates the generation of more refined semantic feature profiles, leading to a clearer material profile. Furthermore, targeting content based on more refined semantic feature profiles improves the accuracy of content targeting.

[0155] In some embodiments, reference Figures 3 to 5 Taking the generative language model as an LLM and the created content as cold material (such as the latest created content) as an example, the embodiments of this application may also include the following contents.

[0156] 1. Obtain the creative intent recognition results of N sample creative content through the large language model 300.

[0157] Optionally, the input data and prompts for each of the N sample creations are obtained. For any sample creation among the N sample creations, the input data and prompts for the sample creation are input into the large language model 300, and the large language model 300 obtains the creation intent recognition result of the sample creation.

[0158] The input data includes basic information about the creator, basic information about the creator's past creative content, and basic information about the new content. The types of creative intent indicated by the prompts include: the degree of match between the creative content and its title, the theme type of the creative content, and the emotional type of the creative content.

[0159] 2. Filter M pairs of data: <input data, creative intent recognition result>.

[0160] 3. Based on M pairs of <input data, creative intent recognition results>, fine-tune the parameters of the large language model 300 to obtain the supervised fine-tuned large language model 300.

[0161] The supervised fine-tuning of the large language model 300 can be used to perform the task of identifying the creative intent of content creation. Compared to the unsupervised large language model 300, the supervised fine-tuning large language model 300 can more accurately identify the creative intent. For example, for clickbait content such as "Data is out! Unexpectedly...", the first intent identification result of the unsupervised large language model 300 is "medium", while the first intent identification result of the supervised fine-tuning large language model 300 is "low".

[0162] 4. Obtain the input data and prompts for the latest created content.

[0163] Optionally, the creator's basic information is stored in a KV (Key-Value) database. When new content appears, the server can retrieve the creator's user account (Key) and obtain the corresponding name and description (Value) from the KV database. The KV database is updated daily to supplement information on new creators.

[0164] For example, the creator's basic information is represented as "xx Daily, height determines influence".

[0165] Basic information about historical content can also be stored in a key-value (KV) database, updated daily to include the latest content created by the creator. The KV database can delete basic information from earlier content based on a certain storage threshold (e.g., 10) to reduce storage pressure.

[0166] For example, the basic information of historical creative content can be represented as "xx luckily qualified, sports news, 0.0001" or "electric vehicles are prohibited from entering the community, social documentary, 0.0023".

[0167] The latest content may be quite long, and you can extract parts of it according to preset rules. For example, the basic information of the latest content can be represented as: "Financial statistics for May have been released, with different levels of monetary statistics. The market generally believes that M1 reflects current economic activity. M1 decreased by 4.2% year-on-year..."

[0168] Basic information about historical content can be stored in the server's memory. When new content is created, the basic information of the creator, the basic information of historical content, and the basic information of the latest content are combined to obtain the input data for the new content.

[0169] The prompts for the latest created content are used to guide the large language model 300 after supervision and fine-tuning. By combining the creator's basic information and the basic information of historical created content, the creative intent of the latest created content can be identified.

[0170] 5. Based on the input data and prompts of the latest created content, the creative intent recognition result of the latest created content is obtained by supervising and fine-tuning the large language model 300.

[0171] Optionally, the creative intent identification results of the latest created content include: the first intent identification result of the latest created content under the matching degree (i.e., the matching degree between the latest created content and the title of the latest created content), the second intent identification result of the latest created content under the theme type (i.e., the theme type of the latest created content), and the third intent identification result of the latest created content under the emotion type (i.e., the emotion type of the latest created content).

[0172] 6. Based on the results of identifying the creative intent of the latest created content, generate a semantic feature profile of the latest created content.

[0173] Obtain the embeddings of the first intent recognition result, the second intent recognition result, and the third intent recognition result. Concatenate the embeddings of the first intent recognition result, the second intent recognition result, and the third intent recognition result together to obtain the semantic feature profile of the latest created content.

[0174] 7. Apply the semantic feature profile of the latest created content to downstream tasks.

[0175] Optionally, a new recall queue can be added, such as an I2i2u (Item-to-Item-to-User Recall) recall queue. I2i2u recall is a recall strategy in a recommendation system that selects items similar to the user's historical preferences from the candidate item library and recommends them to the user.

[0176] For example, the latest created content from the latest content collection that has a high similarity to the user's historical interactive content can be recalled and returned to the user. Here, the similarity between the latest created content and the user's historical interactive content is the similarity between the semantic feature profiles of the latest created content and the semantic feature profiles of the user's historical interactive content.

[0177] Optionally, the semantic feature profile of the latest created content can be used as a supplementary feature, together with the existing features of the latest created content, as input to the recommendation ranking model 500 for click-through rate prediction or negative feedback prediction, etc.

[0178] The technical solution provided in this application identifies the input data of newly created content through a large language model. Because large language models possess powerful text semantic understanding capabilities and vast amounts of universal knowledge, they can more accurately uncover the creative intent hidden behind the newly created content. This facilitates the generation of more refined semantic feature profiles, thereby contributing to the construction of clearer material profiles. Furthermore, delivering newly created content based on more refined semantic feature profiles improves the accuracy of content delivery, thereby enhancing the recall or recommendation effectiveness of the newly created content.

[0179] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0180] refer to Figure 6 This diagram illustrates a block diagram of a creative intent recognition device according to an embodiment of this application. The device has the functionality to implement the method example described above, and can be the computer device described above, or it can be installed within a computer device. Figure 6 As shown, the device 600 includes: an input information acquisition module 601, a recognition result acquisition module 602, and a feature profile generation module 603.

[0181] The input information acquisition module 601 is used to acquire input data and prompt information of the created content. The input data is text information describing the created content, and the prompt information is used to guide the recognition of the input data.

[0182] The recognition result acquisition module 602 is used to input the input data and the prompt information into the generative language model, and the generative language model recognizes the input data under the guidance of the prompt information to obtain the creative intent recognition result of the creative content.

[0183] The feature profile generation module 603 is used to generate a semantic feature profile of the creative content based on the creative intent recognition result.

[0184] In some embodiments, the input data includes: basic information of the creator of the created content, basic information of the creator's historical created content, and basic information of the created content. The prompt information is used to indicate the creative intent of the created content by combining the basic information of the creator and the basic information of the historical created content.

[0185] The recognition result acquisition module 602 is used to input the input data and the prompt information into the generative language model. Under the guidance of the prompt information, the generative language model combines the creator's basic information and the basic information of the historical creation content to identify the basic information of the creation content and obtain the creative intent recognition result of the creation content.

[0186] In some embodiments, the type of creative intent indicated by the prompt information includes at least one of the following: the degree of matching between the creative content and the title of the creative content, the theme type of the creative content, and the emotional type of the creative content;

[0187] The recognition result acquisition module 602 is further configured to have the generative language model recognize the input data under the guidance of the prompt information, and obtain at least one of the following: the first intention recognition result of the creative content under the matching degree, the second intention recognition result of the creative content under the topic type, and the third intention recognition result of the creative content under the emotion type;

[0188] Wherein, the first intent recognition result is used to indicate whether the created content matches the theme of the created content, the second intent recognition result is used to indicate the predicted theme type of the created content, and the third intent recognition result is used to indicate the predicted emotion type of the created content.

[0189] In some embodiments, the first intent recognition result includes the predicted degree of matching between the created content and the theme of the created content;

[0190] The second intent recognition result includes at least one predicted topic type corresponding to the created content, and prediction scores corresponding to the at least one predicted topic type, wherein the prediction score of the predicted topic type is used to indicate the probability that the topic of the created content belongs to the predicted topic type;

[0191] The third intent recognition result includes the predicted emotion type of the created content, which is used to indicate whether the created content has negative emotions.

[0192] In some embodiments, such as Figure 7 As shown, the feature profile generation module 603 includes: an encoded data acquisition submodule 603a and a feature profile generation submodule 603b.

[0193] The encoded data acquisition submodule 603a is used to acquire the encoded data of the first intent recognition result, the encoded data of the second intent recognition result, and the encoded data of the third intent recognition result, respectively, when the first intent recognition result, the second intent recognition result, and the third intent recognition result are all acquired.

[0194] The feature profile generation submodule 603b is used to generate a semantic feature profile of the created content based on the encoded data of the first intent recognition result, the encoded data of the second intent recognition result, and the encoded data of the third intent recognition result.

[0195] In some embodiments, the encoded data acquisition submodule 603a is further configured to:

[0196] For at least one predicted topic type in the second intent recognition result, obtain the encoded data corresponding to each of the at least one predicted topic type;

[0197] Using the prediction scores corresponding to the at least one predicted topic type as weights, the encoded data corresponding to the at least one predicted topic type are weighted and summed to obtain the encoded data of the second intent recognition result.

[0198] In some embodiments, such as Figure 7 As shown, the device 600 also includes: a content creation recall module 604 and a statistical feature prediction module 605.

[0199] The content creation recall module 604 is used to determine the target content to be recalled for the user account from the at least two content creations, based on the semantic feature profiles of each of the content creations and the content creations that the user account has interacted with in the past, when at least two content creations have been obtained.

[0200] Alternatively, the statistical feature prediction module 605 is used to predict the click-through rate or negative feedback rate of the created content based on the semantic feature profile of the created content and the existing features of the created content.

[0201] In some embodiments, such as Figure 7 As shown, the device 600 also includes: a content creation filtering module 606 and a language model adjustment module 607.

[0202] The input information acquisition module 601 is also used to acquire the input data and prompt information of each of the N sample creation contents, where N is an integer greater than 1.

[0203] The recognition result acquisition module 602 is further configured to, for any one of the at least two sample creation contents, input the input data and prompt information of the sample creation content into the generative language model, and under the guidance of the prompt information of the sample creation content, the generative language model predicts the creation intent recognition result of the sample creation content based on the input data of the sample creation content.

[0204] The content filtering module 606 is used to filter out M sample content from the N sample content based on the recognition results of the creative intent of each of the N sample content, where M is a positive integer.

[0205] The language model adjustment module 607 is used to adjust the generative language model based on the creative content of the M samples and the creative intent recognition results of each of the M sample creative content, so as to obtain the adjusted generative language model.

[0206] In summary, the technical solution provided in this application identifies the input data of creative content through a generative language model. Because generative language models possess powerful text semantic understanding capabilities and a vast amount of universally applicable knowledge, they can more accurately uncover the creative intent hidden behind the content. This facilitates the generation of more refined semantic feature profiles, thereby contributing to the construction of clearer material profiles. Furthermore, delivering creative content based on these more refined semantic feature profiles improves the accuracy of content delivery.

[0207] It should be noted that the apparatus provided in the above embodiments is only illustrated by the division of the above functional modules when implementing its functions. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0208] Please refer to Figure 8 This illustration shows a schematic diagram of a computer device according to an embodiment of this application. The computer device can be any electronic device with data computing, processing, and storage functions, and can be implemented as... Figure 1 The terminal device 10 or server 20 in the implementation environment of the scheme shown. Specifically, it may include the following:

[0209] The computer device 800 includes a central processing unit (such as a CPU, GPU, or FPGA) 801, a system memory 804 including RAM (Random-Access Memory) 802 and ROM (Read-Only Memory) 803, and a system bus 805 connecting the system memory 804 and the central processing unit 801. The computer device 800 also includes a basic input / output system 806 that facilitates information transfer between various devices within the server, and a large-capacity storage device 807 for storing the operating system 813, application programs 814, and other program modules 815.

[0210] The basic input / output system 806 includes a display 808 for displaying information and an input device 809 for user input, such as a mouse or keyboard. Both the display 808 and the input device 809 are connected to the central processing unit 801 via an input / output controller 810 connected to the system bus 805. The basic input / output system 806 may also include the input / output controller 810 for receiving and processing input from multiple other devices such as a keyboard, mouse, or electronic stylus. Similarly, the input / output controller 810 also provides output to a display screen, printer, or other types of output devices.

[0211] The mass storage device 807 is connected to the central processing unit 801 via a mass storage controller (not shown) connected to the system bus 805. The mass storage device 807 and its associated computer-readable media provide non-volatile storage for the computer device 800. That is, the mass storage device 807 may include computer-readable media (not shown) such as a hard disk or a CD-ROM (Compact Disc Read-Only Memory) drive.

[0212] Without loss of generality, the computer-readable medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other solid-state storage technologies, CD-ROM, DVD (Digital Video Disc) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that the computer storage medium is not limited to the above-mentioned types. The system memory 804 and mass storage device 807 described above can be collectively referred to as memory.

[0213] According to an embodiment of this application, the computer device 800 can also be connected to a remote computer on a network, such as the Internet, for operation. That is, the computer device 800 can be connected to a network 812 via a network interface unit 811 connected to the system bus 805, or it can also use the network interface unit 811 to connect to other types of networks or remote computer systems (not shown).

[0214] The memory also includes a computer program stored in the memory and configured to be executed by one or more processors to implement the above-described creative intent recognition method.

[0215] In some embodiments, a computer-readable storage medium is also provided, wherein a computer program is stored therein, which, when executed by a processor, implements the above-described creative intent recognition method.

[0216] Optionally, the computer-readable storage medium may include: ROM (Read-Only Memory), RAM (Random-Access Memory), SSD (Solid State Drives), or optical disc, etc. The random access memory may include ReRAM (Resistance Random Access Memory) and DRAM (Dynamic Random Access Memory).

[0217] In some embodiments, a computer program product is also provided, the computer program product comprising a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program, causing the computer device to perform the above-described creative intent recognition method.

[0218] It should be noted that, in this application embodiment, before and during the collection of user-related data, a prompt interface, pop-up window, or voice prompt message can be displayed. This prompt interface, pop-up window, or voice prompt message is used to inform the user that their relevant data is being collected. This ensures that the application only begins executing the steps related to collecting user-related data after receiving confirmation from the user regarding the prompt interface or pop-up window; otherwise (i.e., without receiving confirmation from the user), the steps to collect user-related data end, meaning no user-related data is collected. In other words, all user data collected in this application is processed strictly in accordance with the requirements of relevant national laws and regulations. The informed consent or separate consent of the personal information subject is obtained only with the user's consent and authorization. Subsequent data use and processing are conducted within the scope of laws and regulations and the authorization of the personal information subject. Furthermore, the collection, use, and processing of relevant user data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the creative content, user accounts, and creative intent involved in this application are all obtained with full authorization.

[0219] It should be understood that "multiple" as used herein refers to two or more. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. Furthermore, the step numbers described herein are merely illustrative of one possible execution order between steps. In some other embodiments, the steps may not be executed in numerical order, such as two steps with different numbers being executed simultaneously, or two steps with different numbers being executed in the reverse order of the illustration. This application does not limit this approach.

[0220] The above description is merely an exemplary embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An authoring intent recognition method characterized by, The method includes: The system acquires input data and prompts for the created content. The input data is text information describing the created content, and the prompts are used to guide the recognition of the input data. The input data and the prompt information are input into a generative language model, which then identifies the input data under the guidance of the prompt information to obtain the creative intent recognition result of the creative content. Based on the creative intent recognition results, a semantic feature profile of the creative content is generated.

2. The method of claim 1, wherein, The input data includes: basic information of the creator of the creative content, basic information of the creator's historical creative content, and basic information of the creative content. The prompt information is used to indicate the creative intent of the creative content by combining the basic information of the creator and the basic information of the historical creative content. The step of inputting the input data and the prompt information into a generative language model, and having the generative language model, guided by the prompt information, identify the input data to obtain the creative intent recognition result of the created content, includes: The input data and the prompt information are input into the generative language model. Guided by the prompt information, the generative language model combines the creator's basic information and the basic information of the historical creation content to identify the basic information of the creation content and obtain the creation intention identification result of the creation content.

3. The method according to claim 1 or 2, characterized in that, The type of creative intent indicated by the prompt information includes at least one of the following: the degree of matching between the creative content and the title of the creative content, the theme type of the creative content, and the emotional type of the creative content; The step of inputting the input data and the prompt information into a generative language model, and having the generative language model, guided by the prompt information, identify the input data to obtain the creative intent recognition result of the created content, includes: The generative language model identifies the input data under the guidance of the prompt information and obtains at least one of the following: the first intention recognition result of the creative content under the matching degree, the second intention recognition result of the creative content under the topic type, and the third intention recognition result of the creative content under the emotion type; Wherein, the first intent recognition result is used to indicate whether the created content matches the theme of the created content, the second intent recognition result is used to indicate the predicted theme type of the created content, and the third intent recognition result is used to indicate the predicted emotion type of the created content.

4. The method according to claim 3, characterized in that, The first intent recognition result includes the predicted matching degree between the created content and the theme of the created content; The second intent recognition result includes at least one predicted topic type corresponding to the created content, and prediction scores corresponding to the at least one predicted topic type, wherein the prediction score of the predicted topic type is used to indicate the probability that the topic of the created content belongs to the predicted topic type; The third intent recognition result includes the predicted emotion type of the created content, which is used to indicate whether the created content has negative emotions.

5. The method according to claim 3 or 4, characterized in that, The step of generating a semantic feature profile of the creative content based on the creative intent recognition result includes: When the first intent recognition result, the second intent recognition result, and the third intent recognition result are all obtained, the encoded data of the first intent recognition result, the encoded data of the second intent recognition result, and the encoded data of the third intent recognition result are obtained respectively. Based on the encoded data of the first intent recognition result, the encoded data of the second intent recognition result, and the encoded data of the third intent recognition result, a semantic feature profile of the created content is generated.

6. The method of claim 5, wherein, The method further includes: For at least one predicted topic type in the second intent recognition result, obtain the encoded data corresponding to each of the at least one predicted topic type; Using the prediction scores corresponding to the at least one predicted topic type as weights, the encoded data corresponding to the at least one predicted topic type are weighted and summed to obtain the encoded data of the second intent recognition result.

7. The method according to any one of claims 1 to 6, characterized in that, After generating the semantic feature profile of the creative content based on the creative intent recognition result, the method further includes: Having obtained semantic feature profiles of at least two types of content, the target content to be recalled for the user account is determined from the at least two types of content based on the semantic feature profiles of each type of content and the content that the user account has interacted with in the past. or, Based on the semantic feature profile of the created content and the existing features of the created content, predict the click-through rate or negative feedback rate of the created content.

8. The method according to any one of claims 1 to 7, characterized in that, Before inputting the input data and the prompt information into the generative language model, and having the generative language model identify the input data under the guidance of the prompt information to obtain the creative intent recognition result of the creative content, the method further includes: Obtain the input data and prompts for each of the N sample creation content, where N is an integer greater than 1; For any one of the at least two sample creation contents, the input data and prompt information of the sample creation content are input into the generative language model. Guided by the prompt information of the sample creation content, the generative language model predicts the creation intent recognition result of the sample creation content based on the input data of the sample creation content. Based on the creative intent recognition results of the N sample creative content, M sample creative content are selected from the N sample creative content, where M is a positive integer; Based on the creation content of the M samples and the recognition results of the creation intent of each of the M samples, the generative language model is adjusted to obtain the adjusted generative language model.

9. An authoring intent recognition apparatus characterized by comprising: The device includes: The input information acquisition module is used to acquire input data and prompt information of the created content. The input data is text information describing the created content, and the prompt information is used to guide the recognition of the input data. The recognition result acquisition module is used to input the input data and the prompt information into the generative language model, and the generative language model recognizes the input data under the guidance of the prompt information to obtain the creative intent recognition result of the creative content; The feature profile generation module is used to generate a semantic feature profile of the creative content based on the creative intent recognition result.

10. A computer device, comprising: The computer device includes a processor and a memory, the memory storing a computer program that is loaded and executed by the processor to implement the creative intent recognition method as described in any one of claims 1 to 8.

11. A computer readable storage medium characterized by, The computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the creative intent recognition method as described in any one of claims 1 to 8.

12. A computer program product, characterised in that, The computer program product includes a computer program stored in a computer-readable storage medium, and a processor reads from and executes the computer program to implement the creative intent recognition method as described in any one of claims 1 to 8.

Images