Title recommendation method and device

Abstract

The embodiment of the invention provides a title recommendation method, and the method comprises the steps: displaying a target page which comprises a target input box; obtaining a target text through the target input box; generating a model or a title library through a pre-trained target title, and obtaining a plurality of candidate titles based on the target text; displaying the plurality of candidate titles in the target page; and in response to selection of one candidate title from the plurality of candidate titles, determining that the selected candidate title is a target title. According to the technical scheme, the title generation model or the title library can be utilized to recommend the title for the user according to the keyword input by the user, the title creation efficiency of the user can be improved, and the working time is saved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of computer technology, and in particular, to a title recommendation method, apparatus, computer device, computer-readable storage medium, and computer program product. Background Art

[0002] With the rapid development of the Internet advertising industry, the creativity and quality of advertising titles play a vital role in attracting user attention, increasing click-through rate and ultimately conversion rate.

[0003] However, advertisers often need to invest a lot of time and resources when creating advertising headlines to ensure that the headlines are attractive, which results in low creation efficiency.

[0004] It should be noted that the above content is not necessarily prior art, nor is it intended to limit the scope of patent protection of this application. Summary of the invention

[0005] The embodiments of the present application provide a title recommendation method, apparatus, computer device, computer-readable storage medium, and computer program product to solve or alleviate one or more of the technical problems raised above.

[0006] One aspect of an embodiment of the present application provides a title recommendation method, the method comprising: Displaying a target page, wherein the target page includes a target input box; Obtain target text through the target input box; Obtain multiple candidate titles based on the target text by using a pre-trained target title generation model or title library; Displaying a plurality of candidate titles on the target page; In response to selecting a candidate title from the plurality of candidate titles, the selected candidate title is determined to be a target title.

[0007] Optionally, obtaining the target text through the target input box includes: If the target text is not obtained, obtaining item information of the associated account; A target keyword is determined according to the item information, and the target keyword is determined as the target text.

[0008] Optionally, the target page is a secondary page under the primary page; the method further includes: In response to entering the target page, the text content in the input box of the primary page is used as the target text to be automatically filled into the target input box.

[0009] Optionally, the title library includes a first title library and a second title library, and the second title library is obtained by vectorizing the first title library; correspondingly, a plurality of candidate titles are obtained based on the target text by using a pre-trained title generation model or title library, including: Querying the first title library according to the target text to match and obtain a plurality of first associated titles; According to the text vector of the target text and the vectors of each title in the second title library, a plurality of second associative titles are matched and obtained; A plurality of candidate titles are determined according to the plurality of the first associated titles and the plurality of the second associated titles.

[0010] Optionally, the training operation of the target title generation model includes: Obtain an initial generation model and multiple pieces of initial training data; Cleaning the multiple pieces of initial training data to obtain multiple pieces of target training data; Using the plurality of target training data to train the initial generation model, to obtain a title generation model; Using the pre-acquired preference dataset to adjust the title generation model; Evaluate the titles generated by the adjusted title generation model to obtain evaluation results; When the evaluation result meets the preset requirement, the title generation model is determined as the target title generation model.

[0011] Optionally, cleaning the multiple pieces of initial training data to obtain multiple pieces of target training data includes: Determine a plurality of pieces of data to be cleaned that contain numbers in the initial training data; The numbers in the data to be cleaned are replaced with preset symbols to obtain a plurality of target training data.

[0012] Optionally, at least part of the multiple pieces of target training data are used to indicate the same task objective.

[0013] Optionally, the preference data set includes a plurality of prompts, and a preferred output and a non-preferred output corresponding to each of the prompts; adjusting the title generation model using the pre-acquired preference data set includes: inputting a plurality of the prompts into the title generation model to determine a first probability that the title generation model generates the preferred output and a second probability that the title generation model generates the non-preferred output; Determine a loss rate of the title generation model according to the first probability and the second probability; The title generation model is adjusted according to the loss rate.

[0014] Optionally, the method further comprises: Determine the loss rate according to the first probability, the second probability and a penalty parameter; The penalty parameter is used to increase the loss rate when the second probability exceeds a predetermined threshold.

[0015] Optionally, the method further comprises one or more of the following features: (1) Obtaining community article titles in real time, and updating the obtained community article titles into the title library; (2) generating a plurality of derivative titles according to the target title, and updating the plurality of derivative titles into the title library; (3) Regularly obtain high-frequency words in the community and generate a high-frequency word vector library based on the obtained high-frequency words in the community; When the target text is obtained, the target high-frequency word is matched according to the text vector of the target text and the vectors of each high-frequency word in the high-frequency word vector library; The target title generation model or title library is used to obtain a plurality of candidate titles based on the target text and the target high-frequency words.

[0016] Another aspect of an embodiment of the present application provides a title recommendation device, the device comprising: A first display module, used to display a target page, wherein the target page includes a target input box; A first acquisition module, used to acquire the target text through the target input box; A second acquisition module is used to acquire multiple candidate titles based on the target text by using a pre-trained target title generation model or title library; A second display module, used for displaying a plurality of candidate titles in the target page; The determination module is used for, in response to selecting a candidate title from the plurality of candidate titles, determining the selected candidate title as a target title.

[0017] Another aspect of an embodiment of the present application provides a computer device, including: at least one processor; and a memory communicatively coupled to the at least one processor; Wherein: the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method as described above.

[0018] Another aspect of an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and when the computer instructions are executed by a processor, the method described above is implemented.

[0019] Another aspect of an embodiment of the present application provides a computer program product, including a computer program, which implements the method described above when executed by a processor.

[0020] The embodiments of the present application using the above technical solution may include the following advantages: using a title generation model or a title library to recommend titles to users based on keywords entered by users can improve the efficiency of user title creation and save work time. BRIEF DESCRIPTION OF THE DRAWINGS The accompanying drawings exemplarily illustrate the embodiments and constitute a part of the specification, and together with the text description of the specification, are used to explain the exemplary implementation of the embodiments. The embodiments shown are for illustrative purposes only and do not limit the scope of the claims. In all drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0021] Figure 1 The following schematically shows an operating environment diagram of the title recommendation method according to the first embodiment of the present application; Figure 2 The flowchart of the title recommendation method according to the first embodiment of the present application is schematically shown; Figure 3 A schematic diagram schematically shows a specific implementation effect of the target page according to the first embodiment of the present application; Figure 4 Schematically shows Figure 2 Flow chart of sub-steps of step S202; Figure 5 The following schematically shows the implementation effect of the primary page according to the first embodiment of the present application; Figure 6 A schematic diagram showing the effect of displaying recommended titles after inputting text content into the input box of the primary page according to the first embodiment of the present application is shown; Figure 7 A schematic diagram schematically shows the effect of implementing a secondary page after the text content in the primary page input box is automatically filled into the target input box according to the first embodiment of the present application; Figure 8 Schematically shows Figure 2 Flow chart of sub-steps of step S204; Fig. 9 The following schematically shows a training operation flow chart of a target title generation model according to the first embodiment of the present application; Fig.10 Schematically shows Fig. 9Flow chart of sub-steps of step S902; Fig.11 Schematically shows Fig. 9 Flow chart of sub-steps of step S906; Fig.12 Schematically shows an application example of the title recommendation method according to an embodiment of the present application Fig.13 A block diagram schematically shows a title recommendation device according to the second embodiment of the present application; and Fig.14 The hardware architecture diagram of the computer device according to the third embodiment of the present application is schematically shown. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0023] It should be noted that the descriptions involving "first", "second", etc. in the embodiments of the present application are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in the field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0024] In the description of the present application, it should be understood that the numerical labels before the steps do not indicate the order in which the steps are executed, but are only used to facilitate the description of the present application and to distinguish each step, and therefore should not be understood as a limitation on the present application.

[0025] First, the following terms are explained: Large Language Models: Large Language Models, or LLMs for short, are a class of natural language processing (NLP) models with a large number of parameters that learn the complex patterns and structures of language by pre-training on large-scale text data.

[0026] Generalization: It is a core concept in the field of machine learning and artificial intelligence, which refers to the model's ability to perform on unseen data.

[0027] Vectorization: It is a concept in machine learning and natural language processing (NLP), which is the process of converting data into vector form.

[0028] Feature Vector: A concept in machine learning and data science that is used to represent the attributes or features of each instance (or sample) in a data set.

[0029] ES retrieval: refers to the search and retrieval operations performed using Elasticsearch (ES). Elasticsearch is an open source search engine based on Lucene that provides full-text search capabilities and is distributed and highly scalable.

[0030] EMBD service (Embedding Service): is a technology that converts text, images or other types of data into numerical vectors.

[0031] ANN (Approximate Nearest Neighbor) retrieval: It is an algorithm for fast similarity search in high-dimensional space, which is used to solve the problem of finding the most similar point to a given query point in a large-scale data set.

[0032] MoDS (Model-oriented Data Selection for Instruction Tuning) framework: model-oriented instruction data selection method.

[0033] Human Preference Dataset: Used in the field of Natural Language Processing (NLP) to train and evaluate whether the content generated by the model conforms to human preferences and standards.

[0034] DPO: Direct Preference Optimization, is a preference optimization algorithm based on direct human feedback. It adjusts the model parameters to make it more in line with user preferences by collecting users' direct evaluation of model output, such as satisfaction and preference.

[0035] IPO: Importance-weighted Preference Optimization, identity preference optimization, is a preference optimization algorithm based on user identity. It infers user preferences by analyzing user identity characteristics, such as age, gender, occupation, etc., and adjusts the model output accordingly.

[0036] L2 regularization: also known as Ridge Regression or Weight Decay, is a machine learning model technique used to prevent model overfitting and improve the generalization ability of the model.

[0037] KTO (Kahneman-Tversky Optimisation): A method for aligning large language models (LLMs).

[0038] Penalty parameter: is a hyperparameter used to adjust model complexity and control overfitting by adding an additional term to the loss function.

[0039] Secondly, to facilitate those skilled in the art to understand the technical solutions provided by the embodiments of the present application, the relevant technologies are described below: With the rapid development of digital marketing, the creativity and generation of promotional titles have become key factors in attracting potential customers. On many online platforms, advertising titles not only need to attract users' attention, but also need to be consistent with the culture of the platform and user preferences. However, the traditional advertising title creation process often relies on manual creativity, which is not only time-consuming and costly, but also difficult to ensure the innovation and attractiveness of the title.

[0040] To this end, the embodiment of the present application provides a technical solution for title recommendation. In this technical solution, (1) providing candidate titles to users based on the input target text can improve the relevance of the recommended titles to user needs; (2) through training, adjustment and evaluation and other steps, a model can be obtained that can meet human preferences and the generated titles can meet relevant quality standards. In this way, the efficiency of generating candidate titles based on the target text and the actual effect of the candidate titles can be improved, thereby further improving user satisfaction with the titles; (3) using the preference data set to adjust the title generation model, the titles generated by the model can be made more in line with human preference needs, thereby increasing the probability of the title being adopted; (4) adding a penalty parameter related to the second probability for calculating the loss rate, the loss rate can be increased when the second probability rises excessively, so as to further adjust the title generation model, thereby reducing the probability of the model generating non-preferred answers, improving the effect of the model-generated titles and user satisfaction. See below for details.

[0041] Finally, for ease of understanding, an exemplary operating environment is provided below.

[0042] like Figure 1 As shown, the operating environment diagram includes: server 2, network 4, client 6, where: Server 2 may be comprised of a single or multiple computing devices. The multiple computing devices may include virtualized computing instances. Virtualized computing instances may include virtual machines, such as simulations of computer systems, operating systems, servers, and the like. A computing device may load a virtual machine based on a virtual image and / or other data defining specific software (e.g., an operating system, a dedicated application, a server) for simulation. As the demand for different types of processing services changes, different virtual machines may be loaded and / or terminated on one or more computing devices. A hypervisor may be implemented to manage the use of different virtual machines on the same computing device.

[0043] The server 2 may be configured to communicate with the client 6, etc., via a network 4. The network 4 includes various network devices, such as routers, switches, multiplexers, hubs, modems, bridges, repeaters, firewalls, proxy devices, and / or the like. The network 4 may include physical links, such as coaxial cable links, twisted pair cable links, optical fiber links, combinations thereof, etc., or wireless links, such as cellular links, satellite links, Wi-Fi links, etc.

[0044] Server 2 can provide storage, reading, writing, querying, deleting and other services, such as providing the client with storage title library service, title recommendation service, running target title generation model, etc.

[0045] The client 6 may be an electronic device running an operating system such as Windows, Android™ or iOS, such as a smart phone, a tablet device, a laptop computer, a virtual reality device, a gaming device, a set-top box, a vehicle-mounted terminal, or a smart TV. Based on the above operating system, various application programs may be run, such as an advertisement creation program and a title editing program.

[0046] The client 6 can provide / configure a user access page for manipulating the server 2 or uploading an object, etc.

[0047] It should be noted that the above devices are exemplary, and the number and type of devices are adjustable in different scenarios or according to different needs.

[0048] The following describes the technical solution of the present application through multiple embodiments, taking the client 6 or the server 2 as the execution subject. It should be noted that these embodiments can be implemented in a variety of different forms and should not be interpreted as being limited to the embodiments described here.

[0049] Embodiment 1 Figure 2 The flowchart of the title recommendation method according to the first embodiment of the present application is schematically shown.

[0050] like Figure 2 As shown, the title recommendation method may include steps S200 to S208, wherein: Step S200, displaying a target page, wherein the target page includes a target input box; Step S202, obtaining target text through the target input box; Step S204, obtaining multiple candidate titles based on the target text by using a pre-trained target title generation model or title library; Step S206, displaying a plurality of candidate titles in the target page; Step S208, in response to selecting a candidate title from the plurality of candidate titles, determining the selected candidate title as a target title.

[0051] The title recommendation method provided in this embodiment uses a title generation model or a title library to recommend titles to users based on keywords input by users, which can improve the efficiency of users in creating titles and save work time.

[0052] The following combination Figure 2 , each step in steps S200~S208 and other optional steps are explained in detail.

[0053] Step S200 , displaying a target page, wherein the target page includes a target input box.

[0054] The target page can be in the form of an independent page, a pop-up window, or a sidebar. In addition to the target input box, the target page can also include a button or a drop-down list for selecting related fields or industries, so as to generate a title that is closer to the related field selected by the user. The target page can also include a display area for displaying recommended titles based on the content entered in the target input box. The specific implementation effect of the target page is as follows Figure 3 shown.

[0055] Step S202 , obtain the target text through the target input box.

[0056] The target text can be different types of text such as Chinese characters, numbers or symbols, such as "no krypton gold", "high-quality pictures", etc. In addition to text input, the target text can also be obtained through voice input, image recognition or text extraction.

[0057] In a specific application, the target page can also display recommended titles selected according to preset rules when no content is entered in the target input box. However, the inventors have found that when the user first enters the target page, since the target text for obtaining candidate titles is not entered in the target input box, the recommended titles displayed on the target page may be far from what the user needs. The following two exemplary methods are provided to enhance the relevance of the recommended titles to user needs when no content is entered in the target input box.

[0058] Method 1: Figure 4 As shown, step S202 includes: S400: If the target text is not obtained, obtain item information of the associated account.

[0059] S402: Determine a target keyword according to the item information, and determine the target keyword as the target text.

[0060] Item information may be information related to products that the associated account is selling, has purchased, has collected, or has viewed multiple times.

[0061] For example, if it is obtained that the associated account is selling a certain smart watch, the recommended associated title can be "Upgrade your smart life, special smart watches, limited time discount, don't miss it!", "Explore the infinite possibilities of smart watches, buy now to enjoy early bird discounts!", etc.

[0062] In this embodiment, when the user does not enter the target text, titles related to the product information associated with the account are recommended to the user. In this way, the relevance of the recommended titles to the user's potential needs when no target text is entered on the target page can be improved, as well as the probability of the recommended titles being adopted, thereby improving the user experience and the efficiency of the user's title selection.

[0063] In a second method, the target page is a second-level page under the first-level page, and the method further includes: In response to entering the target page, the text content in the input box of the primary page is used as the target text to be automatically filled into the target input box.

[0064] The first-level page can be the homepage for users to create promotional ideas. Users can complete the entire creation process here, including setting the cover, name and title of the promotional idea. Figure 5 shown.

[0065] When users set a title for a promotional creative through the input box on the first-level page, the first-level page will display recommended titles below the input box based on the content entered by the user. If the recommended title does not meet the user's needs, the user can also enter the target page through specific operations (such as clicking a button) to obtain more recommended titles. The implementation effect of displaying recommended titles after entering text content in the input box on the first-level page is as follows Figure 6 shown.

[0066] In this embodiment, titles are automatically recommended based on the content entered by the user in the primary page input box. This can improve the relevance of the title that the user sees when entering the target page to the user's needs, thereby improving user satisfaction and the efficiency of creating titles. The effect of the secondary page after the text content in the primary page input box is automatically filled into the target input box is as follows Figure 7 shown.

[0067] Step S204 , through a pre-trained target title generation model or title library, multiple candidate titles are obtained based on the target text.

[0068] Specifically, a target title generation model can be used to directly generate multiple original titles based on the target text, or a title library can be searched through the target text to obtain multiple matching existing titles.

[0069] For example, when the input target text is "no krypton gold" and "high-quality graphics", the target title generation model can generate original titles such as "No krypton gold! Games with high-quality graphics? Yes!", "No routines, no krypton gold, exquisite high-quality graphics!", etc., and existing titles such as "People who don't krypton gold can also have super high-quality graphics" and "No krypton gold? High-quality graphics, conscientious good games" can also be retrieved in the title library. According to actual needs, relevant information about the promotion audience of the target title can also be included to provide different candidate titles according to different promotion audiences.

[0070] In this embodiment, candidate titles are provided to the user based on the input target text, which can improve the relevance of the recommended title to the user's needs, while reducing the difficulty of creating titles, improving the efficiency of title creation, and saving the user's creation time.

[0071] In an optional embodiment, the title library includes a first title library and a second title library, and the second title library is vectorized from the first title library. Figure 8 As shown, step S204 includes: S800: Search the first title library according to the target text to match and obtain a plurality of first associated titles.

[0072] S802: Match and obtain a plurality of second associative titles according to the text vector of the target text and the vectors of each title in the second title library.

[0073] S804: Determine a plurality of candidate titles according to a plurality of the first associated titles and a plurality of the second associated titles.

[0074] The first title database can be queried through text matching (such as ES search) and other methods, the feature vector of the target text can be obtained through the text EMBD service, and the vector matching can be performed in the second title database through ANN search and other methods.

[0075] The similarity between the titles matched from the title library and the target text can be determined by text matching, semantic matching or vector matching, and the titles are sorted according to the similarity, and a predetermined number of titles with the highest ranking are selected as candidate titles.

[0076] Combining text matching and vector matching to match the target text with a suitable title can further improve the matching degree between the candidate title and the target text, making the candidate title more suitable for user needs, improving user satisfaction and the efficiency of creating titles.

[0077] In an optional embodiment, if Fig. 9 As shown, the training operation of the target title generation model includes: S900, obtaining an initial generation model and a plurality of initial training data.

[0078] S902: Clean up the multiple pieces of initial training data to obtain multiple pieces of target training data.

[0079] S904: Train the initial generation model using the plurality of target training data to obtain a title generation model.

[0080] S906: Adjust the title generation model using the pre-acquired preference data set.

[0081] S908, evaluating the title generated by the adjusted title generation model to obtain an evaluation result.

[0082] S910: When the evaluation result meets a preset requirement, the title generation model is determined as the target title generation model.

[0083] The initial generation model can be an open source third-party large language model or other available models. The MoDS framework can be used to clean at least part of the initial training data.

[0084] The target training data may include a certain proportion of proprietary task data, commercial domain data, and open domain data, so that the trained title generation model can maintain a high generalization ability. Among them, the proprietary task data includes some existing titles with high click-through rates and their corresponding extracted keywords and other information, which are used to strengthen the initial generation model's compliance with the title generation task. The commercial domain data may include commercial rewriting tasks and rewriting results of the titles, which are used to strengthen the generalization ability of the initial generation model in commercial-related tasks. Open domain data may include tasks and results in various non-commercial scenarios, which can improve the generalization ability of the initial generation model for basic tasks.

[0085] Each target training data may include an instruction and an output, so as to perform model training through supervised learning methods, so that the initial generation model learns how to generate corresponding outputs according to given instructions.

[0086] For example, the target training data can be: {"instruction": "You are a very powerful promotion creative expert. I will give you a set of keywords and related industry information entered by the promoter. Use spaces to separate the keywords. Generate a promotion title related to the industry that can attract users to click. You must promote potential products in a positive way. The title must be fluent and complete, and in line with the temperament of the X site. You don't have to use all of these keywords (when there is only one keyword, try to generate a longer title). Do not return other content unless the brand name is provided in the keyword. Other information such as price, brand, platform, date, etc. that users need to supplement should be replaced with xxx. The generated title must be at least 5 characters and no more than 18 characters. The first-level industry information is household appliances, and the second-level industry is small household appliances. The keywords are: tap water hazards.", "output": "Does your family still drink tap water? See how terrible tap water is when magnified!"} Among them, the instruction part is the instruction, and the output part is the output.

[0087] When evaluating the generated titles, a variety of evaluation criteria may be used. Several exemplary evaluation criteria and their corresponding evaluation methods are provided below.

[0088] Standard one, fluency, is used to evaluate whether the generated title is fluent.

[0089] The titles of real community videos or graphic dynamics can be used as positive samples, and the titles obtained by scrambling the word order or replacing words in the real titles can be used as negative samples to train the binary classification model to obtain an evaluation model for evaluating the fluency of the titles.

[0090] The second criterion, style score, is used to evaluate whether the generated title conforms to the specific community style.

[0091] The titles of real community videos or graphic dynamics can be used as positive samples, and titles randomly generated by a large language model can be used as negative samples to train the binary classification model to obtain an evaluation model for evaluating the title style score.

[0092] Standard three, quality score, is used to evaluate whether the generated title is a high-quality title, such as whether it may receive a higher click-through rate.

[0093] We can input multiple manually screened high-click rate titles into the large language model to determine the reasons why they can achieve high click rates. Then, we use these reasons as the criteria for judging whether the title is high-quality, and use at least two models to judge the generated title, and obtain the quality score of the title based on the judgment results.

[0094] In this embodiment, through the steps of training, adjustment and evaluation, a model can be obtained that can meet human preferences and generate titles that meet relevant quality standards. As a result, the efficiency of generating candidate titles based on the target text and the actual effect of the candidate titles can be improved, thereby further improving user satisfaction with the titles.

[0095] In an optional embodiment, if Fig.10 As shown, step S902 may include: S1000, determining that a plurality of pieces of the initial training data contain data to be cleaned that include numbers.

[0096] S1002, replacing numbers in the data to be cleaned with preset symbols to obtain a plurality of target training data.

[0097] The preset symbol may be an English letter or a special symbol, etc. After the numbers are replaced with the preset symbols, the replaced training data may be verified to ensure the fluency and smoothness of the data.

[0098] For example, the data to be cleaned is: {"instruction": "Your task is to determine whether the given promotion title is a promotion title that can attract people to click, and briefly analyze the reasons. The promotion title is: April Cost-Effective Installation Guide, assemble better computers with less budget, a total of 30 sets!", "output": "Yes. This promotion mainly highlights the advantages of the budget and limits the purchase quantity, which is easy to attract people to click."}, the numbers "4" and "30" can be replaced with the English letter "X" to obtain the target training data as follows: {"instruction": "Your task is to determine whether the given promotional title is a clickable one and briefly analyze the reasons. The promotional title is: Cost-effective installation guide for X month, assemble better computers with less budget, total XX sets!", "output":"Yes. This promotion mainly highlights the advantages of budget and limits the purchase quantity, which is easy to attract clicks."} In this embodiment, the numbers with semantic information in the training data are replaced with symbols without semantic information, which can reduce the dependence of the title generation model on specific numbers and further improve the generalization ability of the model, thereby ensuring the effect of model-generated titles.

[0099] In an optional embodiment, at least part of the target training data among the multiple target training data are used to indicate the same task objective.

[0100] For example, “Please design an advertising headline for the latest smartphone that attracts young consumers.” and “Generate an advertising headline to promote the new smartphone’s high-speed processor on social media.” These two different instructions are used to instruct the model to complete the task goal of “generating an advertising headline for a new smartphone.”

[0101] By using different instructions for the same task objective to train the model, the model can learn multiple ways of expressing the same task. This can further improve the generalization and generation capabilities of the title generation model and enhance the overall stability and reliability of the model.

[0102] In an optional embodiment, the preference data set includes multiple prompts, and a preferred output and a non-preferred output corresponding to each prompt, such as Fig.11 As shown, step S906 includes: S1100, inputting a plurality of the prompts into the title generation model to determine a first probability that the title generation model generates the preferred output and a second probability that the title generation model generates the non-preferred output.

[0103] S1102: Determine a loss rate of the title generation model according to the first probability and the second probability.

[0104] S1104: Adjust the title generation model according to the loss rate.

[0105] The preference data set may be a manually annotated human preference data set, etc. For example, one of the preference data in the preference data set may be: “Prompt: Write a promotional headline for a newly released smartphone that touts long battery life and fast charging technology.

[0106] Preferred output: Ultra-long battery life, ultra-fast charging: New smartphones let you say goodbye to battery anxiety! Non-preferred output: New phone available, features include battery and charging. ” After the above prompts are input into the title recommendation model, the title recommendation model will calculate the probability of generating two outputs, namely the first probability and the second probability, and then calculate the loss rate of the title generation model through a preset formula (such as the DPO formula). According to the value of the loss rate or the gradient it belongs to, it can be determined how to modify the model parameters to adjust the title generation model.

[0107] In this embodiment, the title generation model is adjusted using the preference data set, so that the titles generated by the model are more in line with human preference needs, thereby increasing the probability of the titles being adopted and improving user satisfaction and usage experience.

[0108] In addition to using the preference dataset to adjust the title generation model, other additional optimization measures can be used to improve the adjustment effect. The following are several exemplary optimization measures.

[0109] Optimization 1: Improve the information density of prompts. By streamlining the text content of prompts in the preference dataset and increasing the information content, the differences between input prompts are not limited to different keywords, which can improve the generalization ability of the model during the convergence process of the title generation model.

[0110] Optimization 2: Introduce the IPO algorithm and add the L2 regularization term to the DPO formula. This can avoid the overfitting problem of the DPO algorithm and improve the generalization ability of the title generation model.

[0111] Optimization 3: The KTO algorithm is introduced to decouple positive and negative samples. This allows the model to eliminate the need for paired positive and negative samples and to evaluate the quality of each sample individually, thereby improving the performance of the model and ensuring the lower limit of the model.

[0112] Optimization four, the method further includes: Determine the loss rate according to the first probability, the second probability and a penalty parameter; The penalty parameter is used to increase the loss rate when the second probability exceeds a predetermined threshold.

[0113] During the model adjustment process, when the first probability increases and the second probability decreases, the loss rate obtained will gradually decrease, indicating that the model adjustment effect is good. However, the inventors found that in actual applications, both the first probability and the second probability increase, but because the first probability increases faster than the second probability, the loss rate obtained also decreases, but the second probability, that is, the probability of the model generating a non-preferred answer, increases, which is unacceptable.

[0114] In view of this, in this embodiment, a penalty parameter related to the second probability is added to calculate the loss rate. The loss rate can be increased when the second probability rises excessively to further adjust the title generation model, thereby reducing the probability of the model generating non-preferred answers, improving the effect of the model generating titles and user satisfaction.

[0115] Step S206 , displaying a plurality of the candidate titles in the target page.

[0116] A preset number of candidate titles can be displayed in the form of a list, drop-down list, or pop-up window. When displaying candidate titles, they can be sorted according to the relevance of the candidate titles to the target text or the user's historical selection. It is also possible to allow users to give feedback on the candidate titles while displaying them, such as likes or disinterest, to further optimize the recommendation mechanism.

[0117] Step S208 In response to selecting a candidate title from the plurality of candidate titles, determining the selected candidate title as a target title.

[0118] After the user selects the target title, the user can also rate the candidate titles or provide feedback so that the subsequent recommended titles can be adjusted based on the feedback. At the same time, the user's historical selections can be recorded, the user's preferences can be analyzed, and titles can be recommended to the user based on these preferences in the future. When the user is not satisfied with the displayed candidate titles, the user can also be allowed to refresh or change the displayed candidate titles so that the target page displays a batch of new candidate titles for the user to choose from.

[0119] In order to improve the novelty of candidate titles and make them more in line with current affairs and trends, the following exemplary measures can be used to improve the novelty of titles and generate more attractive titles.

[0120] Measure 1: Obtain community article titles in real time and update the obtained community article titles to the title library. Community articles may include various videos, audios or graphic dynamics on the network platform. By using the latest community article titles, the title library can grasp the latest community trends and improve the novelty of the titles in the title library.

[0121] Measure 2: Generate multiple derivative titles based on the target title and update the multiple derivative titles into the title library. Regularly generate multiple similar titles based on the most recently selected target title, which can increase the number of titles in the title library and improve the relevance of the titles in the title library to the latest user needs.

[0122] Measure 3: Regularly obtain high-frequency words in the community and generate a high-frequency word vector library based on the obtained high-frequency words in the community; When the target text is obtained, the target high-frequency word is matched according to the text vector of the target text and the vectors of each high-frequency word in the high-frequency word vector library; The target title generation model or title library is used to obtain a plurality of candidate titles based on the target text and the target high-frequency words.

[0123] Community high-frequency words refer to hot words or memes that are widely used by users in the community. By using the community high-frequency words that match the target text together with the target text to obtain candidate titles, the candidate titles can more accurately use community hot content that fits user needs, thereby generating more novel and more attractive titles.

[0124] In order to make this application easier to understand, the following Fig.12 An exemplary application is provided. S11, the user enters the ad creative title A (i.e., text content) in the input box of the ad creative creation page (i.e., the first-level page); S12, query the corresponding title library according to the creative title A and its corresponding feature vector to obtain multiple associative titles T1, T2, T3...T10.

[0125] S13, determining T1, T2, and T3 as candidate titles, and displaying them in a list below the input box on the creative creation page.

[0126] S14, the user clicks the button above the input box to display a recommended title pop-up window (i.e., the target page).

[0127] S15, the creative title A is automatically input into the keyword input box (ie, the target input box) of the recommended title pop-up window.

[0128] S16, input the creative title A into the target title generation model, obtain multiple candidate titles L1, L2, L3, and display them in the form of a list below the keyword input box.

[0129] S17, the user continues to input keyword B in the keyword input box.

[0130] S18, input the creative title A and keyword B into the target title generation model, obtain multiple candidate titles L4, L5, L6, and display them in the form of a list below the keyword input box.

[0131] S19, in response to the user selecting the candidate title L5, determining the candidate title L5 as the title that will be ultimately used for the advertisement creative (ie, the target title).

[0132] Embodiment 2 Fig.13The block diagram of the title recommendation device according to the second embodiment of the present application is schematically shown. The device can be divided into one or more program modules, one or more program modules are stored in a storage medium, and are executed by one or more processors to complete the embodiment of the present application. The program module referred to in the embodiment of the present application refers to a series of computer program instruction segments that can complete specific functions. The following description will specifically introduce the functions of each program module in this embodiment. Fig.13 As shown, the device 1000 may include: a first display module 1100, a first acquisition module 1200, a second acquisition module 1300, a second display module 1400, and a determination module 1500, wherein: A first display module 1100 is used to display a target page, wherein the target page includes a target input box; A first acquisition module 1200 is used to acquire a target text through the target input box; The second acquisition module 1300 is used to acquire multiple candidate titles based on the target text by using a pre-trained target title generation model or title library; A second display module 1400 is used to display a plurality of candidate titles in the target page; The determination module 1500 is configured to, in response to selecting a candidate title from the plurality of candidate titles, determine the selected candidate title as a target title.

[0133] As an optional embodiment, the first obtaining module 1200 is further configured to: If the target text is not obtained, obtaining item information of the associated account; A target keyword is determined according to the item information, and the target keyword is determined as the target text.

[0134] As an optional embodiment, the target page is a secondary page under the primary page, and the device 1000 further includes an automatic filling module for: In response to entering the target page, the text content in the input box of the primary page is used as the target text to be automatically filled into the target input box.

[0135] As an optional embodiment, the title library includes a first title library and a second title library, and the second title library is obtained by vectorizing the first title library. Correspondingly, the second acquisition module 1300 is further used for: Querying the first title library according to the target text to match and obtain a plurality of first associated titles; According to the text vector of the target text and the vectors of each title in the second title library, a plurality of second associative titles are matched and obtained; A plurality of candidate titles are determined according to the plurality of the first associated titles and the plurality of the second associated titles.

[0136] As an optional embodiment, the apparatus 1000 further includes a model training module, which is used to: Obtain an initial generation model and multiple pieces of initial training data; Cleaning the multiple pieces of initial training data to obtain multiple pieces of target training data; Using the plurality of target training data to train the initial generation model, to obtain a title generation model; Using the pre-acquired preference dataset to adjust the title generation model; Evaluate the titles generated by the adjusted title generation model to obtain evaluation results; When the evaluation result meets the preset requirement, the title generation model is determined as the target title generation model.

[0137] As an optional embodiment, the model training module is further used for: Determine a plurality of pieces of data to be cleaned that contain numbers in the initial training data; The numbers in the data to be cleaned are replaced with preset symbols to obtain a plurality of target training data.

[0138] As an optional embodiment, at least part of the target training data among the multiple target training data are used to indicate the same task target.

[0139] As an optional embodiment, the preference data set includes multiple prompts, and a preferred output and a non-preferred output corresponding to each of the prompts, and the model training module is further used to: inputting a plurality of the prompts into the title generation model to determine a first probability that the title generation model generates the preferred output and a second probability that the title generation model generates the non-preferred output; Determine a loss rate of the title generation model according to the first probability and the second probability; The title generation model is adjusted according to the loss rate.

[0140] As an optional embodiment, the apparatus 1000 further includes a third determining module, configured to: Determine the loss rate according to the first probability, the second probability and a penalty parameter; The penalty parameter is used to increase the loss rate when the second probability exceeds a predetermined threshold.

[0141] As an optional embodiment, the apparatus 1000 is further used for: (1) Obtaining community article titles in real time, and updating the obtained community article titles into the title library; (2) generating a plurality of derivative titles according to the target title, and updating the plurality of derivative titles into the title library; (3) Regularly obtain high-frequency words in the community and generate a high-frequency word vector library based on the obtained high-frequency words in the community; When the target text is obtained, the target high-frequency word is matched according to the text vector of the target text and the vectors of each high-frequency word in the high-frequency word vector library; The target title generation model or title library is used to obtain a plurality of candidate titles based on the target text and the target high-frequency words.

[0142] Embodiment 3 Fig.14 The schematic diagram of the hardware architecture of a computer device 10000 suitable for implementing the title recommendation method according to the third embodiment of the present application is shown. In some embodiments, the computer device 10000 may be a terminal device such as a smart phone, a wearable device, a tablet computer, a personal computer, a vehicle terminal, a game console, a virtual device, a workbench, a digital assistant, a set-top box, a robot, etc. In other embodiments, the computer device 10000 may be a rack server, a blade server, a tower server, or a cabinet server (including an independent server, or a server cluster composed of multiple servers), etc. Fig.14 As shown, the computer device 10000 includes but is not limited to: a memory 10010, a processor 10020, and a network interface 10030 that can communicate with each other through a system bus. Among them: The memory 10010 includes at least one type of computer-readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (such as SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 10010 can be an internal storage module of the computer device 10000, such as a hard disk or memory of the computer device 10000. In other embodiments, the memory 10010 can also be an external storage device of the computer device 10000, such as a plug-in hard disk equipped on the computer device 10000, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. Of course, the memory 10010 can also include both the internal storage module of the computer device 10000 and its external storage device. In this embodiment, the memory 10010 is generally used to store the operating system and various application software installed in the computer device 10000, such as the program code of the title recommendation method, etc. In addition, the memory 10010 can also be used to temporarily store various data that have been output or will be output.

[0143] In some embodiments, the processor 10020 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other chips. The processor 10020 is generally used to control the overall operation of the computer device 10000, such as performing control and processing related to data interaction or communication with the computer device 10000. In this embodiment, the processor 10020 is used to run the program code stored in the memory 10010 or process data.

[0144] The network interface 10030 may include a wireless network interface or a wired network interface, and the network interface 10030 is generally used to establish a communication link between the computer device 10000 and other computer devices. For example, the network interface 10030 is used to connect the computer device 10000 to an external terminal through a network, and to establish a data transmission channel and a communication link between the computer device 10000 and the external terminal. The network may be a wireless or wired network such as an intranet, the Internet, the Global System of Mobile communication (GSM), Wideband Code Division Multiple Access (WCDMA), 4G network, 5G network, Bluetooth, Wi-Fi, etc.

[0145] It should be pointed out that Fig.14 Only a computer device having components 10010 - 10030 is shown, but it should be understood that implementation of all of the components shown is not a requirement, and more or fewer components may alternatively be implemented.

[0146] In this embodiment, the title recommendation method stored in the memory 10010 can also be divided into one or more program modules and executed by one or more processors (such as processor 10020) to complete the embodiment of the present application.

[0147] Embodiment 4 The embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored, wherein when the computer program is executed by a processor, the steps of the title recommendation method in the embodiment are implemented.

[0148] In this embodiment, the computer-readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the computer-readable storage medium can be an internal storage unit of a computer device, such as a hard disk or memory of the computer device. In other embodiments, the computer-readable storage medium can also be an external storage device of a computer device, such as a plug-in hard disk equipped on the computer device, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. Of course, the computer-readable storage medium can also include both the internal storage unit of the computer device and its external storage device. In this embodiment, the computer-readable storage medium is generally used to store an operating system and various application software installed on the computer device, such as the program code of the title recommendation method in the embodiment. In addition, the computer-readable storage medium can also be used to temporarily store various types of data that have been output or are to be output.

[0149] Embodiment 5 An embodiment of the present application also provides a computer program product, including a computer program, which implements the method in the above embodiment when executed by a processor.

[0150] Obviously, those skilled in the art should understand that the modules or steps of the above-mentioned embodiments of the present application can be implemented by general-purpose computer devices, they can be concentrated on a single computer device, or distributed on a network composed of multiple computer devices, optionally, they can be implemented by executable program codes of computer devices, so that they can be stored in a storage device and executed by the computer device, and in some cases, the steps shown or described can be executed in a different order from that herein, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present application are not limited to any specific combination of hardware and software.

[0151] It should be noted that the above are only preferred embodiments of the present application, and the patent protection scope of the present application is not limited thereto. Any equivalent structure or equivalent process transformation made using the contents of the specification and drawings of the present application, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A title recommendation method, characterized in that: The method comprises: Displaying a target page, wherein the target page includes a target input box; Obtain target text through the target input box; Obtain multiple candidate titles based on the target text by using a pre-trained target title generation model or title library; Displaying a plurality of candidate titles on the target page; In response to selecting a candidate title from the plurality of candidate titles, the selected candidate title is determined to be a target title.

2. The method according to claim 1, characterized in that: Obtaining the target text through the target input box includes: If the target text is not obtained, obtaining item information of the associated account; A target keyword is determined according to the item information, and the target keyword is determined as the target text.

3. The method according to claim 1, characterized in that The target page is a secondary page under the primary page; the method further includes: In response to entering the target page, the text content in the input box of the primary page is used as the target text to be automatically filled into the target input box.

4. The method according to claim 1, characterized in that The title library includes a first title library and a second title library, and the second title library is obtained by vectorizing the first title library; correspondingly, a plurality of candidate titles are obtained based on the target text through a pre-trained title generation model or title library, including: Querying the first title library according to the target text to match and obtain a plurality of first associated titles; According to the text vector of the target text and the vectors of each title in the second title library, a plurality of second associative titles are matched and obtained; A plurality of candidate titles are determined according to the plurality of the first associated titles and the plurality of the second associated titles.

5. The method according to claim 1, characterized in that The training operation of the target title generation model includes: Obtain an initial generation model and multiple pieces of initial training data; Cleaning the multiple pieces of initial training data to obtain multiple pieces of target training data; Using the plurality of target training data to train the initial generation model, to obtain a title generation model; Using the pre-acquired preference dataset to adjust the title generation model; Evaluate the titles generated by the adjusted title generation model to obtain evaluation results; When the evaluation result meets the preset requirement, the title generation model is determined as the target title generation model.

6. The method according to claim 5, characterized in that Cleaning the multiple pieces of initial training data to obtain multiple pieces of target training data includes: Determine a plurality of pieces of data to be cleaned that contain numbers in the initial training data; The numbers in the data to be cleaned are replaced with preset symbols to obtain a plurality of target training data.

7. The method according to claim 5, characterized in that At least part of the target training data among the multiple pieces of target training data are used to indicate the same task target.

8. The method according to claim 5, characterized in that The preference data set includes a plurality of prompts, and a preferred output and a non-preferred output corresponding to each of the prompts; adjusting the title generation model using the pre-acquired preference data set includes: inputting a plurality of the prompts into the title generation model to determine a first probability that the title generation model generates the preferred output and a second probability that the title generation model generates the non-preferred output; Determine a loss rate of the title generation model according to the first probability and the second probability; The title generation model is adjusted according to the loss rate.

9. The method according to claim 8, characterized in that The method further comprises: Determine the loss rate according to the first probability, the second probability and a penalty parameter; The penalty parameter is used to increase the loss rate when the second probability exceeds a predetermined threshold.

10. The method according to claim 1, characterized in that The method may further include one or more of the following features: (1) Obtaining community article titles in real time, and updating the obtained community article titles into the title library; (2) generating a plurality of derivative titles according to the target title, and updating the plurality of derivative titles into the title library; (3) Regularly obtain high-frequency words in the community and generate a high-frequency word vector library based on the obtained high-frequency words in the community; When the target text is obtained, the target high-frequency word is matched according to the text vector of the target text and the vectors of each high-frequency word in the high-frequency word vector library; The target title generation model or title library is used to obtain a plurality of candidate titles based on the target text and the target high-frequency words.

11. A title recommendation device, characterized in that: The device comprises: A first display module, used to display a target page, wherein the target page includes a target input box; A first acquisition module, used to acquire the target text through the target input box; A second acquisition module is used to acquire multiple candidate titles based on the target text by using a pre-trained target title generation model or title library; A second display module, used for displaying a plurality of candidate titles in the target page; The determination module is used for, in response to selecting a candidate title from the plurality of candidate titles, determining the selected candidate title as a target title.

12. A computer device, characterized in that: include: at least one processor; and a memory communicatively connected to the at least one processor; wherein: The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and when the computer instructions are executed by a processor, the method according to any one of claims 1 to 10 is implemented.

14. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

Images