Object feature determination method and device, electronic equipment and storage medium

Abstract

The invention relates to an object feature determination method and device, electronic equipment and a storage medium, and the method comprises the steps: inputting object description information of a sample recommendation object into a large language model, and carrying out the conversion analysis of a target interaction behavior based on a first preset thinking chain prompt text through the large language model, thereby obtaining a conversion reason; the first preset thinking chain prompt text is used for prompting a reasoning process for generating a target interaction behavior for the sample recommendation object, and the conversion reason comprises a reasoning result of at least one dimension in the reasoning process, and then performing feature representation on the sample recommendation object based on the conversion reason to obtain a first object feature of the sample recommendation object; and determining a second object feature of the sample interaction object based on the first object feature of each sample recommendation object in the sample recommendation object sequence of the sample interaction object. According to the invention, the accuracy of object representation oriented to the specific pre-estimation task is improved, the time consumption is shortened, and the training efficiency and accuracy of the recommendation probability pre-estimation model oriented to the specific pre-estimation task are improved.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a method, device, electronic device and storage medium for determining object features. Background Art

[0002] With the development of artificial intelligence technology, in the field of recommendation, the recommendation probability estimation model is usually used to estimate the probability of a user making a certain decision (such as click, purchase, like, comment, etc.) on the recommended object (such as a short video, a product, etc.), and then make recommendations to the user based on the estimated probability.

[0003] When related technologies are used to train recommendation probability prediction models for specific prediction tasks (such as clicks, purchases, likes, comments, etc.), the feature representation of sample users and sample recommendation objects (also called items) is not only time-consuming but also has poor accuracy, resulting in low efficiency in model training and poor estimation accuracy of the model in the prediction tasks. Summary of the invention

[0004] The present disclosure provides an object feature determination method, device, electronic device and storage medium to at least solve the problem that when training a recommendation probability estimation model for a specific estimation task in the related art, the feature representation method of sample users and sample recommended objects is not only time-consuming but also has poor accuracy, resulting in low efficiency of model training and poor estimation accuracy of the model in the estimation task. The technical solution of the present disclosure is as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, a method for determining an object feature is provided, comprising:

[0006] For any sample recommendation object in the sample recommendation object set, obtaining object description information of the sample recommendation object;

[0007] The object description information is input into a large language model, and a conversion analysis of the target interactive behavior is performed through the large language model based on a first preset thought chain prompt text to obtain a conversion reason of the sample recommended object corresponding to the target interactive behavior; the first preset thought chain prompt text is used to prompt the reasoning process of generating the target interactive behavior for the sample recommended object; the conversion reason includes the reasoning result of at least one dimension in the reasoning process, and the reasoning result of each dimension represents the performance of the sample recommended object in generating the target interactive behavior in the dimension;

[0008] Based on the conversion reason of the sample recommended object corresponding to the target interactive behavior, the sample recommended object is characterized to obtain a first object feature of the sample recommended object;

[0009] Based on the first object features of each sample recommended object in the sample recommendation object sequence of the sample interaction object within a historical time period, determine the second object features of the sample interaction object; wherein the first object features of the sample recommendation object and the second object features of the sample interaction object are used to train a recommendation probability prediction model, and the recommendation probability prediction model is used to estimate the probability of generating the target interaction behavior.

[0010] In some exemplary embodiments, the performing feature representation on the sample recommendation object based on the conversion reason of the sample recommendation object corresponding to the target interactive behavior to obtain the first object feature of the sample recommendation object includes:

[0011] The conversion reason of the sample recommendation object corresponding to the target interactive behavior is input into the large language model, and the conversion reason is subjected to keyword extraction processing by the large language model according to a preset keyword extraction prompt to obtain a keyword extraction result corresponding to the sample recommendation object; the preset keyword extraction prompt is used to instruct to perform a keyword extraction task according to a keyword extraction example, and the keyword extraction example represents the output of keywords in a sample text according to a specified format;

[0012] Based on the keyword extraction result corresponding to the sample recommended object, a first object feature of the sample recommended object is generated.

[0013] In some exemplary embodiments, generating the first object feature of the sample recommendation object based on the keyword extraction result corresponding to the sample recommendation object includes:

[0014] Performing text encoding processing on the conversion reasons of the sample recommendation object corresponding to the target interactive behavior to obtain a text encoding result corresponding to the sample recommendation object;

[0015] The keyword extraction result corresponding to the sample recommendation object is used as a sparse feature, and the text encoding result corresponding to the sample recommendation object is used as a dense feature, and the sparse feature and the dense feature are concatenated to obtain a first object feature of the sample recommendation object.

[0016] In some exemplary embodiments, after inputting the object description information into the large language model, performing conversion analysis of the target interactive behavior based on the first preset thought chain prompt text by the large language model, and obtaining the conversion reason of the sample recommended object, the method further includes:

[0017] Obtain a sample recommendation object sequence of the sample interaction object in a historical time period, and determine a positive sample recommendation object and a negative sample recommendation object in the sample recommendation object sequence; the positive sample recommendation object refers to a sample recommendation object for which the sample interaction object generates the target interaction behavior, and the negative sample recommendation object refers to a sample recommendation object for which the sample interaction object does not generate the target interaction behavior;

[0018] Based on the conversion reasons of the positive sample recommended object corresponding to the target interactive behavior and the conversion reasons of the negative sample recommended object corresponding to the target interactive behavior, determine the preference conversion reasons of the sample interactive object when generating the target interactive behavior, and generate the portrait features of the sample interactive object based on the preference conversion reasons of the sample interactive object;

[0019] The determining, based on the first object feature of each sample recommended object in a sample recommended object sequence of the sample interactive object within a historical time period, the second object feature of the sample interactive object comprises:

[0020] The first object feature of each sample recommended object in the sample recommended object sequence and the portrait feature of the sample interactive object are concatenated to obtain the second object feature of the sample interactive object.

[0021] In some exemplary embodiments, before performing feature representation on the sample recommended object based on the conversion reason of the sample recommended object corresponding to the target interactive behavior to obtain the first object feature of the sample recommended object, the method further includes:

[0022] Determining a set of preference conversion reasons based on the preference conversion reasons of each sample interaction object;

[0023] For each conversion reason of the sample recommendation object corresponding to the target interactive behavior, the conversion reason is adjusted based on the preferred conversion reason set to obtain an adjusted conversion reason corresponding to each sample recommendation object; the adjusted conversion reason is a subset of the preferred conversion reason set;

[0024] The adjusted conversion reason corresponding to each of the sample recommended objects is used as the conversion reason of the sample recommended object corresponding to the target interactive behavior, and the step of performing feature representation on the sample recommended objects based on the conversion reason of the sample recommended objects corresponding to the target interactive behavior to obtain the first object feature of the sample recommended objects is performed.

[0025] In some exemplary embodiments, the determining the preferred conversion reason of the sample interaction object when generating the target interaction behavior based on the conversion reason of the positive sample recommendation object corresponding to the target interaction behavior and the conversion reason of the negative sample recommendation object corresponding to the target interaction behavior, and generating the portrait feature of the sample interaction object based on the preferred conversion reason of the sample interaction object includes:

[0026] Acquire a second preset thought chain prompt text; the second preset thought chain prompt text is used to prompt to analyze the preference of the sample interaction object when generating the target interaction behavior according to multiple steps;

[0027] Inputting the conversion reasons of the positive sample recommended object corresponding to the target interactive behavior and the conversion reasons of the negative sample recommended object corresponding to the target interactive behavior into the large language model, performing preference analysis of the sample interactive object based on the second preset thought chain prompt text by the large language model, and obtaining the portrait features of the sample interactive object;

[0028] Among them, the multiple steps include in sequence: determining the positive sample recommended objects that the sample interaction object generates the target interaction behavior and the negative sample recommended objects that do not generate the target interaction behavior, comparing the differences between the conversion reasons of the positive sample recommended objects and the conversion reasons of the negative sample recommended objects, analyzing the preference conversion reasons of the sample interaction object when generating the target interaction behavior based on the differences, and predicting the type of recommended objects and the degree of preference preferred by the sample interaction object based on the preference conversion reasons.

[0029] In some exemplary embodiments, the reasoning process for generating the target interaction behavior for the sample recommendation object includes reasoning from multiple preset dimensions associated with the target interaction behavior.

[0030] In some exemplary embodiments, obtaining the object description information of the sample recommendation object includes:

[0031] Acquire information of multiple modalities of the sample recommendation object;

[0032] Based on one or more information of the multiple modalities, object description information of the sample recommendation object is determined.

[0033] According to a second aspect of an embodiment of the present disclosure, there is provided an object feature determination device, including:

[0034] An object description acquisition unit is configured to acquire object description information of any sample recommended object in the sample recommended object set;

[0035] The conversion analysis unit is configured to input the object description information into a large language model, and perform conversion analysis of a target interactive behavior through the large language model based on a first preset thought chain prompt text to obtain a conversion reason of the sample recommended object corresponding to the target interactive behavior; the first preset thought chain prompt text is used to prompt the reasoning process of generating the target interactive behavior for the sample recommended object; the conversion reason includes a reasoning result of at least one dimension in the reasoning process, and each reasoning result of the dimension represents the performance of the sample recommended object in generating the target interactive behavior in the dimension;

[0036] A first feature representation unit is configured to perform feature representation on the sample recommended object based on the conversion reason of the sample recommended object corresponding to the target interactive behavior, so as to obtain a first object feature of the sample recommended object;

[0037] The second feature representation unit is configured to execute the first object feature of each sample recommendation object in the sample recommendation object sequence of the sample interaction object within a historical time period to determine the second object feature of the sample interaction object; wherein the first object feature of the sample recommendation object and the second object feature of the sample interaction object are used to train a recommendation probability prediction model, and the recommendation probability prediction model is used to estimate the probability of generating the target interaction behavior.

[0038] In some exemplary embodiments, the first feature representation unit includes:

[0039] A keyword extraction unit is configured to input the conversion reason of the sample recommendation object corresponding to the target interaction behavior into the large language model, perform keyword extraction processing on the conversion reason according to a preset keyword extraction prompt through the large language model, and obtain a keyword extraction result corresponding to the sample recommendation object; the preset keyword extraction prompt is used to instruct to perform a keyword extraction task according to a keyword extraction example, and the keyword extraction example represents the output of keywords in a sample text according to a specified format;

[0040] The first object feature generating unit is configured to generate a first object feature of the sample recommended object based on a keyword extraction result corresponding to the sample recommended object.

[0041] In some exemplary embodiments, the first object feature generating unit includes:

[0042] A text encoding unit is configured to perform text encoding processing on the conversion reason of the sample recommendation object corresponding to the target interactive behavior to obtain a text encoding result corresponding to the sample recommendation object;

[0043] The feature splicing unit is configured to execute the keyword extraction result corresponding to the sample recommendation object as a sparse feature, and the text encoding result corresponding to the sample recommendation object as a dense feature, splicing the sparse feature and the dense feature to obtain the first object feature of the sample recommendation object.

[0044] In some exemplary embodiments, the apparatus further comprises:

[0045] A sequence acquisition unit is configured to execute acquisition of a sample recommendation object sequence of a sample interaction object in a historical time period, and determine a positive sample recommendation object and a negative sample recommendation object in the sample recommendation object sequence; the positive sample recommendation object refers to a sample recommendation object for which the sample interaction object generates the target interaction behavior, and the negative sample recommendation object refers to a sample recommendation object for which the sample interaction object does not generate the target interaction behavior;

[0046] A portrait feature generating unit is configured to determine the preference conversion reason of the sample interaction object when generating the target interaction behavior based on the conversion reason of the positive sample recommendation object corresponding to the target interaction behavior and the conversion reason of the negative sample recommendation object corresponding to the target interaction behavior, and generate a portrait feature of the sample interaction object based on the preference conversion reason of the sample interaction object;

[0047] Correspondingly, the second feature representation unit is specifically configured to perform splicing on the first object features of each sample recommended object in the sample recommended object sequence and the portrait features of the sample interactive object to obtain the second object features of the sample interactive object.

[0048] In some exemplary embodiments, the apparatus further comprises:

[0049] A preference reason aggregation unit, configured to execute the preference conversion reason based on each sample interaction object and determine a preference conversion reason set;

[0050] A correction unit is configured to execute, for each of the sample recommendation objects, a conversion reason corresponding to the target interactive behavior, and adjust the conversion reason based on the preferred conversion reason set to obtain an adjusted conversion reason corresponding to each of the sample recommendation objects; the adjusted conversion reason is a subset of the preferred conversion reason set;

[0051] The first feature representation unit is specifically configured to execute the step of using the adjusted conversion reason corresponding to each of the sample recommendation objects as the conversion reason of the sample recommendation object corresponding to the target interactive behavior, and executing the step of performing feature representation on the sample recommendation objects based on the conversion reason of the sample recommendation objects corresponding to the target interactive behavior to obtain the first object feature of the sample recommendation objects.

[0052] In some exemplary embodiments, the portrait feature generation unit is specifically configured to execute acquisition of a second preset thought chain prompt text; the second preset thought chain prompt text is used to prompt the sample interaction object to analyze the preference of the sample interaction object when generating the target interaction behavior according to multiple steps; the conversion reasons of the positive sample recommended object and the conversion reasons of the negative sample recommended object are input into the large language model, and the preference analysis of the sample interaction object is performed by the large language model based on the second preset thought chain prompt text to obtain the portrait feature of the sample interaction object;

[0053] Among them, the multiple steps include in sequence: determining the positive sample recommended objects that the sample interaction object generates the target interaction behavior and the negative sample recommended objects that do not generate the target interaction behavior, comparing the differences between the conversion reasons of the positive sample recommended objects and the conversion reasons of the negative sample recommended objects, analyzing the preference conversion reasons of the sample interaction object when generating the target interaction behavior based on the differences, and predicting the type of recommended objects and the degree of preference preferred by the sample interaction object based on the preference conversion reasons.

[0054] In some exemplary embodiments, the reasoning process for generating the target interaction behavior for the sample recommendation object includes reasoning from multiple preset dimensions associated with the target interaction behavior.

[0055] In some exemplary embodiments, the object description obtaining unit includes:

[0056] A multimodal information acquisition unit, configured to acquire information of multiple modalities of the sample recommendation object;

[0057] The object description determining unit is configured to determine the object description information of the sample recommendation object based on one or more information of the multiple modalities.

[0058] According to a third aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0059] processor;

[0060] a memory for storing instructions executable by the processor;

[0061] The processor is configured to execute the instructions to implement the object feature determination method of the first aspect mentioned above.

[0062] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the object feature determination method of the first aspect described above.

[0063] According to a fifth aspect of an embodiment of the present disclosure, there is provided a computer program product, comprising a computer program, which, when executed by a processor, implements the object feature determination method of the first aspect described above.

[0064] The disclosed embodiment inputs the object description information of any sample recommendation object into a large language model, and uses the large language model to perform a conversion analysis of the target interaction behavior based on a first preset thought chain prompt text to obtain a conversion reason of the sample recommendation object corresponding to the target interaction behavior, wherein the first preset thought chain prompt text is used to prompt the reasoning process of generating the target interaction behavior for the sample recommendation object, and the conversion reason includes the reasoning result of at least one dimension in the reasoning process, and the reasoning result of each dimension represents the performance of the sample recommendation object in generating the target interaction behavior on the dimension, and then based on the conversion reason of the sample recommendation object, the sample recommendation object is characterized to obtain the first object feature of the sample recommendation object, and based on the sample interaction object in the sample recommendation object sequence within the historical time period, each sample The first object feature of the recommended object is used to determine the second object feature of the sample interactive object. Since the above process is an inference process for generating the target interactive behavior, that is, it is oriented towards a specific estimation task, it is able to extract deep information with information gain and easy generalization, and is closely related to the final estimation task, and can remove a large amount of redundant content, thereby improving the accuracy of object representation for specific estimation tasks, and only based on the object description information of the sample recommended object to perform logical reasoning on the target interactive behavior, it imitates the paradigm of pre-trained Embedding without the need for training, saves a lot of time complexity, reduces the time consumption of object representation, and thereby improves the efficiency of using the first object feature and the second object feature to train the recommendation probability estimation model for estimating the probability of generating the target interactive behavior, as well as the estimation accuracy of the model.

[0065] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0067] Figure 1 is a schematic diagram of an application environment of a method for determining object features according to an exemplary embodiment;

[0068] Figure 2 is a flowchart of a method for determining object features according to an exemplary embodiment;

[0069] Figure 3 is a flowchart of another method for determining object features according to an exemplary embodiment;

[0070] Figure 4 is a flowchart of another method for determining object features according to an exemplary embodiment;

[0071] Figure 5 is a structural block diagram of a device for determining object features according to an exemplary embodiment;

[0072] Figure 6 The present invention is a structural block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0073] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0074] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0075] It should also be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data for analysis, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0076] See also Figure 1 , which is a schematic diagram of an application environment according to an exemplary embodiment, the application environment may include a terminal 110 and a server 120, and the terminal 110 and the server 120 may be connected via a wired network or a wireless network.

[0077] The terminal 110 may be a smart phone, a tablet computer, a laptop computer, a desktop computer, etc., but is not limited thereto. The terminal 110 may be installed with client software such as an application (Application, referred to as App) that provides interactive functions. The application may be an independent application or a subprogram in the application. The user of the terminal 110 may log in to the application using pre-registered user information, which may include an account number and a password. Exemplarily, the application may be a social application, a live broadcast application, a video application, etc.

[0078] The server 120 may be a server that provides background services for the application in the terminal 110. The background services may include recommendation services. The recommendation objects of the recommendation services may be based on items in the actual recommendation system, such as videos, commodities, pictures, news, music, etc. The server 120 may store a trained recommendation probability prediction model, which is used to predict the probability of generating a target interaction behavior. The target interaction behavior may characterize the interaction intention of the interaction object. When the interaction object generates the target interaction behavior for the recommendation object, it indicates that the recommendation object has achieved a conversion. The target interaction behavior may be any one of click, purchase, download, like, follow, long broadcast, etc., which may be determined based on actual needs.

[0079] In the disclosed embodiments, server 120 may be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, an edge computing node, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0080] Figure 2 is a flow chart of a method for determining object features according to an exemplary embodiment. Figure 2 As shown, the following steps are included:

[0081] In step S201, for any sample recommended object in the sample recommended object set, object description information of the sample recommended object is obtained.

[0082] The sample recommendation objects in the sample recommendation object set can be used to recommend to the interactive object so as to be exposed and displayed on the terminal of the interactive object. For example, the sample recommendation object can be an Internet product, such as video, commodity, picture, news, music, etc. The interactive object is an object for implementing interactive behaviors, and the interactive behaviors may include browsing, clicking, playing, liking, commenting, purchasing, etc.

[0083] The object description information of the sample recommendation object fully describes the sample recommendation object in text form. Exemplarily, in order to improve the accuracy of determining the object features, the object description information of the sample recommendation object should describe the sample recommendation object as comprehensively as possible. Therefore, the above step S201 may include: obtaining information of multiple modalities of the sample recommendation object; and determining the object description information of the sample recommendation object based on one or more information of the multiple modalities.

[0084] Specifically, multiple modalities may include text, image, voice, etc., wherein the information of the text modality may be the text content in the sample recommendation object, such as the title; the information of the image modality may be the image content in the sample recommendation object identified based on image recognition technology (such as OCR, Optical Character Recognition); the information of the voice modality may be the voice content in the sample recommendation object identified based on automatic speech recognition technology (ASR, Automatic Speech Recognition).

[0085] When determining the object description information of the sample recommendation object based on one or more of the information in the multiple modalities, the sample recommendation object may be described from texts of multiple description dimensions, and the multiple description dimensions may include modal dimensions and attribute dimensions. For example, the modal dimension may include text, voice, and image, and the attribute dimension may include categories, associated objects, etc. Taking the sample recommendation object as a short video advertisement as an example, the multiple description dimensions may include the video title, the content displayed in the video, the subtitles matched with the video, the voice matched with the video, the product recommended in the video, the category to which the product belongs, the price of the product, and the average price of products under the category to which the product belongs.

[0086] In step S203, the object description information of the sample recommended object is input into the large language model, and the large language model performs conversion analysis of the target interactive behavior based on the first preset thought chain prompt text to obtain the conversion reason of the sample recommended object.

[0087] The first preset thought chain prompt text is used to prompt the reasoning process of generating the target interactive behavior for the sample recommendation object. The conversion reason includes the reasoning result of at least one dimension in the reasoning process, and the reasoning result of each dimension represents the performance of the sample recommendation object in generating the target interactive behavior in the dimension.

[0088] The Large Language Model (LLM) is based on deep learning and neural network technology, especially the Transformer architecture. Its specific functions include natural language understanding, text generation, dialogue system, language translation, text summarization, etc. Through training on massive text data, it can predict contextual relationships and achieve semantic understanding and generation. The Chain-of-thought (COT) refers to a series of logically connected thinking paths formed by the model when generating output results.

[0089] Specifically, the first preset thought chain prompt text can be input into the large language model to stimulate the creativity and semantic understanding ability of the large language model, and guide the large language model to perform the conversion analysis task of the target interactive behavior, that is, to reason out the possible conversion reasons for realizing the target interactive behavior step by step, that is, the reasons that will cause the target interactive behavior to occur. The conversion reason can be, for example, a chain structure of "because a, so b; because b, so c". Therefore, the conversion reason of the sample recommended object corresponding to the target interactive behavior can also be called the thought chain of the sample recommended object corresponding to the target interactive behavior, so that the thought chain of the sample recommended object can be extracted for a specific estimation task (i.e., estimating the probability of generating the target interactive behavior). It can be understood that the conversion reason of the sample recommended object corresponding to the target interactive behavior output by the large language model may include content that is not in the object description information but is inferred by the large language model and is associated with the generation of the target interactive behavior.

[0090] In some examples, the first preset thought chain prompt text can be designed as a closed thought chain mode, that is, the reasoning process for generating the target interactive behavior for the sample recommendation object includes reasoning from multiple preset dimensions associated with the target interactive behavior, and then the conversion reason of the sample recommendation object can include the reasoning results of the multiple preset dimensions. Exemplarily, the multiple preset dimensions can be general thought chain dimensions for measuring the sample recommendation object in the recommendation task, such as the thought chain dimensions that are common to most recommendation tasks, such as effectiveness and superiority.

[0091] Taking the example that the sample recommendation object is an e-commerce short video advertisement and the target interactive behavior is purchase, the reasoning performed by the above-mentioned multiple preset dimensions can be questions of multiple different dimensions such as specified answer effectiveness, superiority, scope of application, influence, convenience, etc. In a specific implementation, the first preset thinking chain prompt text may include instruction text, which is used to prompt the large language model to perform a conversion analysis of the target interactive behavior step by step from multiple preset dimensions based on the object description information of the input sample recommendation object according to the thinking method of the thinking chain. For example, the instruction text may be "Please help me analyze how the following short video e-commerce advertisement proves to users that users should buy the products promoted in the advertisement? Please analyze from several aspects: 1. Problem classicity: Is the problem / pain point solved classic and popular? 2. Purchase interest stimulation method: What methods does the advertisement use to stimulate the user's purchase interest? 3. Effectiveness: How does the product effectively solve the specific needs of users? Does it solve problems that cannot be solved by traditional methods? 4. Price advantage: Is there a price advantage? 5. Convenience: Is the ease of use / convenience good (is there an operation demonstration? 6. Superiority: Clearly answer whether the product involves comparison with competing products? And which competing products are compared? Why is the product superior to competing products in terms of effectiveness and cost? If the product does not clearly demonstrate superiority by comparing it with competing products, explain how the subtext in the ad assumes that the product is significantly superior to competing products. 7. Consumption scenario adaptability: Are there specific times, holidays, consumption / use / enjoyment scenarios introduced to make users feel that the product is particularly suitable? 8. Scope of application: Is it applicable to a variety of situations and scenarios? 9. Adaptability: Is there good adaptability for other related products of existing users? 10. Influence: Is there a description of user reputation, sales volume, etc. 11. Brand: What short and memorable words are designed in the ad to impress users, such as slogans, repeated phrases, golden sentences, interesting phrases, etc. Let's think step by step, using common sense and knowledge from various disciplines. "

[0092] Then for a short video advertisement, a specific example of the conversion reason output by the large language model based on the conversion analysis of the purchase prompt text of the first preset thought chain can be "Based on the short video e-commerce advertisement provided, we can analyze from the following aspects:

[0093] 1. Classicity of the problem: Is the problem / pain point solved classic and popular? The main problems solved in this ad are dandruff removal, oil control, and improvement of scalp environment. This is a classic and popular problem because dandruff and oily hair are problems faced by many people.

[0094] 2. Purchase interest stimulation method: What methods did the advertisement use to stimulate the user's purchase interest? The advertisement used XXXX as a spokesperson, and XXXX's identity increased the trustworthiness of the product. In addition, the advertisement showed the actual effect of the product, and demonstrated how to use the shampoo and its effect through a video.

[0095] 3. Effectiveness: How does the product effectively solve the specific needs of users? Does it solve problems that traditional methods cannot solve? The advertisement emphasizes that the main ingredient of the product, selenium disulfide, deeply cleanses hair follicles and improves the scalp environment. This ingredient has a significant effect on improving dandruff, itchy scalp, and oily scalp. Compared with traditional shampoos, the product is effective and has a scientific basis.

[0096] 4. Price advantage: Is there a price advantage? The price of this product is Y yuan, which has a clear price advantage compared to the average price.

[0097] 5. Convenience: Is the usability / convenience good (is there a demonstration)? The advertisement shows a demonstration of the shampoo usage process, which makes it easier for users to understand how to use the product.

[0098] 6. Superiority: Clearly answer whether the product is compared with competing products? Which competing products are compared? Why is the product superior to competing products in terms of effectiveness and cost? The advertisement does not explicitly mention competing products, but by emphasizing the effectiveness and price advantage of the product, it implies that the product is superior to competing products.

[0099] 7. Consumption scenario suitability: Does the advertisement introduce specific time, holidays, consumption / use / enjoyment scenarios to make users feel that the product is particularly suitable. The advertisement does not mention specific time, holidays or consumption scenarios, but the shampoo and hair care products are suitable for daily use.

[0100] 8. Scope of application: Is it applicable to a variety of situations and scenarios? This product is suitable for people with dandruff, itchy scalp, oily scalp and other problems, and is suitable for a variety of situations.

[0101] 9. Adaptability: Whether the product is well adapted to other related products of existing users. The advertisement does not mention the adaptability of the product to other related products.

[0102] 10. Influence: Does it describe user reputation, sales, etc.? The ad mentions the authoritative efficacy test report of the product, which increases the trust of the product. However, it does not mention specific sales and users.

[0103] In the above implementation, since the reasoning process for generating the target interactive behavior for the sample recommendation object includes reasoning from multiple preset dimensions associated with the target interactive behavior, the cost-effectiveness of the demand can be comprehensively solved and the information gain of the output conversion reasons can be improved.

[0104] In other examples, the first preset thought chain prompt text can be designed as an open thought chain mode, that is, to give full play to the creativity of the large language model and openly generate conversion reasons for the sample recommendation object. In a specific implementation, the first preset thought chain prompt text may include instruction text, which is used to prompt the large language model to perform a conversion analysis of the target interactive behavior step by step according to the thinking mode of the thought chain based on the object description information of the input sample recommendation object. For example, the instruction text may be "Please help me analyze how the following short video e-commerce advertisement proves to users that users should buy the products promoted in the advertisement? Please explain how each point can ultimately lead to a purchase decision. Let us think step by step, using common sense and knowledge from various disciplines."

[0105] Then, for a short video advertisement, a specific example of the conversion reason output by the large language model based on the conversion analysis of the purchase prompt text of the first preset thought chain can be "This short video e-commerce advertisement demonstrates to users the reasons for buying the shampoo from the following aspects:

[0106] 1. XXXX endorsement: Through the recommendation of XXXX, the credibility of the advertisement is improved, making it easier for users to believe in the effectiveness of the product.

[0107] 2. Shampoo ingredients: The "selected golden ratio selenium disulfide" mentioned in the advertisement lets users know that the product contains effective ingredients, thereby increasing their willingness to buy.

[0108] 3. Deep cleansing of hair follicles: The advertisement emphasizes the function of "deep cleansing of hair follicles", making users believe that this shampoo can effectively improve the scalp environment.

[0109] 4. Oil control and anti-dandruff: The advertisement repeatedly emphasizes the "oil control and anti-dandruff" effect, making users believe that this shampoo can solve common problems such as oily scalp, dandruff, and itchy scalp.

[0110] 5. Domestic brands: By mentioning "domestic brands", the advertisement links the product with national pride, making users more willing to support domestic products.

[0111] 6. Demonstration of actual effects: The advertisement shows a lady using shampoo in the bathroom, allowing users to intuitively see the product effects.

[0112] 7. Price comparison: The advertisement compares the product price with similar products to show that the product is reasonably priced and has a high cost-effectiveness.

[0113] In practical applications, the reasoning process for generating the target interactive behavior for the sample recommendation object may also include assigning a reasoning role to the large language model, so as to prompt the large language model to capture the professional knowledge possessed by the reasoning role, and to perform reasoning based on the professional knowledge in the reasoning process to improve the professionalism and accuracy of the conversion analysis of the target interactive behavior. For example, the large language model may be assigned a reasoning role through a text guide such as "You are a professional short video e-commerce advertising analyst, proficient in sales, photography, short video production skills, marketing psychology, and logical reasoning."

[0114] In step S205, the sample recommended object is characterized based on the conversion reason of the sample recommended object corresponding to the target interactive behavior to obtain a first object feature of the sample recommended object.

[0115] In step S207, based on the first object feature of each sample recommended object in the sample recommended object sequence of the sample interactive object within the historical time period, the second object feature of the sample interactive object is determined.

[0116] The first object feature of the sample recommendation object and the second object feature of the sample interaction object are used to train a recommendation probability estimation model, and the recommendation probability estimation model is used to estimate the probability of generating the target interaction behavior.

[0117] The historical time period may be set based on actual needs, for example, the most recent week or the most recent month.

[0118] Specifically, each sample recommended object in the sample recommended object sequence may be represented by its corresponding first object feature, thereby obtaining a first object feature sequence, and the first object feature sequence is used as the second object feature of the sample interactive object.

[0119] Since the first preset thought chain prompt text of the embodiment of the present disclosure is used to prompt the reasoning process for generating the target interactive behavior for the sample recommended object, that is, it is oriented to a specific estimation task, and the output conversion reason includes the reasoning result of at least one dimension in the reasoning process, the reasoning result of each dimension represents the correlation between the sample recommended object and the target interactive behavior in the dimension, so that deep information with information gain and easy generalization can be extracted, and it is closely related to the final estimation task, a large amount of redundant content can be removed, and the accuracy of object representation is improved, and logical reasoning of the target interactive behavior is performed only based on the sample recommended object, imitating the paradigm of pre-trained Embedding without the need for training, saving a lot of time complexity, reducing the time consumption of object representation, and thereby improving the efficiency of using the first object feature and the second object feature to train the recommendation probability estimation model for estimating the probability of generating the target interactive behavior and the estimation accuracy of the model.

[0120] In some exemplary embodiments, in order to further improve the accuracy of object features to improve the training effect of the probability estimation model, the above step S205 may include:

[0121] The conversion reason of the sample recommendation object corresponding to the target interactive behavior is input into the large language model, and the conversion reason is subjected to keyword extraction processing by the large language model according to a preset keyword extraction prompt to obtain a keyword extraction result corresponding to the sample recommendation object; the preset keyword extraction prompt is used to instruct to perform a keyword extraction task according to a keyword extraction example, and the keyword extraction example represents the output of keywords in a sample text according to a specified format;

[0122] Based on the keyword extraction result corresponding to the sample recommended object, a first object feature of the sample recommended object is generated.

[0123] Specifically, based on the summarization function of the large language model and the preset keyword extraction prompt of the embodiment of the present disclosure, the large language model can output the corresponding keyword extraction result based on the keyword extraction example in the preset keyword extraction prompt, so as to further filter out redundant information and retain the key information related to generating the target interactive behavior, so that the first object feature of the sample recommendation object generated based on the keyword extraction result and the second object feature based on the first object feature better match the specific estimation task of the recommendation probability estimation model (i.e., estimating the probability of generating the target interactive behavior), which is conducive to improving the training effect of the recommendation probability estimation model trained based on the first object feature and the second object feature for the specific estimation task.

[0124] In some exemplary embodiments, in order to further improve the accuracy of the object representation for a specific estimation task so as to improve the training effect of the recommendation probability estimation model for the specific estimation task, when generating the first object feature of the sample recommendation object based on the keyword extraction result corresponding to the sample recommendation object, the following may be included:

[0125] Performing text encoding processing on the conversion reasons of the sample recommendation object corresponding to the target interactive behavior to obtain a text encoding result corresponding to the sample recommendation object;

[0126] The keyword extraction result corresponding to the sample recommendation object is used as a sparse feature, and the text encoding result corresponding to the sample recommendation object is used as a dense feature, and the sparse feature and the dense feature are concatenated to obtain a first object feature of the sample recommendation object.

[0127] Specifically, a pre-trained language representation model such as a BERT model can be used to perform the above-mentioned text encoding processing, that is, the conversion reason of the sample recommendation object corresponding to the target interaction behavior is input into the BERT model for text encoding processing to obtain the text encoding result output by the BERT model.

[0128] After obtaining the keyword extraction results and text encoding results of the sample recommendation object respectively, the accuracy of the first object feature can be further improved by taking the keyword extraction results as sparse features and the text encoding results as dense features and splicing them, and using the spliced ​​result as the first object feature of the sample recommendation object.

[0129] In some exemplary embodiments, Figure 3 A flow chart of another object feature determination method is provided. In step S203, the object description information is input into a large language model. After the large language model performs a conversion analysis of the target interactive behavior based on a first preset thought chain prompt text and obtains the conversion reason of the sample recommended object, the method may further include:

[0130] In step S301, a sample recommended object sequence of a sample interactive object in a historical time period is obtained, and positive sample recommended objects and negative sample recommended objects in the sample recommended object sequence are determined.

[0131] The sample interaction object may be any sample interaction object in the sample interaction object set. The sample recommendation object sequence of the sample interaction object in the historical time period refers to the sample recommendation objects recommended to the sample interaction object in the historical time period, such as the sample recommendation objects exposed in the terminal of the sample interaction object. The sample interaction object may perform interactive behaviors such as browsing, clicking, purchasing, downloading, etc. on the exposed sample recommendation objects.

[0132] Among them, determining the positive sample recommended objects and negative sample recommended objects in the sample recommended object sequence may be to generate corresponding label information for each sample recommended object, and the label information represents whether the sample interactive object generates a target interactive behavior for the corresponding sample recommended object. Exemplarily, the label information may be "1" or "0", where "1" indicates that the sample interactive object generates a target interactive behavior for the sample recommended object, and "0" indicates that the sample interactive object does not generate a target interactive behavior for the sample recommended object. Taking the sample recommended object as a short video advertisement and the target interactive behavior as a purchase as an example, if the sample interactive object purchases the product associated with the short video advertisement based on the exposed short video advertisement, the label information of the short video advertisement is "1", and conversely, if the sample interactive object does not purchase the product associated with the short video advertisement based on the exposed short video advertisement, the label information of the short video advertisement is "0".

[0133] Among them, the positive sample recommended object in the sample recommended object sequence refers to the sample recommended object whose sample interaction object generates the target interaction behavior, such as the sample recommended object whose label information is "1"; the negative sample recommended object in the sample recommended object sequence refers to the sample recommended object whose sample interaction object does not generate the target interaction behavior, such as the sample recommended object whose label information is "0".

[0134] In step S303, based on the conversion reasons of the positive sample recommended object corresponding to the target interactive behavior and the conversion reasons of the negative sample recommended object corresponding to the target interactive behavior, the preferred conversion reasons of the sample interactive object when generating the target interactive behavior are determined, and the portrait features of the sample interactive object are generated based on the preferred conversion reasons of the sample interactive object.

[0135] Specifically, the preference conversion reason of the sample interaction object when generating the target interaction behavior may be the conversion reason of the positive sample recommendation object corresponding to the target interaction behavior but not the conversion reason of the negative sample recommendation object corresponding to the target interaction behavior. For example, the sample recommendation object sequence of the sample interaction object User1 is {(item1,1), (item2,0), (item3,0)}, then item1 is the positive sample recommendation object, item2 and item3 are the negative sample recommendation objects, assuming that the conversion reason of item1 corresponding to the target interaction behavior is (R1,R2), the conversion reason of item2 corresponding to the target interaction behavior is (R2,R3), and the conversion reason of item3 corresponding to the target interaction behavior is (R4,R5), then the preference conversion reason of the sample interaction object User1 when generating the target interaction behavior is R1.

[0136] Among them, the portrait features of the sample interaction object may include its preference conversion reasons, and may also include other information associated with its preference conversion reasons, such as the recommended object type associated with its preference conversion reasons.

[0137] Based on this, the aforementioned step S207 can be implemented as Figure 3 In step S305 shown in the figure, the first object features of each sample recommended object in the sample recommended object sequence and the portrait features of the sample interactive object are spliced ​​to obtain the second object features of the sample interactive object.

[0138] Specifically, each sample recommended object in the sample recommended object sequence of the sample interactive object is represented by its corresponding first object feature, and then concatenated with the portrait feature of the sample interactive object, so as to obtain the second object feature of the sample interactive object.

[0139] The above-mentioned implementation method determines the preference conversion reason of the sample interaction object when generating the target interaction behavior based on the conversion reasons of the positive and negative sample recommendation objects in the sample recommendation object sequence of the sample interaction object, and generates a portrait feature of the sample interaction object based on this, and then combines the portrait feature and the first object feature of each sample recommendation object in the sample recommendation object sequence to obtain the second object feature of the sample interaction object, thereby improving the accuracy of the sample interaction object representation for specific estimation tasks, and thus is conducive to improving the training effect of the recommendation probability estimation model for specific estimation tasks.

[0140] In some exemplary implementations, in order to further save the time spent on training the probability estimation model and improve the model training efficiency, the reasoning capability of the large language model may be used to implement the aforementioned step S303. Specifically, the implementation of step S303 may include:

[0141] Obtain a second preset thought chain prompt text, the second preset thought chain prompt text is used to prompt to analyze the preference of the sample interaction object when generating the target interaction behavior according to multiple steps, the multiple steps sequentially include: determining the positive sample recommendation object that the sample interaction object generates the target interaction behavior and the negative sample recommendation object that does not generate the target interaction behavior, comparing the difference between the conversion reasons of the positive sample recommendation object and the conversion reasons of the negative sample recommendation object, analyzing the preference conversion reasons of the sample interaction object when generating the target interaction behavior based on the difference, and predicting the type of recommendation object and the degree of preference preferred by the sample interaction object based on the preference conversion reasons;

[0142] The conversion reasons of the positive sample recommended object corresponding to the target interactive behavior and the conversion reasons of the negative sample recommended object corresponding to the target interactive behavior are input into the large language model, and the preference analysis of the sample interactive object is performed by the large language model based on the second preset thinking chain prompt text to obtain the portrait features of the sample interactive object.

[0143] Specifically, when the large language model performs preference analysis of the sample interaction object based on the second preset thought chain prompt text, it can reason step by step according to the above-mentioned multiple steps and output the reasoning results of each step, so that the reasoning results of the multiple steps can be used as the portrait features of the sample interaction object.

[0144] The above implementation method allows LLM to comprehensively judge why the conversion reasons of the positive sample recommendation object corresponding to the target interactive behavior are more reasonable, and allows LLM to summarize the thought chain that the user is more likely to generate conversion. This can not only improve the accuracy of reasoning and make the portrait features of the sample interactive objects more accurate, but also is equivalent to model training based on pre-trained embeddings without the need for training, saving a lot of time complexity, which is conducive to improving the training efficiency and training effect of the probability estimation model.

[0145] Still taking the example of the sample recommendation object being a short video advertisement and the target interactive behavior being purchase, the second preset thinking chain prompt text can be "Please help me analyze the following short video e-commerce advertisements. For a specific user A, he purchased the product in advertisement 1 but not the product in advertisement 2. Please first clearly answer which products in the advertisements the user purchased and which products in the advertisements he did not purchase. Please compare the differences between the advertisements with and without purchases, and rethink from a higher level to analyze which argumentation methods and arguments the user is more likely to accept in making a purchasing decision? Finally, please guess the set of brand archetypes preferred by the user and the probability of each archetype (list concisely). Let us think step by step, using common sense and world knowledge."

[0146] In a specific implementation, the second preset thought chain prompt text may also include assigning a reasoning role to the large language model, thereby prompting the large language model to capture the professional knowledge possessed by the reasoning role, and to perform reasoning in the preference analysis based on the professional knowledge, so as to improve the professionalism and accuracy of the analysis of the preferences of the sample interaction objects when generating the target interaction behavior. For example, the second preset thought chain prompt text may guide the large language model to assign a reasoning role through the following text: "You are a professional short video e-commerce advertising analyst, proficient in sales, photography, short video production skills, marketing psychology, and logical reasoning."

[0147] Exemplarily, the second preset thought chain prompt text may also include multiple preset recommended object types, thereby guiding the large language model to predict the recommended object type preferred by the sample interaction object based on the preference conversion reason, from the multiple preset recommended object types. The degree of preference may be represented by probability.

[0148] In some exemplary embodiments, in order to further improve the accuracy of object representation and the training efficiency of the probability estimation model, Figure 4 The flowchart of another method for determining object features is shown in FIG. 1 . In the aforementioned step S205, before the sample recommended object is characterized based on the conversion reason of the sample recommended object corresponding to the target interactive behavior and the first object feature of the sample recommended object is obtained, the method may further include:

[0149] In step S401, a preference conversion reason set is determined based on the preference conversion reasons of each sample interaction object.

[0150] Specifically, based on the aforementioned steps S301 to S303, the preference conversion reasons of each sample interaction object in the sample interaction object set can be obtained, and the preference conversion reasons of each sample interaction object can be aggregated to obtain a preference conversion reason set. For example, the preference conversion reason of the sample interaction object User1 when generating the target interaction behavior is {R1}, the preference conversion reason of the sample interaction object User2 when generating the target interaction behavior is {R2, R3}, and the preference conversion reason of the sample interaction object User3 when generating the target interaction behavior is {R2, R4}, then the preference conversion reason set is {R1, R2, R3, R4}, assuming that the conversion reason of the sample recommendation object item1 is {R1, R3, R5}, then the adjusted conversion reason of the sample recommendation object item1 is {R1, R3}, that is, the conversion reasons of the sample recommendation object item1 that do not belong to the preference conversion reason set are eliminated, so that the remaining is the adjusted conversion reason of the sample recommendation object item1.

[0151] In step S403, for each conversion reason of the sample recommendation object corresponding to the target interactive behavior, the conversion reason is adjusted based on the preferred conversion reason set to obtain an adjusted conversion reason corresponding to each sample recommendation object.

[0152] Among them, the adjusted conversion reason corresponding to each sample recommendation object is a subset of the preferred conversion reason set, and then when executing the aforementioned step S205, the adjusted conversion reason corresponding to each sample recommendation object is used as the conversion reason of the sample recommendation object corresponding to the target interactive behavior, and the step of performing feature representation on the sample recommendation object based on the conversion reason of the sample recommendation object corresponding to the target interactive behavior is executed to obtain the first object feature of the sample recommendation object.

[0153] The above implementation method adjusts the conversion reasons of each sample recommendation object corresponding to the target interactive behavior by combining the preferred conversion reason set, thereby improving the fit between the adjusted conversion reasons and the target interactive behavior, and improving the accuracy of object representation for specific estimation tasks, which is conducive to improving the accuracy of the recommendation probability estimation model for specific estimation tasks.

[0154] In terms of depth, the disclosed embodiments can extract the transformation reasons from specific multimodal information to display, and then to the deep transformation reasons behind it, and step by step the information becomes more essential, stable, and generalizable; in terms of breadth, such a chain of thinking can mine various rational, irrational, and subtextual information in the text, which is a big step forward in breadth, and these reasons are all covered by the three major systems of the user's mind: symbols, meanings, and value systems. The disclosed embodiments are applicable to any "out of nothing" recommendation system. In addition, the disclosed embodiments can save a lot of time complexity for training recommendation probability estimation models for specific estimation tasks.

[0155] Figure 5 is a structural block diagram of an object feature determination device according to an exemplary embodiment. Figure 5 , the object feature determination device 500 comprises:

[0156] The object description acquisition unit 510 is configured to acquire object description information of any sample recommended object in the sample recommended object set;

[0157] The conversion analysis unit 520 is configured to input the object description information into the large language model, and perform conversion analysis of the target interactive behavior through the large language model based on the first preset thought chain prompt text to obtain the conversion reason of the sample recommended object corresponding to the target interactive behavior; the first preset thought chain prompt text is used to prompt the reasoning process of generating the target interactive behavior for the sample recommended object; the conversion reason includes the reasoning result of at least one dimension in the reasoning process, and each reasoning result of the dimension represents the performance of the sample recommended object in generating the target interactive behavior in the dimension;

[0158] A first feature representation unit 530 is configured to perform feature representation on the sample recommended object based on the conversion reason of the sample recommended object corresponding to the target interactive behavior, so as to obtain a first object feature of the sample recommended object;

[0159] The second feature representation unit 540 is configured to execute the first object feature of each sample recommendation object in the sample recommendation object sequence of the sample interaction object within a historical time period to determine the second object feature of the sample interaction object; wherein the first object feature of the sample recommendation object and the second object feature of the sample interaction object are used to train a recommendation probability prediction model, and the recommendation probability prediction model is used to estimate the probability of generating the target interaction behavior.

[0160] In some exemplary embodiments, the first feature representation unit 530 includes:

[0161] A keyword extraction unit is configured to input the conversion reason of the sample recommendation object corresponding to the target interactive behavior into the large language model, perform keyword extraction processing on the conversion reason according to a preset keyword extraction prompt through the large language model, and obtain a keyword extraction result corresponding to the sample recommendation object; the preset keyword extraction prompt is used to indicate that the keyword extraction task is performed according to a keyword extraction example, and the keyword extraction example represents the output of keywords in the sample text according to a specified format;

[0162] The first object feature generating unit is configured to generate a first object feature of the sample recommended object based on a keyword extraction result corresponding to the sample recommended object.

[0163] In some exemplary embodiments, the first object feature generating unit includes:

[0164] A text encoding unit is configured to perform text encoding processing on the conversion reason of the sample recommendation object corresponding to the target interactive behavior to obtain a text encoding result corresponding to the sample recommendation object;

[0165] The feature splicing unit is configured to execute the keyword extraction result corresponding to the sample recommendation object as a sparse feature, and the text encoding result corresponding to the sample recommendation object as a dense feature, splicing the sparse feature and the dense feature to obtain the first object feature of the sample recommendation object.

[0166] In some exemplary embodiments, the apparatus 500 further includes:

[0167] A sequence acquisition unit is configured to execute acquisition of a sample recommendation object sequence of a sample interaction object in a historical time period, and determine a positive sample recommendation object and a negative sample recommendation object in the sample recommendation object sequence; the positive sample recommendation object refers to a sample recommendation object for which the sample interaction object generates the target interaction behavior, and the negative sample recommendation object refers to a sample recommendation object for which the sample interaction object does not generate the target interaction behavior;

[0168] A portrait feature generating unit is configured to determine the preferred conversion reason of the sample interactive object when generating the target interactive behavior based on the conversion reason of the positive sample recommended object corresponding to the target interactive behavior and the conversion reason of the negative sample recommended object corresponding to the target interactive behavior, and generate a portrait feature of the sample interactive object based on the preferred conversion reason of the sample interactive object;

[0169] Correspondingly, the second feature representation unit 540 is specifically configured to perform splicing of the first object feature of each sample recommended object in the sample recommended object sequence and the portrait feature of the sample interactive object to obtain the second object feature of the sample interactive object.

[0170] In some exemplary embodiments, the apparatus 500 further includes:

[0171] A preference reason aggregation unit is configured to execute the preference conversion reason based on each sample interaction object and determine a preference conversion reason set;

[0172] A correction unit is configured to execute, for each conversion reason of the sample recommendation object corresponding to the target interactive behavior, adjusting the conversion reason based on the preferred conversion reason set, to obtain an adjusted conversion reason corresponding to each sample recommendation object; the adjusted conversion reason is a subset of the preferred conversion reason set;

[0173] The first feature representation unit 530 is specifically configured to execute the step of using the adjusted conversion reason corresponding to each of the sample recommendation objects as the conversion reason of the sample recommendation object corresponding to the target interactive behavior, and performing feature representation on the sample recommendation object based on the conversion reason of the sample recommendation object corresponding to the target interactive behavior to obtain the first object feature of the sample recommendation object.

[0174] In some exemplary embodiments, the portrait feature generation unit is specifically configured to execute acquisition of a second preset thought chain prompt text; the second preset thought chain prompt text is used to prompt the sample interaction object to analyze the preference when generating the target interaction behavior according to multiple steps; the conversion reason of the positive sample recommendation object corresponding to the target interaction behavior and the conversion reason of the negative sample recommendation object corresponding to the target interaction behavior are input into the large language model, and the preference analysis of the sample interaction object is performed based on the second preset thought chain prompt text by the large language model to obtain the portrait feature of the sample interaction object;

[0175] Among them, the multiple steps include in sequence: determining the positive sample recommendation object that the sample interaction object generates the target interaction behavior and the negative sample recommendation object that does not generate the target interaction behavior, comparing the difference between the conversion reason of the positive sample recommendation object and the conversion reason of the negative sample recommendation object, analyzing the preference conversion reason of the sample interaction object when generating the target interaction behavior based on the difference, and predicting the type of recommendation object and the degree of preference preferred by the sample interaction object based on the preference conversion reason.

[0176] In some exemplary embodiments, the reasoning process for generating the target interactive behavior for the sample recommendation object includes reasoning from multiple preset dimensions associated with the target interactive behavior.

[0177] In some exemplary implementations, the object description acquisition unit 510 includes:

[0178] A multimodal information acquisition unit, configured to acquire information of multiple modalities of the sample recommendation object;

[0179] The object description determining unit is configured to determine the object description information of the sample recommendation object based on one or more information of the multiple modalities.

[0180] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0181] In an exemplary embodiment, an electronic device is also provided, including a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to implement the object feature determination method provided in any of the above embodiments when executing the instructions stored in the memory.

[0182] The electronic device may be a terminal, a server or a similar computing device. For example, the electronic device is a server. Figure 6 is a block diagram of an electronic device for executing an object feature determination method according to an exemplary embodiment. Figure 6 As shown, the server 600 may have relatively large differences due to different configurations or performances, and may include one or more central processing units (CPU) 610 (the processor 610 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 630 for storing data, and one or more storage media 620 (such as one or more mass storage devices) for storing application programs 623 or data 622. Among them, the memory 630 and the storage medium 620 can be short-term storage or permanent storage. The program stored in the storage medium 620 may include one or more modules, each of which may include a series of instruction operations on the server. Furthermore, the central processing unit 610 can be configured to communicate with the storage medium 620 and execute a series of instruction operations in the storage medium 620 on the server 600. The server 600 may also include one or more power supplies 660, one or more wired or wireless network interfaces 650, one or more input and output interfaces 640, and / or one or more operating systems 621, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0183] The input / output interface 640 may be used to receive or send data via a network. The specific example of the network may include a wireless network provided by a communication provider of the server 600. In one example, the input / output interface 640 includes a network adapter (Network Interface Controller, NIC), which may be connected to other network devices via a base station so as to communicate with the Internet. In one example, the input / output interface 640 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0184] It can be understood by those skilled in the art that Figure 6 The structure shown is for illustration only and does not limit the structure of the above electronic device. Figure 6 More or fewer components as shown, or with Figure 6 Different configurations shown.

[0185] In an exemplary embodiment, a computer-readable storage medium is also provided. When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device can execute the object feature determination method provided in any of the above embodiments.

[0186] In an exemplary embodiment, a computer program product is also provided, including a computer program, and when the computer program is executed by a processor, the object feature determination method provided in any of the above embodiments is implemented.

[0187] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0188] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method for determining object features, characterized in that: include: For any sample recommendation object in the sample recommendation object set, obtaining object description information of the sample recommendation object; The object description information is input into a large language model, and a conversion analysis of the target interactive behavior is performed by the large language model based on a first preset thought chain prompt text to obtain a conversion reason of the sample recommended object corresponding to the target interactive behavior; the first preset thought chain prompt text is used to prompt the reasoning process of generating the target interactive behavior for the sample recommended object; The conversion reason includes the reasoning result of at least one dimension in the reasoning process, and the reasoning result of each dimension represents the performance of the sample recommendation object in the dimension in generating the target interactive behavior; Based on the conversion reason of the sample recommended object corresponding to the target interactive behavior, the sample recommended object is characterized to obtain a first object feature of the sample recommended object; Based on the first object features of each sample recommended object in the sample recommendation object sequence of the sample interaction object within a historical time period, determine the second object features of the sample interaction object; wherein the first object features of the sample recommendation object and the second object features of the sample interaction object are used to train a recommendation probability prediction model, and the recommendation probability prediction model is used to estimate the probability of generating the target interaction behavior.

2. The method according to claim 1, characterized in that The feature representation of the sample recommended object based on the conversion reason of the sample recommended object corresponding to the target interactive behavior to obtain the first object feature of the sample recommended object includes: The conversion reason of the sample recommendation object corresponding to the target interactive behavior is input into the large language model, and the conversion reason is subjected to keyword extraction processing by the large language model according to a preset keyword extraction prompt to obtain a keyword extraction result corresponding to the sample recommendation object; the preset keyword extraction prompt is used to instruct to perform a keyword extraction task according to a keyword extraction example, and the keyword extraction example represents the output of keywords in a sample text according to a specified format; Based on the keyword extraction result corresponding to the sample recommended object, a first object feature of the sample recommended object is generated.

3. The method according to claim 2, characterized in that The step of generating the first object feature of the sample recommendation object based on the keyword extraction result corresponding to the sample recommendation object includes: Performing text encoding processing on the conversion reasons of the sample recommendation object corresponding to the target interactive behavior to obtain a text encoding result corresponding to the sample recommendation object; The keyword extraction result corresponding to the sample recommendation object is used as a sparse feature, and the text encoding result corresponding to the sample recommendation object is used as a dense feature, and the sparse feature and the dense feature are concatenated to obtain a first object feature of the sample recommendation object.

4. The method according to claim 1, characterized in that: After inputting the object description information into the large language model, performing conversion analysis of the target interactive behavior based on the first preset thought chain prompt text by the large language model, and obtaining the conversion reason of the sample recommended object, the method further includes: Obtain a sample recommendation object sequence of the sample interaction object in a historical time period, and determine a positive sample recommendation object and a negative sample recommendation object in the sample recommendation object sequence; the positive sample recommendation object refers to a sample recommendation object for which the sample interaction object generates the target interaction behavior, and the negative sample recommendation object refers to a sample recommendation object for which the sample interaction object does not generate the target interaction behavior; Based on the conversion reasons of the positive sample recommended object corresponding to the target interactive behavior and the conversion reasons of the negative sample recommended object corresponding to the target interactive behavior, determine the preference conversion reasons of the sample interactive object when generating the target interactive behavior, and generate the portrait features of the sample interactive object based on the preference conversion reasons of the sample interactive object; The determining, based on the first object feature of each sample recommended object in a sample recommended object sequence of the sample interactive object within a historical time period, the second object feature of the sample interactive object comprises: The first object feature of each sample recommended object in the sample recommended object sequence and the portrait feature of the sample interactive object are concatenated to obtain the second object feature of the sample interactive object.

5. The method according to claim 4, characterized in that Before performing feature representation on the sample recommendation object based on the conversion reason of the sample recommendation object corresponding to the target interactive behavior to obtain the first object feature of the sample recommendation object, the method further includes: Determining a set of preference conversion reasons based on the preference conversion reasons of each sample interaction object; For each conversion reason of the sample recommendation object corresponding to the target interactive behavior, the conversion reason is adjusted based on the preferred conversion reason set to obtain an adjusted conversion reason corresponding to each sample recommendation object; the adjusted conversion reason is a subset of the preferred conversion reason set; The adjusted conversion reason corresponding to each of the sample recommended objects is used as the conversion reason of the sample recommended object corresponding to the target interactive behavior, and the step of performing feature representation on the sample recommended objects based on the conversion reason of the sample recommended objects corresponding to the target interactive behavior to obtain the first object feature of the sample recommended objects is performed.

6. The method according to claim 4, characterized in that The method of determining the preferred conversion reason of the sample interaction object when generating the target interaction behavior based on the conversion reason of the positive sample recommended object corresponding to the target interaction behavior and the conversion reason of the negative sample recommended object corresponding to the target interaction behavior, and generating the portrait feature of the sample interaction object based on the preferred conversion reason of the sample interaction object includes: Acquire a second preset thought chain prompt text; the second preset thought chain prompt text is used to prompt to analyze the preference of the sample interaction object when generating the target interaction behavior according to multiple steps; Inputting the conversion reasons of the positive sample recommended object corresponding to the target interactive behavior and the conversion reasons of the negative sample recommended object corresponding to the target interactive behavior into the large language model, performing preference analysis of the sample interactive object based on the second preset thought chain prompt text by the large language model, and obtaining the portrait features of the sample interactive object; Among them, the multiple steps include in sequence: determining the positive sample recommended objects that the sample interaction object generates the target interaction behavior and the negative sample recommended objects that do not generate the target interaction behavior, comparing the differences between the conversion reasons of the positive sample recommended objects and the conversion reasons of the negative sample recommended objects, analyzing the preference conversion reasons of the sample interaction object when generating the target interaction behavior based on the differences, and predicting the type of recommended objects and the degree of preference preferred by the sample interaction object based on the preference conversion reasons.

7. The method according to claim 1, characterized in that The reasoning process of generating the target interaction behavior for the sample recommendation object includes reasoning from multiple preset dimensions associated with the target interaction behavior.

8. The method according to claim 1, characterized in that The obtaining of the object description information of the sample recommendation object comprises: Acquire information of multiple modalities of the sample recommendation object; Based on one or more information of the multiple modalities, object description information of the sample recommendation object is determined.

9. An object feature determination device, characterized in that: include: An object description acquisition unit is configured to acquire object description information of any sample recommended object in the sample recommended object set; A conversion analysis unit is configured to input the object description information into a large language model, and perform conversion analysis of a target interactive behavior through the large language model based on a first preset thought chain prompt text to obtain a conversion reason of the sample recommended object corresponding to the target interactive behavior; the first preset thought chain prompt text is used to prompt the reasoning process of generating the target interactive behavior for the sample recommended object; The conversion reason includes the reasoning result of at least one dimension in the reasoning process, and the reasoning result of each dimension represents the performance of the sample recommendation object in the dimension in generating the target interactive behavior; A first feature representation unit is configured to perform feature representation on the sample recommended object based on the conversion reason of the sample recommended object corresponding to the target interactive behavior, so as to obtain a first object feature of the sample recommended object; The second feature representation unit is configured to execute the first object feature of each sample recommendation object in the sample recommendation object sequence of the sample interaction object within a historical time period to determine the second object feature of the sample interaction object; wherein the first object feature of the sample recommendation object and the second object feature of the sample interaction object are used to train a recommendation probability prediction model, and the recommendation probability prediction model is used to estimate the probability of generating the target interaction behavior.

10. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the object feature determination method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the object feature determination method according to any one of claims 1 to 8.

12. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the object feature determination method according to any one of claims 1 to 8 is implemented.

Images