Personalized content recommendation method and system based on large model and electronic equipment

Abstract

The embodiment of the invention relates to the technical field of artificial intelligence, and provides a personalized content recommendation method and system based on a large model and electronic device.The method comprises the steps that in response to a received content query instruction, at least one large language model is used for generating a question text corresponding to the content query instruction; determining a corresponding text block according to the similarity between each question text and the text block in the content library; based on a first prompt word generated according to each question text and the corresponding text block, using at least one large language model to generate a recommendation result corresponding to the question text; and outputting a recommendation result. On the basis of the scheme, the to-be-recommended content can be subjected to text division, the multiple problem texts are generated according to information such as user requirements, progressive retrieval is performed on the text library through the multiple problem texts, then the recommendation result meeting the personalized requirements of the user is generated on the basis of the retrieval result, the adaptability of the recommended content to the user is improved, and the user experience is improved. And the personalized requirements of the user are met.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of artificial intelligence technology, and in particular to a large-model-based personalized content recommendation method, system, and electronic device. Background Art

[0002] With the rapid development of mobile communications technology, telecom operators are offering an increasingly diverse range of packages and services, encompassing a variety of combinations including voice, data, and value-added services. Faced with this complex system of packages and services, users struggle to quickly select the one they need. Traditional recommendation methods also struggle to meet user needs, leading to inefficient selection and even incurring additional costs.

[0003] Common content recommendation methods can generate recommendation results similar to the packages / services in the user's historical usage records based on the user's historical usage records, or determine the corresponding recommendation results based on the similarity of user attributes in the user group. They can also use user characteristics, such as social attributes and service preferences, to obtain the user's adaptability to different packages / services to determine the recommendation results.

[0004] However, conventional content recommendation methods lack analysis of dynamic features such as user behavior trends and changes in user preferences, and insufficient analysis of user personalized needs. The recommendation results lack long-term adaptability and cannot meet users' personalized needs. Summary of the Invention

[0005] The embodiments of the present application provide a large-model-based personalized content recommendation method, system, and electronic device, which can improve the adaptability of recommendation results to users to meet the personalized needs of users by splitting and matching content and user needs.

[0006] To achieve the above objectives, the present invention adopts the following technical solutions:

[0007] In a first aspect, an embodiment of the present application provides a personalized content recommendation method based on a large model, comprising: in response to receiving a content query instruction, using at least one large language model to generate multiple question texts corresponding to the content query instruction; determining the text block corresponding to each question text based on the similarity between each question text and a text block in a content library, the content library including text blocks corresponding to the content to be recommended; generating a first prompt word based on each question text and the text block corresponding to the question text; based on the first prompt word, using at least one large language model to generate a recommendation result corresponding to the question text; and outputting the recommendation result.

[0008] Based on this solution, the recommended content can be divided into texts, and when the user searches, a large language model can be used to generate multiple question texts based on information such as user needs. The generated multiple question texts can be used to perform progressive searches on the divided text blocks, and then based on the search results, recommendation results that meet the user's personalized needs can be recursively generated to improve the adaptability of the recommended content to the user and meet the user's personalized needs.

[0009] In one possible embodiment, before generating a plurality of question texts corresponding to the content query instruction using at least one large language model in response to receiving the content query instruction, the method further includes: obtaining a content library, the content library also including text corresponding to the content to be recommended; dividing the text corresponding to the content to be recommended into at least one first text block that meets a preset text length; determining at least one second text block in each first text block based on the semantics of the text in the first text block; and storing the first text block and the second text block in the content library.

[0010] Based on this solution, the text in the content library can be divided into multiple granularities to achieve a more refined text content matching process, improve the accuracy of information retrieval for different needs during the text retrieval process, and improve the coherence of the final text content context.

[0011] In another possible implementation, the text block corresponding to each question text is determined based on the similarity between each question text and a text block in a content library, including: converting the question text into a question vector; obtaining the vector similarity between the question vector and each second text block in the content library; and determining the first text block corresponding to the second text block whose vector similarity is greater than a preset threshold as the text block corresponding to the question text.

[0012] Based on this solution, it is possible to determine text blocks with relatively complete content and high coherence based on segmented text blocks, so as to avoid missing content during the retrieval process and increase the relevance between the retrieval results and the question text.

[0013] In another possible implementation, storing the first text block and the second text block in a content library includes: determining a first mapping relationship between the second text block and the first text block based on acquisition sources of the first text block and the second text block, and determining a second mapping relationship between the first text block and the content to be recommended; and storing the text content, the first mapping relationship, and the second mapping relationship corresponding to the first text block and the second text block in the content library.

[0014] Based on this solution, the text content in the content library can be divided into multiple granularities to generate two text blocks with different amounts of content. In the retrieval process, the text blocks with higher precision can be matched with text blocks with higher coherence and completeness, thereby improving the effect of text retrieval and increasing the relevance between the final text blocks and the question text.

[0015] In another possible implementation, storing the first text block and the second text block in a content library includes: converting each second text block into a corresponding text vector; determining a third mapping relationship between the text vector and the first text block based on the acquisition sources of the first text block and the second text block, and determining a second mapping relationship between the first text block and the content to be recommended; and storing the text content, text vector, second mapping relationship, and third mapping relationship corresponding to the first text block in the content library.

[0016] Based on this solution, the vector corresponding to the second text block can be determined to optimize the calculation process in the similarity acquisition process and improve the efficiency of text similarity retrieval.

[0017] In another possible implementation, at least one large language model is used to generate multiple question texts corresponding to content query instructions, including: obtaining user history information, where the user history information is used to identify content that the user has searched for; generating a second prompt word based on the user history information and the content query instruction; and generating multiple question texts based on the second prompt word using at least one large language model.

[0018] Based on this solution, user needs can be finely divided to search the content library for different needs, so that the recommended content finally generated is more in line with the user's personalized needs.

[0019] In another possible implementation, a first prompt word is generated based on each question text and a text block corresponding to the question text, including: determining a first text based on the generation order of the question texts, where the first text is any question text corresponding to a content query instruction generated using at least one large language model; if the first text has a corresponding second text, generating a first prompt word based on the first text, the text block corresponding to the first text, and a recommendation result corresponding to the second text, where the second text is the question text generated before the first text; if the first text is the first question text corresponding to the content query instruction generated using at least one large language model, generating a first prompt word based on the first text and the text block corresponding to the first text.

[0020] Based on this solution, the order of generating the first prompt word can be obtained by determining the generation order of each question text, and the first prompt word can be optimized based on the recommendation results corresponding to the question text, thereby recursively generating the first prompt word and improving the final recommendation results to better meet the user's personalized needs.

[0021] In another possible implementation, outputting a recommendation result includes: determining a third text based on the generation order of the question text, the third text being the question text corresponding to the content query instruction finally generated using at least one large language model; and outputting a recommendation result corresponding to the third text.

[0022] Based on this solution, the final generated recommendation results can be output to provide users with recommendation results that are highly relevant to the content query instructions they input, thereby optimizing the user experience of using the content recommendation function.

[0023] In a second aspect, an embodiment of the present application also provides a personalized content recommendation system based on a large model, comprising: a receiving module configured to receive a content query instruction; a content inference module configured to generate, in response to receiving the content query instruction, a plurality of question texts corresponding to the content query instruction using at least one large language model; a text matching module configured to determine, based on the similarity between each question text and a text block in a content library, a text block corresponding to the content to be recommended; a text generation module configured to generate a first prompt word based on each question text and the text block corresponding to the question text; the content inference module is further configured to generate, based on the first prompt word, a recommendation result corresponding to the question text using at least one large language model; and an output module configured to output the recommendation result.

[0024] In a third aspect, an embodiment of the present application further provides an electronic device comprising a memory and a processor; the memory and the processor are coupled; the memory is used to store computer program code, and the computer program code comprises computer instructions, and when the processor is used to execute the computer instructions, the electronic device executes a method as described in any one of the first aspects above.

[0025] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a computer, the method of any one of the first aspects is implemented.

[0026] It can be understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A flowchart of a large-model-based personalized content recommendation method provided in an embodiment of the present application;

[0028] Figure 2 A schematic diagram of a process for determining a content library provided in an embodiment of the present application;

[0029] Figure 3 A schematic diagram of a process for storing text blocks provided in an embodiment of the present application;

[0030] Figure 4 A schematic diagram of a process for generating a question text provided in an embodiment of the present application;

[0031] Figure 5 A schematic diagram of a process for matching text blocks based on text similarity provided in an embodiment of the present application;

[0032] Figure 6 A schematic diagram of determining a text block corresponding to a question text provided in an embodiment of the present application;

[0033] Figure 7 A schematic diagram of a process for generating a first prompt word provided in an embodiment of the present application;

[0034] Figure 8 A schematic diagram of generating a first prompt word and a recommendation result provided in an embodiment of the present application;

[0035] Figure 9 A schematic diagram of a large-model-based personalized content recommendation system provided in an embodiment of the present application;

[0036] Figure 10 A schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0037] The following will describe the technical solutions in the embodiments of the present application in conjunction with the accompanying drawings. To facilitate the clear description of the technical solutions in the embodiments of the present application, the first, second, etc. descriptions in the embodiments of the present application are only used for illustration and to distinguish the described objects. There is no order, nor does it represent a special limitation on the number of devices in the embodiments of the present application, and it does not constitute any limitation on the embodiments of the present application.

[0038] The embodiments of the present application provide a large-model-based personalized content recommendation method, system, and electronic device. The method divides the recommended content into text segments and uses a large language model to generate multiple questions based on information such as user needs when the user searches. The divided text is progressively searched, and based on the recommendation results corresponding to each question, recommendation results that meet the user's personalized needs are recursively generated to improve the adaptability of the recommended content to the user.

[0039] Figure 1 A flowchart of a large-model-based personalized content recommendation method provided in an embodiment of the present application.

[0040] The execution subject of each step in the large-model-based personalized content recommendation method provided in the embodiment of the present application can be the large-model-based personalized content recommendation system provided in the embodiment of the present application. Specifically, the large-model-based personalized content recommendation system can be implemented through software and / or hardware and integrated into an electronic device, which can be a terminal device (such as a smart phone, a personal computer, etc.), or a server (such as a local server or a cloud server, or a server cluster, etc.), or a processor, chip, etc. The embodiment of the present application does not limit the type of electronic device integrated with the large-model-based personalized content recommendation system. Figure 1 As shown, the large model-based personalized content recommendation method provided in the embodiment of the present application may include the following steps S100 to S500.

[0041] S100: In response to receiving a content query instruction, using at least one large language model to generate a plurality of question texts corresponding to the content query instruction.

[0042] When a user applies the content recommendation method and system provided in the embodiments of the present application, they can use a portable terminal device, such as a smartphone, to communicate with the electronic device that hosts the content recommendation system, and send content query instructions to the content recommendation system based on the communication connection, so that the content recommendation system can receive the content query instructions sent by the user. In some embodiments, the user can also directly control the electronic device that hosts the content recommendation system to directly input content query instructions into the content recommendation system using the electronic device. The embodiments of the present application do not limit the method for inputting content query instructions into the content recommendation system.

[0043] In embodiments of the present application, a content query instruction may include a query requirement submitted by a user. The query requirement may be textual information entered by the user, or may be content determined by the content recommendation system based on the user data corresponding to the user who sent the content query instruction after receiving the content query instruction. The content recommendation system may determine the user's query requirement by analyzing the content query instruction, and then perform the subsequent process of searching the content library for content to be recommended.

[0044] Figure 2 A schematic diagram of a process for determining a content library provided in an embodiment of the present application.

[0045] In the embodiment of the present application, the content in the content library may be a database determined by the content recommendation system before receiving the content query instruction. Taking the example of the content recommendation system recommending package / service information to the user, the content to be recommended included in the content library may be package / service information, and the text block corresponding to the content to be recommended is the text content corresponding to the package / service information. Figure 2As shown, the process of the content recommendation system determining the content in the content library may include the following steps S101 to S104.

[0046] S101: Obtain a content library.

[0047] In this embodiment, the content library includes text corresponding to the content to be recommended, so that the system can select corresponding text in the content library in the subsequent process by using the question text determined by the large language model to obtain text corresponding to the content to be recommended that better meets user needs.

[0048] Specifically, the content library can be a database set in an electronic device that carries the content recommendation system provided by the embodiment of the present application. The text corresponding to the content to be recommended in the content library can be manually entered by the user, or can be obtained based on the content in the storage medium placed by the user in the electronic device, or can be obtained from the Internet based on a network connection. The embodiment of the present application does not limit the method of obtaining the content to be recommended in the content library.

[0049] In some embodiments, taking the package / service information as an example of the content to be recommended, the content in the content library can be that the electronic device obtains the package / service information from the operator's database storing the package / service information through a communication connection, and then obtains the text content corresponding to the package / service information to obtain the text corresponding to the content to be recommended.

[0050] In another part of the embodiments, the electronic device carrying the content recommendation system can establish a communication connection with the operator's database that stores package / service information, so as to use the database as the content library in the embodiments of the present application, so that when the system obtains the text corresponding to the content to be recommended in the content library, it can obtain the corresponding text by reading the content in the database.

[0051] S102: Divide the text corresponding to the content to be recommended into at least one first text block having a preset text length.

[0052] In an embodiment of the present application, the content recommendation system can avoid missing texts by traversing the texts corresponding to the content to be recommended in the content library, thereby improving the system's recognition efficiency of text content.

[0053] After obtaining the text, the system may divide the text corresponding to each content to be recommended to obtain at least one first text block corresponding to each content to be recommended. The content recommendation system may obtain the first text block by dividing the text into text blocks of a preset text length.

[0054] Specifically, because there may be multiple contents to be recommended in the content library, and there are differences in storage addresses between different contents to be recommended, when the system divides the text blocks, it can divide the text corresponding to each content to be recommended, and the text corresponding to one content to be recommended can be divided into at least one first text block.

[0055] Taking the example of the text corresponding to the content to be recommended containing 200 characters, if the preset text length in the system is 50, when the system divides the text corresponding to the content to be recommended, it can divide the text corresponding to the content to be recommended into 4 first text blocks each containing 50 characters.

[0056] Furthermore, to prevent the system from segmenting content within the same sentence, the content recommendation system can identify punctuation marks, separators, and other elements in the text to determine sentence spacing, thus avoiding segmenting text between two punctuation marks or separators into two first text blocks. This can reduce matching errors caused by incorrect sentence segmentation when selecting text based on similarity.

[0057] Therefore, the system can determine the preset text length as a length range, so that when the text content is segmented according to the text length, the first text block that meets the length requirements is determined based on the text length between adjacent punctuation marks and separators and the preset text length range. Taking the preset text length range of 40 to 60 as an example, when the system divides the text, if it is segmented at the first separator, the resulting text length is 47, and the text length between the first separator and the second separator is 15, the system can determine the text content before the first separator as a first text block, and determine the text between the first separator and the second separator as the content in the next first text block. Among them, the first separator is any punctuation mark or separator in the text, and the second separator is a punctuation mark or separator adjacent to the first separator and located after the first separator.

[0058] In some embodiments, when there are two delimiters that can be divided within the preset text length range, the system can determine the length of the text in the first text block based on the difference in text length between the delimiter and the middle value in the preset text length range. Also, taking the preset text length range of 40 to 60 as an example, when the system divides the text, if it is divided at the first delimiter, the resulting text length is 47, and the text length between the first delimiter and the second delimiter is 12. At this time, the first delimiter and the second delimiter are also within the preset text length range. The content recommendation system can select the text content before the delimiter with a smaller interval as the first text block based on the middle value of the text length range, i.e. 50, and based on the text length between the delimiter and the text corresponding to the length. For example, in this example, the system can determine the text before the first delimiter as the first text block.

[0059] Furthermore, because the text length in the content to be recommended is not fixed, and the division process is performed from front to back based on the reading order of the text, after multiple first text blocks are divided out of each content to be recommended, if the text length of the last remaining text content does not meet the preset text length range, the system can divide this part of the text content into a new first text block to avoid missing content when dividing the text blocks.

[0060] In another embodiment, when dividing text blocks based on a preset text length, the content recommendation system may also determine the text content closest to the preset text length as a first text block based on the text length interval between each delimiter or punctuation mark and the text corresponding to the preset text length. For example, if the preset text length is 50, when the system divides the text, if the text length is 47 when split at the first delimiter, and the text length between the first and second delimiters is 12, the system may determine the text before the first delimiter as the first text block.

[0061] It should be understood that the specific numerical value of the preset text length and the text division method based on the preset text length in the above embodiments are only a few feasible implementation methods in the embodiments of this application. The embodiments of this application do not limit the specific numerical value of the preset text length and the specific method of text division.

[0062] S103: Determine at least one second text block in each first text block based on the semantics of the text in the first text block.

[0063] After obtaining the first text blocks, the content recommendation system can determine the different content contained in each first text block based on the semantics of the text in each first text block, thereby determining at least one second text block in each first text block. The second text block includes at least a portion of the continuous text content in the first text block, and a first text block can be divided into multiple second text blocks.

[0064] In an embodiment of the present application, the system can segment the second text block based on semantics by identifying the semantics corresponding to the text in the first text block. Taking package / service information in the content library as an example, the system can segment the text corresponding to the package / service content by identifying the content in the first text block, such as dynamic traffic, call duration, and charging standards. This allows the system to accurately match the second text block with the content in the content query instruction entered by the user, avoiding the loss of special information in the package / service content, and making the content ultimately recommended by the system more consistent with the user's information.

[0065] In some embodiments of the present application, the content recommendation system may further determine different second text blocks based on separators between text in the first text block. For example, if the first text block includes a separator, the system may determine the text content before the separator and the text content after the separator as a second text block, respectively, thereby separating two second text blocks from the first text block.

[0066] It should be noted that the above-mentioned division methods are only several feasible implementation methods in the embodiments of this application. When the content library includes text corresponding to other types of information, the system may analyze the semantics and divide the second text block in a different way from the above-mentioned embodiments, so that its division method is more in line with the characteristics corresponding to this type of information. This application will not elaborate on this.

[0067] S104: Storing the first text block and the second text block in a content library.

[0068] After the content recommendation system completes the division of the text blocks, it can store the first and second text blocks in the content library for the system to access during subsequent queries. It should be noted that the technical solution of storing the obtained first and second text blocks in the content library is only implemented in the scenario where the content library is set up in the electronic device that hosts the content recommendation system. In this case, the content recommendation system uses the read and write permissions of the content library to enable the system to read or write the corresponding content in the content library.

[0069] In another embodiment of the present application, the content recommendation system can obtain content from a database located in another device or storage medium through a communication connection. In this case, the content recommendation system usually only has read permission on the database and cannot store the first text block and the second text block in the database. Therefore, before storing the first text block and the second text block, the system can query whether there is a content library to store the first text block and the second text block in the storage medium to which it has read and write permission. If the content library does not exist in the system, the content recommendation system can build a content library in the storage medium to which it has read and write permission and store the first text block and the second text block in the content library.

[0070] Figure 3 A flowchart of storing text blocks provided in an embodiment of the present application.

[0071] In the embodiment of the present application, since the content recommendation system can call the first text block and the second text block in the subsequent steps to match the text content similarity, the system needs to process the text blocks accordingly during the storage process to optimize the retrieval efficiency in the subsequent retrieval process. Figure 3 As shown in (a), the process of storing text blocks provided by the embodiment of the present application may include the following steps S1041a and S1042a.

[0072] S1041a: Determine a first mapping relationship between the second text block and the first text block, and determine a second mapping relationship between the first text block and the content to be recommended, based on the acquisition sources of the first text block and the second text block.

[0073] After obtaining the first text block and the second text block, the content recommendation system can determine the second mapping relationship and the first mapping relationship based on the acquisition source of the first text block, that is, the content to be recommended to which the first text block belongs, and the acquisition source of the second text block, that is, the first text block corresponding to the second text block.

[0074] In an embodiment of the present application, the first mapping relationship may be determined when the content recommendation system divides the second text block from the first text block, and the second mapping relationship may be determined when the content recommendation system divides the first text block from the text corresponding to the content to be recommended. This embodiment of the present application does not limit the method for obtaining the first mapping relationship and the second mapping relationship.

[0075] S1042a: Storing the text contents, the first mapping relationship, and the second mapping relationship corresponding to the first text block and the second text block in a content library.

[0076] After obtaining the first mapping relationship and the second mapping relationship, the content recommendation system may store the first text block, the second text block, the first mapping relationship, and the second mapping relationship together in the content library.

[0077] For example, the first mapping relationship stored in the content library may be a correspondence between the storage address of the second text block and the first text block in the content library, and the second mapping relationship stored in the content library may be a correspondence between the storage address of the first text block and the text corresponding to the content to be recommended. In this way, when the system can call the second text block, it calls the first text block at the storage address corresponding to the storage address corresponding to the second text block based on the storage address corresponding to the second text block.

[0078] Based on this embodiment, the system can establish a correspondence between text blocks, and achieve decoupling between text contents by dividing the content to be recommended into text blocks of multiple granularities, so as to facilitate the system to implement a progressive multi-granularity retrieval process for the recommended content in subsequent steps.

[0079] In some embodiments of the present application, the content recommendation system needs to use vector similarity to compare the similarity between the text block and the content query instruction, such as Figure 3 As shown in (b), the process of the content recommendation system storing the first text block and the second text block in the content library may include the following steps S1041b to S1043b.

[0080] S1041b: Convert each second text block into a corresponding text vector.

[0081] In an embodiment of the present application, after the content recommendation system divides and obtains the second text block, it can use the vectorization tool built into the system to convert the second text block into a corresponding text vector. Exemplarily, the vectorization tool built into the content recommendation system can be a static word vector model (word embeddings, Word2Vec) or a dynamic context-aware model (bidirectional encoder representations from transformers, BERT), wherein the static word vector model can convert words into fixed-length vectors, and the dynamic context-aware model can obtain a vector representation of each input sentence.

[0082] It should be noted that the vectorization tool used in the embodiments of the present application is not limited to the above two models, and can also be other tools with the function of vectorizing text. This application does not impose any restrictions on this.

[0083] S1042b: Determine a third mapping relationship between the text vector and the first text block, and determine a second mapping relationship between the first text block and the content to be recommended, based on the acquisition sources of the first text block and the second text block.

[0084] After determining the text vector corresponding to each second text block, the system can determine the second mapping relationship and the first mapping relationship based on the acquisition source of the first text block, that is, the content to be recommended to which the first text block belongs, and the acquisition source of the second text block, that is, the first text block corresponding to the second text block.

[0085] Furthermore, the system can also determine a third mapping relationship between the text vector and the first text block based on the text vector corresponding to the second text block and the first mapping relationship, so that the system can obtain the text content corresponding to the first text block by matching the vector corresponding to the second text block.

[0086] S1043b: Store the text content, text vector, second mapping relationship, and third mapping relationship corresponding to the first text block in a content library.

[0087] After obtaining the second and third mapping relationships, the content recommendation system can store the text content, text vector, second mapping relationship, and third mapping relationship corresponding to the first text block in the content library. At this time, the system may not store the text content corresponding to the second text block, but only store the text vector, reducing the load on the content library. At the same time, during the retrieval process, it reduces the number of mapping relationship queries, allowing the system to directly map to the corresponding first text block through the text vector.

[0088] Furthermore, the second mapping relationship stored in the content library may be a correspondence between the first text block in the content library and the storage address of the text corresponding to the content to be recommended, and the third mapping relationship stored in the content library may be a correspondence between the text vector of the second text block in the content library and the storage address of the first text block. In this way, when the system determines the text content based on vector similarity, it calls the first text block at the storage address corresponding to the text vector based on the storage address corresponding to the text vector.

[0089] Based on this embodiment, the system can establish a correspondence between the text vector and the first text block, which facilitates the system's call of content in the content library, optimizes the text retrieval process, and facilitates the system to implement a progressive multi-granularity retrieval process for recommended content in subsequent steps.

[0090] Figure 4 A flowchart of question text generation provided in an embodiment of the present application.

[0091] In some embodiments of the present application, in a scenario where the content library is set up on an electronic device that carries the content recommendation system, the content recommendation system may also store the user's personal data, the content contained in the content query instructions input by the user in the past, and the recommendation results corresponding to the content query instructions input by the user in the past in the content library. In this way, after receiving the content query instruction, the system can query the content library for the user's corresponding personal data, the content contained in the content query instructions input in the past, and the recommendation results corresponding to the content query instructions input by the user in the past based on the user who sent the content query instruction, thereby improving the efficiency of the system in obtaining user information. At the same time, the system can also determine the corresponding question text through the historical information and the content in the currently input content query instruction, thereby optimizing the system's query process for the text in the content library. Figure 4 As shown, the process of generating question text in the embodiment of the present application may include the following steps S110 to S130.

[0092] S110: Obtain user history information.

[0093] User history information is used to identify content that a user has searched for. This information can be the recommendation results stored in the content library by the system after the user has sent historical content query instructions to the system and obtained corresponding recommendation results. After receiving the content query instruction input by the user, the content recommendation system can query the content library for the user's corresponding user history information based on the user data in the content query instruction.

[0094] In some embodiments of the present application, because there is a scenario where the user uses the content recommendation system for the first time, the user's corresponding user history information is not stored in the system at this time. At this time, the system can obtain the user's user history information by collecting the corresponding information from the portable terminal device to which the user sends the content query instruction.

[0095] Taking the content recommendation system recommending package / service information to users as an example, when the user sends a content query instruction, the existing package / service information in the portable terminal device that sends the instruction will be sent to the system together with the content query instruction. After obtaining the historical information in the content query instruction, the system can store it in the content library, so that the system can obtain the user's historical information when generating the question text.

[0096] Furthermore, the content query instruction may also include the query requirements of the user in this query. After receiving the content query instruction, the content recommendation system can obtain the user's query requirements in the content query instruction, so that the system can generate questions that meet the user's personalized needs based on the user's current needs and user historical information, and obtain corresponding recommendation results based on the questions, thereby improving the personalization of content recommendations and optimizing the user's experience of using the system for content recommendation.

[0097] S120: Generate a second prompt word based on the user history information and the content query instruction.

[0098] After obtaining the user's historical information and content query instructions, a second prompt word for generating the question text can be generated based on the obtained historical information and query requirements. For example, the content query system can include a corresponding second prompt word template within it, so that after obtaining the historical information and content query instructions, the system fills the obtained historical information and query requirements into the second prompt word template to obtain the corresponding second prompt word.

[0099] In some embodiments of this application, the second prompt word template may include multiple elements. Specifically, the second prompt word template may include task information, input prompts, and output prompts. Task information refers to the task that the large language model needs to perform, input prompts are information that the large language model needs to refer to when performing the task in the task information, and output prompts can be used to indicate the content generated by the large language model and the format of the generated content.

[0100] Because the question text determined in step S100 in the embodiment of the present application can be a detailed division of the information input by the user and the information obtained by the system query, the task information in the second prompt word template can be "Please assist the user in determining multiple question texts for the content to be recommended in the content library."

[0101] The input prompt in the second prompt template can be dynamically filled based on the user's historical information and content query instructions obtained by the system. The system can directly fill the input prompt with the historical information and query requirements it obtains. For example, the input prompt content can be "Based on the following user's historical information and query requirements, analyze the user's corresponding content requirements."

[0102] For example, in a scenario where the content recommendation system can recommend packages / service information to users, the format of historical information and query requirements that can be filled in the second prompt word template can be as follows:

[0103]

[0104]

[0105] Historical Information

[0106] 1.{}

[0107] 2.{}”.

[0108] The "type" refers to the type of content the user is searching for, such as packages / services or other content, and this application does not impose any restrictions on this. Furthermore, because the content corresponding to the query requirement may include more or less content than in the template, when generating the second prompt word, the system can use the content that can be filled into the second prompt word template to generate the corresponding second prompt word.

[0109] In some embodiments of the present application, because the content library may include different types of recommended content, after obtaining the content library and the second text block, the content recommendation system can use the large language model to divide the second text block in the content library to summarize the various information categories corresponding to the text therein, thereby obtaining items corresponding to the query requirements and updating them into the second prompt word template. When the content recommendation system uses the second prompt word template to generate the second prompt word, it can obtain the items in the query content, fill the items in the query requirements into the corresponding positions in the second prompt word template, and set other items corresponding to the query requirements in the input prompt to null values, so that the system can generate the corresponding input prompt based on the content filled into the template.

[0110] In an embodiment of the present application, the output prompt in the second prompt word template may be "Give three different question texts that can reflect the user's needs", where the number of generated question texts can be a fixed value or a value obtained by the system in the content query instruction. On this basis, the content of the second prompt word generated by the content recommendation system may be "Please assist the user in determining multiple question texts for the content to be recommended in the content library. Based on the user's historical information and query needs described below, analyze the user's corresponding content needs and provide three different question texts that can reflect the user's needs."

[0111] It should be noted that the method of generating the second prompt word using the second prompt word template given in the embodiment of the present application is only an example in the embodiment of the present application, and the embodiment of the present application does not limit the generation process of the second prompt word.

[0112] S130: Based on the second prompt word, use at least one large language model to generate multiple question texts.

[0113] After receiving the second prompt, the content recommendation system can use the large language model to generate multiple question texts corresponding to the number of prompts output in the second prompt. The large language model used in the embodiments of this application can be any one or more large language models with language and text understanding and reasoning capabilities. The embodiments of this application do not limit the type of large language model used.

[0114] The content recommendation system can utilize at least one large language model to perform an inference process based on the task information in the second prompt and the input prompt, inferring and obtaining multiple question texts corresponding to the user's needs. The number of question texts generated by the large language model is the same as the number of question texts provided in the output prompt. In this way, the large language model can be used to break down and refine the user's content needs, and searches can be performed sequentially based on these refined needs, optimizing the system's process for determining recommendation results and increasing the relevance of the final recommendation results to the user's needs.

[0115] S200: Determine the text block corresponding to each question text based on the similarity between each question text and the text blocks in the content library.

[0116] The content library includes text blocks corresponding to the content to be recommended. It should be understood that the content library in this embodiment is the same as the content library in the previous embodiment, and this application will not elaborate on this. After obtaining the question text, the content recommendation system can determine the text block corresponding to each question text based on the similarity between each question text and the corresponding second text block in the content library.

[0117] Figure 5 A flowchart of a method for matching text blocks based on text similarity is provided in an embodiment of the present application.

[0118] In some embodiments of this application, when comparing the similarity between the question text and the text blocks in the content library, the content recommendation system can use the similarity between the vectors corresponding to the question text and the text blocks to obtain the similarity between the texts. Figure 5 As shown, the process of matching text blocks based on text similarity may include the following steps S210 to S230.

[0119] S210: Convert the question text into a question vector.

[0120] After obtaining the question text, the system can convert each question text into a corresponding question vector, allowing the system to subsequently compare the question vector with the text vector corresponding to the second text block in the content library. It should be understood that the vectorization tool used to vectorize the question text in the embodiment of the present application is the same as the vectorization tool used by the system to vectorize the second text block. This ensures that the generation rules for question vectors and text vectors are consistent, avoiding the problem of texts with high vector similarity but low actual content similarity.

[0121] S220: Obtain similarity between the question vector and the vector corresponding to each second text block in the content library.

[0122] After obtaining the second question vector, the content recommendation system can traverse the text vectors in the content library based on the question vector to obtain the vector similarity between the question vector and the text vector. Exemplarily, the vector similarity obtained by the content recommendation system is the cosine similarity between the question vector and the text vector. Specifically, the vector similarity can be calculated based on the quotient between the product of the vectors and the product of the vector moduli.

[0123] It should be understood that the maximum value of vector similarity is 1, at which point the question vector and the text vector are exactly the same, and the minimum value is 0, at which point the question vector and the text vector are completely different. The vector similarity calculated by the content recommendation system can be any value, and this application does not impose any restrictions on this.

[0124] When obtaining vector similarities, the content recommendation system can determine the similarity ranking of the text vectors corresponding to the question vector based on the obtained vector similarities. For example, the system can construct a similarity list during the process of obtaining vector similarities and sort the text vectors based on the obtained vector similarities. After the traversal is complete, the system can select the text vector corresponding to the second text block with the highest vector similarity.

[0125] S230: Determine the first text block corresponding to the second text block whose vector similarity is greater than a preset threshold as the text block corresponding to the question text.

[0126] In an embodiment of the present application, based on the value range of vector similarity, in order to make the question vector and the text vector have a higher correlation, the preset threshold can be any value such as 0.9, 0.85, 0.8, etc. The embodiment of the present application does not impose any other restrictions on the specific value of the preset threshold of vector similarity.

[0127] After the system obtains a second text block with a high vector similarity, the system can determine the first text block corresponding to the second text block as the text block corresponding to the question text based on the correspondence between the first and second text blocks. Because the content recommendation system can screen multiple second text blocks with vector similarities greater than a preset threshold, multiple first text blocks can also be obtained based on multiple second text blocks. The system can determine the corresponding first text block as the text block corresponding to the question text by reading the text content corresponding to the corresponding first text block from the content library.

[0128] Furthermore, because the content recommendation system can generate a corresponding order of text vectors based on vector similarity from high to low during the process of generating vector similarity, the content recommendation system can also determine the corresponding first text block by selecting the text vectors that are ranked at the top of the text vector order. For example, the content recommendation system can select the first five text vectors in the text vector order as candidate texts, and then determine the text block corresponding to the question text based on these five text vectors.

[0129] Figure 6 A schematic diagram of determining a text block corresponding to a question text provided in an embodiment of the present application.

[0130] like Figure 6 As shown, in some embodiments of the present application, after obtaining a question vector corresponding to a question text, the content recommendation system can query the content library for a text vector corresponding to a second text block with a high vector similarity based on the question vector. A question vector can correspond to multiple text vectors that meet a preset vector similarity threshold, and these text vectors can correspond to the same first text block or different first text blocks.

[0131] Specifically, such as Figure 6 As shown, for a question vector corresponding to a question text, the system determines the number of second text blocks to be five based on vector similarity, namely, second text blocks a to second text blocks e in the figure. The first mapping relationships of second text blocks a, b, and c all point to first text block a, and the first mapping relationships of second text blocks d and e all point to first text block b. At this point, the system can execute subsequent step S300 based on first text block a, first text block b, and the question text to generate a first prompt word corresponding to the question text.

[0132] Based on this solution, the content recommendation system can match the segmented text content to a relatively complete first text block with high contextual coherence, so as to achieve a more refined text content matching process, improve the accuracy of information retrieval for different needs during the text retrieval process, and improve the coherence of the final text content context.

[0133] S300: Generate a first prompt word according to each question text and the text block corresponding to the question text.

[0134] After obtaining the text block corresponding to each question text, the content recommendation system can generate a first prompt word based on the obtained question text and the corresponding text block, so that the user can determine the recommendation result based on the matched text block.

[0135] In an embodiment of the present application, the system can set a first prompt word template. After the system obtains the question text and determines the corresponding text block based on the question text, the obtained question text and the corresponding text block can be filled into the first prompt word template to obtain the first prompt word.

[0136] Specifically, the first prompt word template can have a similar structure to the second prompt word template, also having multiple elements. For example, the first prompt word template can include task information, input prompts, and output prompts. The content included in the first prompt word template has the same function as that in the second prompt word template, and this application does not elaborate on this.

[0137] In this embodiment of the present application, when generating a first prompt word, the content recommendation system may use the question text used when generating the first prompt word as the task information for the first prompt word, and the text block corresponding to the question text as the input prompt. The output prompt may be "obtain a recommendation result that meets the above content in the content library and output the text content corresponding to the recommendation result."

[0138] See also Figure 6 The matching result of a question text and a text block in the content library shown in the figure can correspond to two first text blocks. Therefore, when the system generates the first prompt word, it can use the question text as task information and the text content in the corresponding two first text blocks as input prompts, which together with the output prompts constitute a complete first prompt word.

[0139] It should be understood that the above-mentioned method for generating the first prompt word is only one feasible implementation method in the embodiment of the present application, and the embodiment of the present application does not limit the method for generating the first prompt word.

[0140] Figure 7 A schematic diagram of a process for generating a first prompt word provided in an embodiment of the present application.

[0141] Because the above embodiment provides an example of a content recommendation system generating multiple question texts based on a content query instruction input by a user, when the system processes the content query instruction, if the content recommendation system directly determines the recommendation results separately based on the first prompt word corresponding to each question text, multiple recommendation results will be generated, and there may be a situation where each recommendation result only meets part of the user's needs, which is not conducive to the system determining the optimal recommendation method for the user.

[0142] Therefore, in some embodiments of this application, when the content recommendation system generates recommendation results using a large language model inference, it can adopt a recursive generation method to update the first prompt word corresponding to the question text, thereby improving the accuracy of the recommendation results determined by the system by optimizing the content of the prompt word. Figure 7As shown, the process of generating the first prompt word in the embodiment of the present application may include the following steps S310 to S330.

[0143] S310: Determine the first text based on the generation order of the question texts.

[0144] In the embodiments of the present application, the first text is any question text generated using at least one large language model corresponding to a content query instruction. The system can determine any generated question text as the first text. It should be understood that the first text is merely an identifier of the question text in the system and does not limit the content or generation order of the question text. Instead, it serves only as an identifier to describe the question text in the embodiments.

[0145] After determining the first text, the system can determine whether there is a corresponding question text before the first text based on the position of the first text in the generation order of multiple question texts, and thus generate a corresponding first prompt word based on the position of the first text in the generation order of multiple question texts.

[0146] S320: If the first text has a corresponding second text, generate a first prompt word according to the first text, the text block corresponding to the first text, and the recommendation result corresponding to the second text.

[0147] After determining the first text, the system can determine whether the first text has a second text by determining the position of the first text in the generation order of multiple question texts. The second text is the question text generated before the first text. Like the first text, the second text also serves as an identifier for the question text in the system. It does not limit the content and generation order of the question text, but only serves as an identifier to express the question text in the embodiment.

[0148] Figure 8 A schematic diagram of generating a first prompt word and a recommendation result provided in an embodiment of the present application.

[0149] Take the content recommendation system as an example to generate three question texts based on a received content query instruction, such as Figure 8 As shown, if the first text is the second question text in the generation order, the first text has a second text, and the second text corresponding to the first text is the first question text in the generation order; and if the first text is the third question text in the generation order, the first text has a second text, and the second text corresponding to the first text is the second question text in the generation order.

[0150] In this embodiment of the present application, the order of the first text can be determined to determine the generation method of the first prompt word corresponding to different question texts. If the first text does not have a second text, that is, the first text is the first question text in the question text generation order, the system can execute steps S330 and S400 to obtain the recommendation result corresponding to the first text.

[0151] The system can then use the second question text in the generation order as the first text, and the first question text that has generated the corresponding recommendation result as the second text, execute step S320, determine the first text as the task information in the first prompt word, and determine the first text block corresponding to the first text and the recommendation result obtained by the second text in the aforementioned step as the input prompt in the first prompt word, so that the result corresponding to the previously executed question text and the text block corresponding to the first text together constitute the context information when inferring the recommendation result corresponding to the first text.

[0152] Finally, the system may use the third question text in the generation order as the first text, and the second question text for which the corresponding recommendation result has been generated as the second text, execute step S320, determine the first text as the task information in the first prompt word, and determine the first text block corresponding to the first text and the recommendation result obtained by the second question text in the aforementioned step as the input prompt in the first prompt word, so that the result corresponding to the second question text and the text block corresponding to the first text together constitute the context information when inferring the recommendation result corresponding to the first text.

[0153] Based on this solution, the system can implement recursive queries based on multiple question texts, and recursively optimize the system's query effect in the content library, so that the final recommendation results meet the user's personalized needs.

[0154] It should be understood that the above-mentioned method of generating the first prompt word is also only a feasible implementation method in the embodiment of the present application. The embodiment of the present application does not limit the number of question texts and the method of recursively generating content.

[0155] S330: If the first text is a first question text corresponding to a content query instruction generated using at least one large language model, a first prompt word is generated according to the first text and a text block corresponding to the first text.

[0156] In the case where the first text is the first question text corresponding to the content query instruction generated by the system using at least one large language model, the first text has no corresponding second text. Therefore, when generating the first prompt word, it can directly generate the first prompt word by determining the first text as task information and the first text block corresponding to the first text as input prompt, so that the subsequent step S400 calls the first prompt word to generate the corresponding recommendation result.

[0157] S400: Based on the first prompt word, use at least one large language model to generate a recommendation result corresponding to the question text.

[0158] In an embodiment of the present application, after generating a first prompt word based on a question text, the content recommendation system can invoke at least one large language model to generate recommendation results corresponding to the question text, thereby determining recommendation results in the content library that are relevant to the user's query requirements. The recommendation results can include text corresponding to any content to be recommended in the content library. When the content recommendation system subsequently generates the first prompt word, it can determine the text contained in the recommendation results as contextual information used to generate the input prompt for the first prompt word.

[0159] In some embodiments of the present application, after obtaining multiple question texts, the system may execute the vector similarity calculation process for each question text respectively to obtain the text block corresponding to each question text respectively. The system may then execute steps S300 and S400 in sequence according to the generation order of the question texts to recursively generate the first prompt word and thereby determine the corresponding recommendation result.

[0160] like Figure 8 As shown, taking the example of a content recommendation system that splits a user's content needs into three question texts based on a large language model, the system can first generate a first prompt word based on the question text that ranks first in the question text generation order and its corresponding text block, and then input it into the large language model to obtain a recommendation result for the first question text.

[0161] After obtaining the recommendation result corresponding to the first question text, the system can generate a second prompt word by combining the recommendation result with the second question text in the question text generation order and its corresponding text block, and then generate a recommendation result corresponding to the second question text through the large language model.

[0162] Then, the content recommendation system can use the recommendation result corresponding to the second question text as an input of the first prompt word of the third question text, and jointly generate the first prompt word together with the third question text and the text block corresponding to the third question text, so that the system can infer and obtain the final recommendation result through the first prompt word composed of the recommendation result, the question text and the text block corresponding to the question text, so as to improve the relevance between the recommendation result obtained by the system and the user.

[0163] In some embodiments of the present application, the content recommendation system may, based on the number of determined question texts, execute steps S310, S330, and S400 in sequence after step S200 is completed, and then return to execute steps S320 and S400 until the system generates corresponding recommendation results for each question text.

[0164] S500: Output recommendation results.

[0165] In an embodiment of the present application, after the content recommendation system obtains the recommendation results corresponding to each question text, the system can select one of the recommendation results for output to provide the user with a recommendation result for the content query instruction input by the user.

[0166] In some embodiments of the present application, the system can output the recommendation results corresponding to each question text it obtains to the user, present multiple recommendation results to the user in the form of a list, provide multiple recommendation results for the user to choose from, and optimize the content recommendation experience.

[0167] In another embodiment of the present application, the system can also determine the question text corresponding to the content query instruction that is finally generated by the system using at least one large language model as the third text based on the generation order of the question text, and output the recommendation results corresponding to the third text to provide users with content generated after multiple recursions, so that the recommendation results generated by the content recommendation system are more in line with the user's personalized needs.

[0168] Specifically, based on the solution for generating recommendation results in the above implementation, the system can use the recommendation result finally generated by the large language model as the final recommendation result of the system. When the system executes step S500, the recommendation result can be directly output to provide the user with the recommendation result.

[0169] Based on the solution in the embodiment of the present application, the recommended content can be divided into multi-granularity texts, and when the user searches, a large language model can be used to generate multiple question texts based on information such as user needs, and the divided text blocks can be progressively searched to recursively generate recommendation results that meet the user's personalized needs based on the recommendation results corresponding to each question text, so as to improve the adaptability of the recommended content to the user and meet the user's personalized needs.

[0170] Figure 9 A schematic diagram of a large-model-based personalized content recommendation system provided in an embodiment of the present application.

[0171] Corresponding to the embodiment of the aforementioned personalized content recommendation method based on a large model, the present application also provides an embodiment of a personalized content recommendation system based on a large model. Figure 9 As shown, the large model-based personalized content recommendation system 100 may include:

[0172] The receiving module 110 is configured to receive a content query instruction.

[0173] The content reasoning module 120 is configured to generate a plurality of question texts corresponding to the content query instruction using at least one large language model in response to receiving the content query instruction.

[0174] The text matching module 130 is configured to determine the text block corresponding to each question text based on the similarity between each question text and the text blocks in the content library, wherein the content library includes the text blocks corresponding to the content to be recommended.

[0175] The text generation module 140 is configured to generate a first prompt word according to each question text and the text block corresponding to the question text.

[0176] The content reasoning module 120 is further configured to generate a recommendation result corresponding to the question text based on the first prompt word using at least one large language model.

[0177] The output module 150 is configured to output the recommendation result.

[0178] Figure 10 A schematic diagram of an electronic device provided in an embodiment of the present application.

[0179] In some embodiments, an electronic device may include one or more processors and a memory coupled to the processor. The memory is configured to store one or more computer programs. The computer program code includes computer instructions, and when the processor executes the computer instructions, the electronic device performs the large model-based personalized content recommendation method described in the above embodiments.

[0180] like Figure 10 As shown, the electronic device 200 includes a processor 201 and a memory 202. Exemplarily, the electronic device 200 may further include a communication interface 203 and a communication bus 204.

[0181] The processor 201, the memory 202 and the communication interface 203 communicate with each other via the communication bus 204. The communication interface 203 is used to communicate with other devices such as clients or other servers.

[0182] In some embodiments, the processor 201 is configured to execute the computer program 205, specifically, to execute the relevant steps in the above-mentioned embodiment of the large model-based personalized content recommendation method. Specifically, the computer program 205 may include computer program code, which includes computer executable instructions.

[0183] For example, the processor 201 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement some embodiments of the present application. The electronic device 200 may include one or more processors, which may be processors of the same type, such as one or more CPUs, or different types of processors, such as one or more CPUs and one or more ASICs.

[0184] In some embodiments, the memory 202 is used to store the computer program 205. The memory 202 may include a high-speed RAM memory, and may also include a non-volatile memory (NVM), such as at least one disk memory.

[0185] The computer program 205 can be specifically called by the processor 201 to enable the electronic device 200 to perform the operation of the large model-based personalized content recommendation method.

[0186] Some embodiments of the present application further provide a computer-readable storage medium, which stores at least one executable instruction. When the executable instruction runs on the electronic device 200, the electronic device 200 executes the large model-based personalized content recommendation method in the above embodiment.

[0187] The executable instructions can be specifically used to enable the electronic device 200 to perform the operation of the personalized content recommendation method based on the large model.

[0188] For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc (CD-ROM), a magnetic tape, a floppy disk, an optical data storage device, and the like.

[0189] The beneficial effects that can be achieved by the readable storage medium provided in some embodiments of the present application can be referred to the beneficial effects of the corresponding large model-based personalized content recommendation method provided above, and will not be repeated here.

[0190] The implementation methods described above are only specific implementation methods of the present application and are not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solution of the present application should be included in the scope of protection of the present application.

Claims

1. A personalized content recommendation method based on a large model, characterized in that: include: In response to receiving a content query instruction, generating a plurality of question texts corresponding to the content query instruction using at least one large language model; Determining the text block corresponding to each question text based on the similarity between each question text and a text block in a content library, wherein the content library includes text blocks corresponding to the content to be recommended; generating a first prompt word according to each question text and the text block corresponding to the question text; Based on the first prompt word, using the at least one large language model to generate a recommendation result corresponding to the question text; The recommendation result is output.

2. The large model-based personalized content recommendation method according to claim 1, characterized in that: Before generating a plurality of question texts corresponding to the content query instruction using at least one large language model in response to receiving the content query instruction, the method further includes: Obtaining the content library, wherein the content library also includes text corresponding to the content to be recommended; Dividing the text corresponding to the content to be recommended into at least one first text block having a preset text length; Determining at least one second text block in each of the first text blocks based on the semantics of the text in the first text blocks; The first text block and the second text block are stored in the content library.

3. The large model-based personalized content recommendation method according to claim 2, characterized in that: Determining the text block corresponding to each question text based on the similarity between each question text and the text blocks in the content library includes: Convert the question text into a question vector; Obtaining a similarity between the question vector and a vector corresponding to each second text block in the content library; The first text block corresponding to the second text block whose vector similarity is greater than a preset threshold is determined as the text block corresponding to the question text.

4. The large model-based personalized content recommendation method according to claim 2, characterized in that: The storing the first text block and the second text block in the content library includes: Determining a first mapping relationship between the second text block and the first text block, and determining a second mapping relationship between the first text block and the content to be recommended, based on the acquisition sources of the first text block and the second text block; The text contents corresponding to the first text block and the second text block, the first mapping relationship, and the second mapping relationship are stored in the content library.

5. The large model-based personalized content recommendation method according to claim 2, characterized in that: The storing the first text block and the second text block in the content library includes: Convert each of the second text blocks into a corresponding text vector; determining, based on the acquisition sources of the first text block and the second text block, a third mapping relationship between the text vector and the first text block, and a second mapping relationship between the first text block and the content to be recommended; The text content corresponding to the first text block, the text vector, the second mapping relationship, and the third mapping relationship are stored in the content library.

6. The large model-based personalized content recommendation method according to claim 2, characterized in that: Generating a plurality of question texts corresponding to the content query instruction using at least one large language model includes: Obtaining user history information, where the user history information is used to identify content that the user has searched for; generating a second prompt word based on the user history information and the content query instruction; Based on the second prompt word, a plurality of question texts are generated using the at least one large language model.

7. The large model-based personalized content recommendation method according to any one of claims 1 to 6, characterized in that: The step of generating a first prompt word according to each question text and the text block corresponding to the question text includes: Determining a first text based on the generation order of the question texts, where the first text is any question text corresponding to the content query instruction generated using at least one large language model; If the first text has a corresponding second text, generating the first prompt word according to the first text, the text block corresponding to the first text, and the recommendation result corresponding to the second text, where the second text is the question text generated before the first text; If the first text is a first question text corresponding to the content query instruction generated using at least one large language model, the first prompt word is generated according to the first text and the text block corresponding to the first text.

8. The large model-based personalized content recommendation method according to claim 7, characterized in that: The outputting the recommendation result includes: Determining a third text based on the generation order of the question texts, where the third text is the question text corresponding to the content query instruction that is last generated using at least one large language model; The recommendation result corresponding to the third text is output.

9. A personalized content recommendation system based on a large model, characterized by: include: A receiving module configured to receive a content query instruction; a content reasoning module configured to generate a plurality of question texts corresponding to the content query instruction using at least one large language model in response to receiving the content query instruction; a text matching module configured to determine the text block corresponding to each question text based on similarity between each question text and a text block in a content library, wherein the content library includes text blocks corresponding to the content to be recommended; A text generation module is configured to generate a first prompt word according to each question text and the text block corresponding to the question text; The content reasoning module is further configured to generate a recommendation result corresponding to the question text based on the first prompt word using the at least one large language model; The output module is configured to output the recommendation result.

10. An electronic device, characterized in that: It includes a memory and a processor; the memory and the processor are coupled; the memory is used to store computer program code, and the computer program code includes computer instructions. When the processor is used to execute the computer instructions, the electronic device executes the large model-based personalized content recommendation method as described in any one of claims 1 to 8.

Images