Search

CN-121998706-A - Game and short play advertisement title generation and optimization method and system based on AI

CN121998706ACN 121998706 ACN121998706 ACN 121998706ACN-121998706-A

Abstract

The invention discloses a game and short play advertisement title generation and optimization method and system based on AI, wherein the method specifically comprises the steps of collecting multi-source heterogeneous data and processing by using a user interest migration model to generate a time sequence user interest vector; the method comprises the steps of constructing a knowledge graph based on product data, fusing time-ordered user interest vectors to perform node weighting to generate depth product feature vectors, inputting the depth product feature vectors into a hybrid generation framework comprising a transducer model and a generation countermeasure network to generate candidate titles, evaluating and iteratively optimizing the candidate titles through a multi-target rewarding model intelligent agent to obtain target titles, and executing self-adaptive intelligent delivery on the target titles. The invention realizes automatic generation, accurate customization and continuous optimization of the advertisement titles, and improves the creation efficiency, accuracy and individuation degree of the advertisement titles, thereby improving the click rate and conversion rate of advertisements in the gaming and short play industries.

Inventors

  • WANG CHUANPENG
  • Zeng Baijian

Assignees

  • 安徽三七极域网络科技有限公司

Dates

Publication Date
20260508
Application Date
20251224

Claims (10)

  1. 1. An AI-based game and short play advertisement title generation and optimization method, which is characterized by comprising the following steps: collecting multi-source heterogeneous data and processing the multi-source heterogeneous data by using a user interest migration model to generate a time-sequence user interest vector; constructing a knowledge graph based on the product data, fusing time-ordered user interest vectors, weighting nodes, and generating deep product feature vectors; inputting the depth product feature vector into a hybrid generation architecture comprising a transducer model and a generation countermeasure network, generating candidate titles, and introducing a constraint generation mechanism in the generation process to ensure title compliance; Evaluating and iteratively optimizing candidate titles through a multi-target rewarding model intelligent agent to obtain target titles; and executing self-adaptive intelligent delivery on the target title according to the channel characteristic knowledge base, the output result of the delivery opportunity prediction model and the real-time user image.
  2. 2. The method according to claim 1, wherein the acquiring multi-source heterogeneous data and processing using a user interest migration model generates a time-sequenced user interest vector, specifically comprising: Collecting multi-source heterogeneous data containing text, vision and user behavior sequences from a plurality of internet platforms, cleaning the multi-source heterogeneous data and fusing the multi-mode data with the user as a center to form a user behavior time sequence; inputting the user behavior time sequence into a user interest migration model for training, and carrying out weighted coding on key behaviors in the user behavior time sequence through a time sequence attention mechanism to extract preliminary user interest characteristics; In the model training process, the user interest migration model is stripped from the preliminary user interest characteristics through an adaptive training strategy in the resistance domain to obtain generalized user interest representations irrelevant to a specific data source platform, and a time-sequence user interest vector is output.
  3. 3. The method of claim 1, wherein the creating a knowledge graph based on the product data and fusing the time-ordered user interest vectors for node weighting, and generating the deep product feature vector, specifically comprises: based on multi-mode description data of a target product, automatically identifying and extracting product entities and relations among the product entities by using a natural language processing technology, and constructing a fine-grained product knowledge graph; Mapping entity nodes in the fine-grained product knowledge graph to a semantic vector space to obtain a static vector representation of each entity node; Calculating semantic relativity between time-sequence user interest vectors and static vector representations of all entity nodes, and taking the semantic relativity as dynamic importance weights of all entity nodes corresponding to target users; and injecting the dynamic importance weight into a graph neural network, carrying out weighted information propagation and global aggregation on the fine-grained product knowledge graph through the graph neural network, and outputting a deep product feature vector containing the objective characteristics of the product and the subjective interests of the user.
  4. 4. The method according to claim 1, wherein the inputting the depth product feature vector into the hybrid generation architecture comprising a transducer model and generating an antagonism network, generating candidate titles, and introducing a constraint generation mechanism in the generation process to ensure title compliance, specifically comprises: inputting the depth product feature vector into a transducer model to generate an initial title sequence for representing product core information; inputting the initial title sequence into a generator for generating an countermeasure network for stylized rendering and creative enhancement to obtain an optimized title sequence; And calling a constraint generation module, and guiding and filtering the vocabulary selection and the sequence structure in the generation process in real time according to a preset compliance rule.
  5. 5. The method of claim 4, wherein the inputting the depth product feature vector into the transducer model generates an initial header sequence for characterizing product core information, specifically comprising: Mapping the depth product feature vector into a conditional embedding vector with the same dimension as the word vector through a projection layer of a transducer model; At each time step, converting the generated partial sequence of the title words into word vectors, combining the word vectors with conditional embedding vectors, and inputting the word vectors to an encoder of a transducer model together; Performing context coding on the combined word vector and the conditional embedded vector through an encoder, and outputting probability distribution of the next candidate word on the whole word list; obtaining a generated word of the current time step according to the probability distribution sampling, and adding the generated word to a part of the sequence of the topic words to update input; and iteratively generating the generated words of each time step until the sequence termination condition is met, and outputting an initial title sequence.
  6. 6. The method according to claim 4, wherein the inputting the initial title sequence into a generator for generating an countermeasure network for stylized rendering and creative enhancement results in an optimized title sequence, specifically comprising: inputting the initial title sequence into a generating countermeasure network, encoding the initial title sequence through a generator for sequence-to-sequence style rendering, and fusing depth product feature vectors to preserve product core information; Extracting a predefined style vector from a preset high-performance advertisement topic library, and performing stylized decoding rewriting on the encoded initial topic sequence through the predefined style vector to obtain the stylized initial topic sequence; Inputting the stylized initial title sequence and the real high-performance advertisement titles in the high-performance advertisement title library into a multitask discriminator together, and generating corresponding discriminating signals by executing a plurality of tasks including true and false discrimination and attraction assessment; Updating parameters of the generator by using the discrimination signal through an antagonistic training mechanism, and outputting an optimized title sequence based on the updated generator.
  7. 7. The method of claim 5, wherein the calling constraint generating module guides and filters the vocabulary selection and the sequence structure in the generating process in real time according to a preset compliance rule, and specifically comprises: constructing a multi-level compliance rule base comprising a vocabulary level negative list, a pattern level regular rule and a semantic level compliance classifier; In each time step of word-by-word generation of the transducer model, calling a constraint generation module, dynamically shielding candidate word probability distribution output by the transducer model according to a vocabulary level negative list, and correcting a violation pattern existing in a title word sequence output by the transducer model according to a pattern level regular rule; Inputting the currently generated title word sequence into a semantic level compliance classifier, and obtaining a scoring signal for representing sequence compliance risk; and taking the scoring signal as a reward, and updating parameters of the transducer model through a strategy gradient reinforcement learning algorithm to guide the transducer model to generate an initial title sequence with low compliance risk.
  8. 8. The method according to claim 1, wherein the evaluating and iteratively optimizing candidate titles by the multi-objective rewards model agent to obtain objective titles comprises: utilizing historical advertisement putting data to construct and train a multi-target rewarding model intelligent body according to a neural network comprising a shared coding layer and a multi-task output head; inputting candidate titles to be evaluated into a trained multi-target rewarding model intelligent agent, obtaining multi-dimensional rewarding vectors corresponding to each candidate title, and calculating to obtain a comprehensive rewarding value of each candidate title according to preset business weights; Based on the comprehensive rewarding value, a reinforcement learning algorithm is adopted to calculate and update model parameters of a hybrid generation architecture comprising a transducer model and a generated countermeasure network, so that the updated hybrid generation architecture outputs a target title capable of obtaining a higher comprehensive rewarding value.
  9. 9. The method according to claim 1, wherein the performing adaptive intelligent delivery on the target title according to the channel characteristic knowledge base, the output result of the delivery opportunity prediction model, and the real-time user image specifically comprises: inquiring a pre-constructed channel characteristic knowledge base according to the target advertisement channel identification to obtain channel characteristics including format specifications, group user portraits and historical expression characteristics of the target advertisement channel; inputting real-time user image and historical behavior time sequence data of a target user into a preset release opportunity prediction model to obtain an opportunity score for representing the release suitability at the current moment; The target title, channel characteristics, opportunity scores and real-time user images to be put in are input to a dynamic decision engine together, and expected utility values of the target title aiming at the target advertisement channel, the target user and the current opportunity are calculated comprehensively through a multi-target decision function; and according to the expected utility value, the target title is adaptively put into a target user through a target advertisement channel.
  10. 10. An AI-based game and short-play advertisement title generation and optimization system, the system comprising: The data acquisition module is used for acquiring multi-source heterogeneous data and processing the multi-source heterogeneous data by using the user interest migration model to generate a time-sequence user interest vector; The vector weighting module is used for constructing a knowledge graph based on the product data, fusing time-ordered user interest vectors to carry out node weighting, and generating a depth product feature vector; The title generation module is used for inputting the depth product feature vector into a hybrid generation architecture comprising a transducer model and a generation countermeasure network, generating candidate titles, and introducing a constraint generation mechanism in the generation process to ensure title compliance; the title optimization module is used for evaluating and performing iterative optimization on the candidate titles through the multi-target rewarding model intelligent agent to obtain target titles; And the title throwing module is used for executing self-adaptive intelligent throwing on the target title according to the channel characteristic knowledge base, the output result of the throwing opportunity prediction model and the real-time user image.

Description

Game and short play advertisement title generation and optimization method and system based on AI Technical Field The invention relates to the technical field of artificial intelligence, in particular to a game and short-play advertisement title generation and optimization method and system based on AI. Background In the era of vigorous development of digital marketing at present, the game and short play industries are taken as important components of entertainment industry, and the accuracy and effectiveness of advertisement delivery play a key role in promotion of products and promotion of market share. The advertisement title is used as a primary element for attracting the attention of the user and exciting the interest of the user, and the creation quality of the advertisement directly influences the click rate and conversion rate of the advertisement. However, current games and short-play titles rely mainly on manual work, which presents a number of significant problems. First, manual writing of titles is extremely inefficient. Under the background of rapid iteration of advertisement putting requirements, the manual creation needs to consume a great deal of time and effort to design, write and modify, a great deal of titles meeting the requirements are difficult to generate in a short time, the timeliness requirement of the market on advertisement putting cannot be met in time, and the best popularization opportunity can be missed. Second, manual authoring lacks solid data support. The creation of titles often relies on the personal experience and subjective feeling of the creator, lacking in-depth analysis of large amounts of user data, product data, and market trend data. The creation mode lacking scientific basis makes the generated title difficult to accurately fit the interests and demands of target users, and the core selling points of the product cannot be effectively highlighted, so that the attractiveness and persuasion of advertisements are insufficient. Furthermore, the hot spots in the gaming and theatrical industries change rapidly and new gameplay, theatrical materials and popular cultural elements continue to emerge. The latest popular trends are difficult to capture in time by manual creation and are skillfully integrated into the title creation, so that the titles are easy to appear old and outdated, and a user group pursuing freshness cannot be attracted, thereby reducing the propagation effect of advertisements. Furthermore, manual creation has significant shortcomings in terms of degree of personalization. Different user groups have different interest preferences, consumption habits and aesthetic criteria, and different delivery channels (such as social media platforms, video platforms, search engines and the like) also have unique format specifications and user group characteristics, so that the liveness and focus of users in different time periods are different. The manual creation is difficult to accurately customize personalized titles according to various factors, so that the pertinence and the adaptability of advertisement titles are poor, and the effect of advertisement putting cannot be fully exerted. The prior art, while attempting to solve the problem of advertisement banner creation to some extent, still faces a number of limitations. On the one hand, the automation degree is low, the advertisement title generation process lacks efficient automation tools and methods, a great deal of manual intervention is still required, and large-scale and rapid title generation cannot be realized. On the other hand, the accuracy is not enough, the matching degree between the generated title and the characteristics of the target user and the product is not high, the core information of the product can not be accurately transmitted, resonance of the target user is difficult to cause, and then the click rate and the conversion rate of the advertisement are lower. In addition, the prior art lacks an effective data feedback and optimization mechanism, can not continuously optimize and improve the title according to the actual effect of advertisement delivery, is difficult to continuously improve the quality and effect of the title, and limits the overall benefit of advertisement delivery. Disclosure of Invention The invention aims to provide an AI-based game and short play advertisement title generation and optimization method and system, which realize automatic generation, accurate customization and continuous optimization of advertisement titles, and improve creation efficiency, accuracy and individuation degree of the advertisement titles, thereby improving click rate and conversion rate of advertisements in the game and short play industries, and solving at least one of the problems in the prior art. In a first aspect, the present invention provides a method for generating and optimizing AI-based game and short-play advertisement titles, the method specifically comprising: collecting mu