Message pushing method, system and device and electronic equipment

Abstract

The invention discloses a message pushing method, system and device, electronic equipment and a computer readable storage medium. The method comprises the following steps: searching and acquiring a plurality of messages related to information to be concerned from a plurality of preset information source services according to the information to be concerned preset by a user; generating a summary text corresponding to each message through a large language model; obtaining a text vector corresponding to each message; sequentially aiming at the text vector corresponding to each message, searching in a vector database and judging the similarity between the text vector corresponding to the message and a database text vector; and determining a target message based on a similarity retrieval result, storing a text vector corresponding to the target message into a vector database, and pushing a summary text corresponding to the target message to the user. According to the method, the technical problems of high labor cost, incapability of realizing customization, incapability of automatically summarizing message contents and high possibility of repeated pushing in the prior art are solved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a message push method, system, device, electronic device, and computer-readable storage medium. Background Art

[0002] With the development of information technology and the Internet, Internet information has shown explosive growth, and message push has gradually become an important way for users to conveniently obtain Internet information.

[0003] At present, message push mostly relies on manual search on various Internet service platforms, and then manually screening and summarizing the searched messages, which undoubtedly increases the labor cost of message push. And due to the limitations of manual search capabilities, the pushed messages cannot be customized, resulting in a poor user experience. In order to solve the problems existing in the manual message push method, an automated message push method has emerged, but the current automated message push method is relatively simple, mainly relying on web crawlers to crawl links to relevant messages on various service platforms, and push the links to users, so that users can enter various service platforms based on the links to read the corresponding messages. Since the actual content of the message cannot be understood by crawling the links of relevant messages based on web crawlers, the message content will not be summarized, which often causes users to spend a lot of time reading the entire content of the message, and it is also easy to push repeatedly.

[0004] Therefore, the existing message push technology has technical problems such as high labor costs, inability to achieve customization, inability to automatically summarize message content, and prone to repeated push. Summary of the invention

[0005] The present application provides a message push method, system, device, electronic device, and computer-readable storage medium to solve the technical problems of the prior art, such as high labor cost, inability to achieve customization, inability to automatically summarize message content, and easy occurrence of repeated push.

[0006] In a first aspect, an embodiment of the present application provides a message push method, the method comprising: searching and obtaining multiple messages related to the information to be paid attention to from multiple preset source services according to the information to be paid attention to preset by the user; inputting each of the messages into a large language model, and generating a summary text corresponding to each of the messages through the large language model; vectorizing the summary text corresponding to each of the messages to obtain a text vector corresponding to each of the messages; searching and judging the similarity between the text vector corresponding to the message and the database text vector stored in the vector database in turn for each of the text vectors corresponding to the message; wherein the vector database stores multiple database text vectors, and the database text vector is the text vector corresponding to the pushed message; based on the similarity retrieval result of the text vector corresponding to the message, determining the target message from the messages, storing the text vector corresponding to the target message in the vector database, and pushing the summary text corresponding to the target message to the user.

[0007] In a second aspect, an embodiment of the present application provides a message push system, the system comprising: a module for customizing information to be paid attention to and information sources, a message collection module, a text processing module, a text database, a vector processing module, a vector database, and a push module; the module for customizing information to be paid attention to and information sources is used to provide users with a window interface for customizing information to be paid attention to and information source services; the message collection module is used to search and obtain multiple messages related to the information to be paid attention to from multiple customized information source services based on the information to be paid attention to customized by the user; the text processing module is used to input each of the messages into a large language model, and generate a summary text corresponding to each of the messages through the large language model; the text database is used to store a message list corresponding to each of the messages, and the message list records at least the information of the message. The summary text of the message; the vector processing module is used to vectorize the summary text corresponding to each of the messages to obtain the text vector corresponding to each of the messages; and for each of the text vectors corresponding to the message, the similarity between the text vector corresponding to the message and the database text vector stored in the vector database is retrieved and judged in the vector database; the vector database is used to store database text vectors, and the database text vectors are text vectors corresponding to the pushed messages; the push module is used to determine the target message from the messages based on the similarity retrieval results of the text vectors corresponding to the messages, store the text vectors corresponding to the target messages in the vector database, and push the summary text corresponding to the target messages to the user.

[0008] In a third aspect, an embodiment of the present application provides a message pushing device, the device comprising: an acquisition unit, a first processing unit, a second processing unit, a judgment unit, and a pushing unit; the acquisition unit is used to search and acquire multiple messages related to the information to be paid attention to from multiple preset source services according to the information to be paid attention to preset by the user; the first processing unit is used to input each of the messages into a large language model, and generate a summary text corresponding to each of the messages through the large language model; the second processing unit is used to vectorize the summary text corresponding to each of the messages to obtain a text vector corresponding to each of the messages; the judgment unit is used to retrieve and judge the similarity between the text vector corresponding to the message and the database text vector stored in the vector database in turn for each of the text vectors corresponding to the message; wherein the vector database stores multiple database text vectors, and the database text vector is the text vector corresponding to the pushed message; the pushing unit is used to determine the target message from the message based on the similarity retrieval result of the text vector corresponding to the message, store the text vector corresponding to the target message in the vector database, and push the summary text corresponding to the target message to the user.

[0009] In a fourth aspect, an embodiment of the present application provides an electronic device, comprising: a memory and a processor; the memory is used to store one or more computer instructions; the processor is used to execute the one or more computer instructions to implement the above method.

[0010] In a fifth aspect, an embodiment of the present application provides a computer-readable storage medium having one or more computer instructions stored thereon, which, when executed by a processor, executes the above method.

[0011] Compared with the prior art, the message push method provided by the present application includes: searching and obtaining multiple messages related to the information to be paid attention to from multiple preset information source services according to the information to be paid attention to preset by the user; inputting each message into a pre-trained large language model, and generating a summary text corresponding to each message through the large language model; vectorizing the summary text corresponding to each message to obtain a text vector corresponding to each message; searching and judging the similarity between the text vector corresponding to the message and the database text vector stored in the vector database in turn for the text vector corresponding to each message; based on the similarity retrieval result of the text vector corresponding to the message, determining the target message from the message, storing the text vector corresponding to the target message in the vector database, and pushing the summary text corresponding to the target message to the user. First, the method is an automated message push method, which does not require manual participation and does not consume labor costs. Second, the method can search for messages related to the information to be paid attention to from the preset information source service according to the information to be paid attention to preset by the user, and at the same time meets the customization of the message direction and the source of the message, and enhances the user experience. Third, the method applies the large language model to the processing of the message text, realizes the understanding of the actual content of the message, and summarizes the actual content of the message. Fourth, the method realizes the screening of duplicate messages by vectorizing the summary text of the message and setting a vector database for storing the text vectors of the pushed messages. Therefore, the message push method provided by the present application solves the technical problems of the prior art, such as high labor cost, inability to achieve customization, inability to automatically summarize the message content, and easy occurrence of repeated push. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is an application system diagram of the message push method provided in the embodiment of the present application;

[0013] Figure 2 is a flow chart of the message push method provided by the first embodiment of the present application;

[0014] Figure 3 is a schematic diagram of a message briefing provided in the first embodiment of the present application;

[0015] Figure 4 It is a schematic diagram of a message push page provided in the first embodiment of the present application;

[0016] Figure 5 It is a flowchart of the message push method provided in the first embodiment of the present application;

[0017] Figure 6 is a schematic diagram of a message push system provided in a second embodiment of the present application;

[0018] Figure 7is a structural diagram of a message push device provided in the third embodiment of the present application;

[0019] Figure 8 It is a schematic diagram of the structure of an electronic device provided in the fourth embodiment of the present application. DETAILED DESCRIPTION

[0020] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0021] With the development of information technology and the Internet, Internet information has shown explosive growth, and message push has gradually become an important way for users to easily obtain Internet information. Message push platforms usually obtain information from various Internet service platforms and push it to users through the client of the message push platform, saving users the process of searching for information from various service platforms.

[0022] At present, message push mostly relies on manual search on various Internet service platforms, and then manually screening and summarizing the searched messages to form message briefs and push them to users. This undoubtedly increases the labor cost of message push. And due to the limitations of manual search capabilities, the pushed messages cannot be customized, that is, the messages pushed to users may not be the direction of interest to users, and users still need to spend time searching for messages of interest by themselves, resulting in a poor user experience of the message push platform.

[0023] In order to solve the problems existing in the manual message push method, an automated message push method has emerged. However, the current automated message push method is relatively crude and mainly relies on web crawlers to crawl links of relevant messages on various service platforms and push the links to users, so that users can enter various service platforms based on the links to read the corresponding messages. Since the links of relevant messages crawled by web crawlers cannot understand the actual content of the messages, on the one hand, it is impossible to process the messages, such as summarizing the messages, which often causes users to spend a lot of time reading the entire content of the messages. On the other hand, it is also easy to push repeatedly, that is, the same messages from different links are pushed to users, causing users to read the same content repeatedly.

[0024] In view of this, the present application provides a message push method. First, the method is an automated message push method that does not require human participation and does not consume labor costs. Second, the method can search for messages related to the information of interest from the preset source service based on the user's preset information to be paid attention to, while satisfying the customization of the message direction and source, and enhancing the user experience. Third, the method applies a large language model to the processing of message text, realizes the understanding of the actual content of the message, and summarizes the actual content of the message. Fourth, the method realizes the screening of duplicate messages by vectorizing the summary text of the message and setting a vector database that stores the text vectors of the pushed messages.

[0025] The message push method, system, device, electronic device, and computer-readable storage medium described in the present application are further described in detail below in conjunction with specific embodiments and drawings.

[0026] Figure 1 is an application system diagram of the message push method provided in the embodiment of the present application. Figure 1 As shown, the system includes a user terminal 101 and a server terminal 102, and the user terminal 101 can be any device such as a smart phone, a tablet computer, a laptop computer, a desktop computer, a personal digital assistant (PDA), etc. The server terminal 102 can be an independent server, deploying the message push method provided by the present application, or it can be a server group composed of multiple servers, each of which deploys a module of the message push method provided by the present application. For example: the server group may include a message acquisition server, a message processing server, a message push server, etc. Of course, the server terminal 102 can also be a cloud server, and the message push method provided by the present application is deployed on the cloud server, so that the content of interest to the user is sent to the user terminal 101 in the form of message push based on the method on a regular basis (e.g., daily), so that the user can quickly understand the content of interest to him / her based on a message push service platform without searching by himself / herself.

[0027] The first embodiment of the present application provides a message push method, which is deployed in Figure 1 The server 102 shown is used to push messages to users.

[0028] Figure 2 is a flow chart of the message push method provided in this embodiment. Figure 2 The message push method provided in this embodiment is described in detail. The embodiments described below are used to explain the technical solution of this application and are not intended to be used as limitations for actual use.

[0029] like Figure 2As shown, the message push method provided in this embodiment includes the following steps S210 to S250:

[0030] Step S210, based on the information to be concerned preset by the user, searching and acquiring multiple messages related to the information to be concerned from multiple preset information source services.

[0031] The information to be paid attention to is user-defined. Optionally, when the user uses the message push service corresponding to the method for the first time, the server 102 sends multiple alternative information types to the user 101. The user 101 presents the multiple alternative information types to the user through the screen in the form of optional options. Specifically, a window interface for customizing the information to be paid attention to is presented on the screen. The user selects at least one alternative information type in the window interface. The user 101 sends the alternative information type selected by the user to the server 102. The server 102 determines the received alternative information type as the information to be paid attention to preset by the user. Optionally, the window interface also includes an input box. When the multiple alternative information types presented in the window interface lack the information type that the user is interested in, the user can enter the information type that he is interested in in the input box. The user 101 sends the information type entered by the user to the server 102. The server 102 determines the received information type as the information to be paid attention to preset by the user. The customization method of the information to be paid attention to is not limited here.

[0032] In a specific implementation, the information to be paid attention to preset by the user can be one or more. However, no matter how many pieces of information to be paid attention to there are, the method provided in this embodiment can be used to process them sequentially or synchronously, so that all the content that the user is interested in can be pushed to the user within a message push service platform.

[0033] The source service can be understood as the source platform of messages related to the information to be paid attention to. The source service can be some professional websites (such as official websites of some professional companies, famous universities, and research institutes), public accounts, subscription accounts, etc., and there is no specific limitation here. In an optional implementation, the source service can be the default of the server, that is, when the method provided by this embodiment is deployed on the server, the URLs and parsing methods of multiple source services are deployed on the server at the same time. When the user's information to be paid attention to is determined, the server will use the source service related to the information to be paid attention to as the source of message acquisition. Of course, during the maintenance and update process of the method provided in this embodiment, available sources can also be added to the server to expand the sources of message acquisition. In another optional implementation, the source service can also be selected or added by the user, that is, when the user uses the message push service corresponding to the method for the first time, the server 102 can send multiple alternative source services related to the information to be paid attention to to the user terminal 101 according to the information to be paid attention to customized by the user, and the user terminal 101 presents these alternative source services to the user through the screen in the form of optional options. Specifically, a window interface for customizing the source service is presented on the screen, and the user selects at least one alternative source service in the window interface. The user terminal 101 sends the alternative source service selected by the user to the server terminal 102, and the server terminal 102 determines the received alternative source service as the preset source service. Optionally, the window interface also includes an input box. When the multiple alternative source services presented in the window interface lack the information service that the user is concerned about, the user can enter the website address and resolution method of the source service he is concerned about in the input box, and the user terminal 101 sends the website address and resolution method of the source service entered by the user to the server terminal 102, and the server terminal 102 determines the source service as the preset source service. Optionally, if the user cannot provide a parsing method, the server 102 will help the user obtain the parsing method of the source service. Optionally, the window interface also includes a default control. When the user does not want to customize the source service, the default control can be triggered to enable the server 102 to use multiple candidate source services related to the information to be followed or the first few with the highest popularity among the multiple candidate source services as the preset source services. The customization method of the source service is not limited here.

[0034] Step S220: input each message into the large language model, and generate a summary text corresponding to each message through the large language model.

[0035] The large language model refers to an artificial intelligence system with a large number of parameters obtained through machine learning training, which is specifically used to process natural language tasks. The model contains hundreds of millions to trillions of parameters, which enables it to perform well in processing complex language structures, can generate or understand language based on context, and provide a more human-like understanding and generation capability. It can perform new tasks without additional training (zero-sample learning), or only require a small number of examples to quickly learn new tasks (few-sample learning). In this embodiment, the large language model can be a mature large language model, or it can be obtained by training a general large language model with a few samples.

[0036] In a specific implementation, in order to make the output content of the large language model meet the requirements of developers, the large language model will be fine-tuned based on prompt words before application so that the large language model understands the content that needs to be output based on the input message, such as the summary text corresponding to the output message.

[0037] Step S230, vectorizing the summary text corresponding to each message to obtain a text vector corresponding to each message.

[0038] Vectorization can be understood as the process of converting text into vectors, i.e., text embedding vectors, which can be optionally operated through vectorization models. In a specific implementation, the summary text corresponding to the message is vectorized using the Bge-large-zh model, which is a text embedding model that can convert the input natural language into a vector, i.e., a long string of numbers, to facilitate subsequent processing by the computer.

[0039] In the method provided in this embodiment, the message itself can be processed using a vector model, that is, the entire text of the message is converted into a vector. However, since the number of characters in the entire text of the message is too large, the length of the converted vector will also be long, which may affect the subsequent vector-based processing steps and make the vector processing result inaccurate. Therefore, in a specific implementation, the summary text corresponding to the message is used as input data, input into the vector model, and the vector output by the vector model for the summary text is obtained. In this embodiment, the vector converted based on the summary text is defined as a text vector. On the one hand, since the summary text is the summary content obtained by understanding the message through a large language model, the summary text can better represent the core content of the message. Therefore, the text vector output for the summary text can represent the message itself; on the other hand, since the number of characters in the summary text is much smaller than the entire text of the message, using the text vector corresponding to the summary text is more conducive to subsequent operations than using the vector corresponding to the entire text of the message.

[0040] Step S240, for each text vector corresponding to the message, retrieve and determine the similarity between the text vector corresponding to the message and the database text vector stored in the vector database in turn; wherein the vector database stores a plurality of database text vectors, and the database text vector is the text vector corresponding to the pushed message.

[0041] In the method described in this embodiment, a vector database is provided, which is used to store text vectors. These text vectors are text vectors corresponding to messages that have been pushed to users. In this embodiment, the text vectors stored in the vector database are defined as database text vectors. In an optional implementation, the message that has been pushed to the user can be a message pushed to the user in the historical push stage. For example, the message push service corresponding to the method described in this embodiment is a daily push service, then every day before the current push day belongs to the historical push stage. In another optional implementation, the message that has been pushed to the user can be a message pushed to the user in the current push stage. For example, the message that has been processed and pushed to the user among the multiple messages searched and determined on the current push day. Optionally, when the push mode is different, such as after all the multiple messages that have been searched and obtained are processed, a briefing is made and pushed to the user. In this push mode, the message that has been pushed to the user can be a message that has been determined to be pushed to the user in the current push stage, that is, a message that has been processed and determined to be pushed to the user.

[0042] In an optional implementation, a nearest neighbor algorithm can be used to calculate the similarity between the text vector corresponding to the message and each database text vector in the vector database. The nearest neighbor algorithm is an instance-based learning method that determines the similarity between different data instances by measuring the distance between them. In this embodiment, the text vector corresponding to the message can be used as a new instance, and the nearest neighbor algorithm can be used to quickly compare the new instance with multiple database instances (database text vectors) in the vector database for similarity, and select the first number of database instances (database text vectors) with the highest similarity.

[0043] Step S250, based on the similarity search results of the text vectors corresponding to the messages, determine the target message from the messages, store the text vectors corresponding to the target message in a vector database, and push the summary text corresponding to the target message to the user.

[0044] The target message refers to a non-repetitive message. Specifically, based on the similarity search result of the text vector corresponding to the message, it can be determined whether there is a database text vector similar to the text vector in the vector database, and then determine whether the message corresponding to the text vector is a non-repetitive message. Based on this, in this embodiment, a first preset similarity threshold is provided, and the threshold is used as a standard for judging whether the text vector is similar to the database text vector. The first preset similarity threshold is an adjustable value, which can be adjusted according to the type of information to be paid attention to preset by the user, the push results of the historical push stage, etc., and there is no specific limitation. Exemplarily, the first preset similarity threshold is set to 99%.

[0045] In an optional implementation, based on the similarity search results of the text vectors corresponding to the messages, determining the target message from the messages may specifically include the following steps S251-1 to S251-3:

[0046] Step S251 - 1 , based on the similarity search results of the text vectors corresponding to the message, obtain a first number of database text vectors with the highest similarity.

[0047] Step S251-2, in response to the similarity between each of the first number of database text vectors and the text vector corresponding to the message being less than a first preset similarity threshold, it is determined that there is no database text vector in the vector database that is similar to the text vector corresponding to the message.

[0048] Step S251 - 3 , in response to the absence of a database text vector similar to the text vector corresponding to the message in the vector database, determining the message as a target message.

[0049] Exemplarily, a search is performed in the vector database for the text vector a corresponding to message A. When the similarities between text vector a and database text vector 1, database text vector 2, database text vector 3, database text vector 4, and database text vector 5 are all less than a first preset similarity threshold (assumed to be 99%), it means that there is no message that is substantially identical to message A among the messages that have been pushed to the user, and message A is not a duplicate message. Message A can be determined as the target message and pushed to the user.

[0050] In another optional implementation, based on the similarity search results of the text vectors corresponding to the messages, determining the target message from the messages may specifically include the following steps S252-1 to S252-3:

[0051] Step S252-1, based on the similarity search results of the text vectors corresponding to the message, obtain a first number of database text vectors with the highest similarity.

[0052] Step S252-2, in response to the similarity between any database text vector in the first number of database text vectors and the text vector corresponding to the message being greater than or equal to a first preset similarity threshold, determining whether there is a database text vector in the vector database that is similar to the text vector corresponding to the message.

[0053] Step S252-3, in response to the presence of a database text vector in the vector database that is similar to the text vector corresponding to the message, but the time interval between the acquisition time of the message and the acquisition time of the pushed message corresponding to the database text vector is greater than or equal to the first preset time threshold, the message is determined as a target message.

[0054] Correspondingly, in response to the presence of a database text vector in the vector database that is similar to the text vector corresponding to the message, and the time interval between the acquisition time of the message and the acquisition time of the pushed message corresponding to the database text vector is less than a first preset time threshold, the message is determined to be a duplicate message and deleted.

[0055] Exemplarily, assuming that there are 100 database text vectors stored in the vector database, it is necessary to determine whether there is a database text vector similar to the text vector a corresponding to the message A in the vector database. First, the similarity between the text vector a and the 100 database text vectors is calculated. Secondly, based on the calculated similarity, 5 (the first number) database text vectors with the highest similarity are selected from the 100 database text vectors. It is assumed that the 5 selected database text vectors with the highest similarity are database text vector 1, database text vector 2, database text vector 3, database text vector 4, and database text vector 5. Again, determine the size of the similarity between text vector a and database text vector 1, database text vector 2, database text vector 3, database text vector 4, and database text vector 5 and the first preset similarity threshold (assuming 99%); when it is determined that the similarity between text vector a and database text vector 1, database text vector 2, database text vector 3, database text vector 4, and database text vector 5 is less than the first preset similarity threshold, determine that there is no database text vector similar to text vector a in the vector database; when it is determined that the similarity between text vector a and at least one of database text vector 1, database text vector 2, database text vector 3, database text vector 4, and database text vector 5 is greater than or equal to the first preset similarity threshold (assuming that the similarity between text vector a and database text vector 1 is 99.5%), determine that there is a database text vector similar to text vector a in the vector database. In the case where there is no database text vector similar to text vector a in the vector database, message A can be determined as the target message. In the case where there is a database text vector similar to text vector a in the vector database, it is necessary to continue to determine the time interval between the acquisition time of message A and the acquisition time of the pushed message 1 corresponding to database text vector 1. Specifically, when the time interval between the acquisition time of message A and the acquisition time of message 1 is greater than or equal to the first preset time threshold, message A is determined as the target information; when the time interval between the acquisition time of message A and the acquisition time of message 1 is less than the first preset time threshold, message A is determined as a duplicate message and message A is deleted.

[0056] It is understandable that each message has its life cycle, from the occurrence of the event to media coverage, to public attention, and finally to the gradual decrease of public attention. Therefore, the push of the same message by each source service basically occurs within the life cycle of the message, that is, repeated push usually occurs within the life cycle of the message, such as one week, two weeks, etc. After the life cycle, the possibility of each source service to push the message again is greatly reduced, almost non-existent. Therefore, in the method provided in this embodiment, in addition to using whether there is a database text vector similar to the text vector corresponding to the message in the vector database as a benchmark for judging whether the message is a repeated message, a time threshold is added based on the common phenomenon that the message has a life cycle, so as to determine whether the database text vector similar to the text vector corresponding to the message exists in the vector database. As another benchmark for judging whether the message is a repeated message. In this embodiment, the time threshold is defined as a first preset time threshold. Specifically, the first preset time threshold is preset according to the general law of the message life cycle. Of course, it can also be adjusted according to the message push requirements and message push results, and there is no specific limitation.

[0057] Exemplarily, assume that the first preset time threshold is 2 weeks, that is, 14 days. Search the text vector a corresponding to message A in the vector database. When no database text vector similar to text vector a is found in the vector database, it is determined that message A is not a duplicate message and message A is pushed to the user. When a database text vector 1 similar to text vector a is found in the vector database, and the time interval between the acquisition time of message A corresponding to text vector a and the acquisition time of message 1 corresponding to database text vector 1 is 3 weeks, since 3 weeks is greater than 2 weeks, it is determined that message A is not a duplicate message and message A is pushed to the user. When a database text vector 2 similar to text vector a is found in the vector database, and the time interval between the acquisition time of message A corresponding to text vector a and the acquisition time of message 2 corresponding to database text vector 2 is 1 week, since 1 week is less than 2 weeks, it is determined that message A is a duplicate message and message A is deleted and not pushed to the user.

[0058] In an optional implementation, when the push method is different, for example, after all the multiple messages searched and obtained are processed, a briefing is made and pushed to the user. In this push method, the messages determined to be duplicate messages may not be directly deleted. After all the multiple messages searched and obtained are processed, the duplicate messages ranked lower in terms of popularity, attention, and professionalism may be deleted based on the popularity of each duplicate message in the search engine, or the attention and professionalism of the source service corresponding to each duplicate message, and only the duplicate message ranked first is retained and pushed to the user in the form of a briefing.

[0059] For example, suppose that among the multiple messages searched and obtained, message 1, message 2, message 3, and message 4 are duplicate messages with the same core content. According to the popularity judgment, the popularity of message 2 is greater than that of message 1, message 3, and message 4. Therefore, message 1, message 3, and message 4 are deleted, and message 2 is pushed to the user.

[0060] In this embodiment, after the target message is determined, the text vector corresponding to the target message is stored in the vector database as a database text vector, so as to determine whether other messages are duplicate messages of the target message.

[0061] In this embodiment, after the target message is determined, the summary text corresponding to the target message is pushed to the user, so that the user can quickly understand the core content of the message based on the summary text without having to read the full text of the message.

[0062] In another optional implementation provided by the present embodiment, based on the vector database, it is not only possible to screen whether a message is a duplicate message, but also possible to provide support for historical related messages for the messages that have been determined to be sent to the user, so that when the user views the currently pushed message, he can also view the historical pushed messages associated with the message, thereby enhancing the understanding of the content of the message and the understanding of the development history of the technology corresponding to the message.

[0063] Based on this, before the step of pushing the summary text corresponding to the target message to the user, the method provided in this embodiment may further include the following steps: based on the vector database, determining whether there is a pushed message related to the target message.

[0064] It can be understood that similarity refers to the degree of similarity between a message and a pushed message, and relevance refers to the degree of relevance between a message and a pushed message. Similar messages must be related, and related messages will also have a certain degree of similarity in their text vectors. Specifically, for text vectors, the vector consistency corresponding to similarity is greater than the vector consistency corresponding to relevance. This embodiment provides an optional method for determining the degree of relevance between a message and a pushed message, using vector similarity as the first screening step, specifically, screening out several database text vectors that have a certain degree of similarity with the text vector corresponding to the message; using text relevance as the second screening step, specifically, screening out the pushed messages with the highest relevance to the message from several pushed messages corresponding to several database text vectors. Based on this method, efficient screening of pushed messages related to the message can be achieved, and the number of screenings can be one or more, without specific limitation.

[0065] Based on this, in a specific implementation, judging whether there is a pushed message related to the target message based on the vector database may specifically include the following steps S11 to S13:

[0066] Step S11, calculating the similarity between the text vector corresponding to the target message and each database text vector.

[0067] Step S12, obtaining a second number of database text vectors with the highest similarity.

[0068] Step S13, in response to the similarity between any database text vector among the second number of database text vectors and the text vector corresponding to the target message being greater than or equal to a second preset similarity threshold, and the time interval between the acquisition time of the target message and the acquisition time of the pushed message corresponding to the database text vector being greater than or equal to the second preset time threshold, it is determined that there is a pushed message related to the target message.

[0069] Optionally, a nearest neighbor algorithm can be used to calculate the similarity between the text vector corresponding to the target message and each database text vector, and the second number of database text vectors with the highest similarity can be selected. The second number is an adjustable value and can be set according to the evaluation requirements of the relevant message, and is not limited here.

[0070] In this embodiment, the second preset similarity threshold is used as a benchmark for determining whether the text vector and the database text vector are similar to a certain extent. The second preset similarity threshold is an adjustable value, which can be adjusted according to the type of information to be paid attention to preset by the user, the push results of the historical push stage, etc., and is not specifically limited. Exemplarily, the second preset similarity threshold is set to 80%, which is less than the first preset similarity threshold used to determine whether the text vector and the database text vector are similar.

[0071] In this embodiment, in order to prevent the filtered pushed messages from being duplicate messages with the target message, the requirements for the acquisition time of the target message and the pushed message are added to the criteria for filtering the pushed messages related to the target message. It can be understood that when the similarity between the text vector corresponding to the target message and the database text vector corresponding to the pushed message is greater than the first preset similarity threshold (e.g., 99%), the similarity between the text vector corresponding to the target message and the database text vector corresponding to the pushed message must be greater than the second preset similarity threshold (e.g., 80%). At this time, if the time interval between the acquisition time of the target message and the acquisition time of the pushed message is too small (e.g., less than two weeks), the two messages are likely to be duplicate messages. Therefore, in this embodiment, the pushed messages whose similarity is greater than or equal to the second preset similarity threshold and whose time interval between the acquisition times is greater than or equal to the second preset time threshold are regarded as the pushed messages related to the target message.

[0072] The second preset time threshold is preset according to the general rules of the message life cycle. Generally, the second preset time threshold is the same as the first preset time threshold. Of course, it can also be adjusted according to the push requirements for related messages and the push results of related messages.

[0073] Based on this, the summary text corresponding to the target message is pushed to the user, which can be specifically: in response to the existence of a pushed message related to the target message, the pushed message is associated with the summary text corresponding to the target message and presented on the push page of the target message, and the push page is pushed to the user. The push page can be understood as a carrier for carrying the target message, and the push page can carry the target message in the form of a picture, a combination of text boxes, etc. In this embodiment, the push page not only carries the summary text corresponding to the target message, but also carries the related messages corresponding to the target message. Exemplarily, the title or summary text of the related message is presented in a linkable form below the summary text corresponding to the target message.

[0074] In an optional implementation, since there may be multiple pushed messages whose similarity is greater than or equal to the second preset similarity threshold and whose time interval between acquisition times is greater than or equal to the second preset time threshold, in order to further push the most relevant pushed messages to the user, the multiple pushed messages may be re-arranged in terms of relevance at the text level, thereby selecting the first few pushed messages with the highest relevance and pushing them to the user. Specifically, the following steps S21 to S22 may be included:

[0075] Step S21, sorting the pushed messages related to the target message by relevance.

[0076] Step S22: associate the third number of pushed messages with the highest relevance with the summary text corresponding to the target message and present them on a push page of the target message, and push the push page to the user.

[0077] Optionally, a reranking model (bge-reranker-large model) can be used to rerank the relevance of several pushed messages related to the target message. The reranking model can perform a semantic comparison between the target message and the pushed message text based on semantic understanding and calculate the relevance between the two. The higher the relevance, the higher the ranking.

[0078] Rearranging several played messages related to the target message at the text level can be understood as a refined sorting process. After sorting, the third number of pushed messages with the highest sorting (i.e., the highest relevance) can be associated with the summary text corresponding to the target message and pushed to the user. The third number is an adjustable value, which can be set according to the number of associated pushed messages, and is not limited here. Associated playback can take many forms, such as presenting a linkable title of the pushed message below the summary text corresponding to the target message, presenting a link logo next to the summary text corresponding to the message, etc., without specific limitation.

[0079] Exemplarily, assuming that 100 database text vectors are stored in the vector database, it is necessary to determine whether there is a played message related to message A based on the vector database. Calculate the similarity between the text vector a corresponding to message A and the 100 database text vectors. According to the calculated similarity, select 8 (second number) database text vectors with the highest similarity from the 100 database text vectors. Assuming that the similarity between database text vector 1, database text vector 3, database text vector 5, database text vector 7 and text vector a among the 8 database text vectors is greater than 80% (second preset similarity threshold), and the time interval between the acquisition time of message A and the acquisition time of pushed message 1 corresponding to database text vector 1, pushed message 3 corresponding to database text vector 3, pushed message 5 corresponding to database text vector 5, and pushed message 7 corresponding to database text vector 7 is greater than 2 weeks (second preset time threshold), it is determined that there is a pushed message related to message A. Next, a reranking model (e.g., bge-reranker-large model) is used to rerank the played messages 1, 3, 5, and 7 related to message A. The reranking result is that message 7 is greater than message 1, greater than message 5, and greater than message 3. Based on this, the two (third number) pushed messages with the highest relevance, i.e., message 7 and message 1, are pushed to the user as the associated text of message A.

[0080] In an optional implementation, determining duplicate messages and determining related messages based on a vector database can be performed simultaneously. Specifically, the similarity between the text vector corresponding to the message and each database text vector is calculated, and the M (M is a positive integer, and the first number is the same as the second number) database text vectors with the highest similarity are selected to determine whether there is a database text vector among the M database text vectors whose similarity with the text vector is greater than or equal to 99% (the first preset similarity threshold), and the time interval between the acquisition of the corresponding messages is less than t (the first preset time threshold is the same as the second preset time threshold, such as 2 weeks). If yes, determine that the message is a duplicate message and delete the message. If not, determine whether there is a database text vector among the M database text vectors whose similarity with the text vector is greater than or equal to 80% (the second preset similarity threshold), and the time interval between the acquisition of the corresponding messages is greater than or equal to t (the first preset time threshold is the same as the second preset time threshold, such as 2 weeks). If not, determine that the message has no related pushed messages, and push the summary text corresponding to the message to the user. If yes, the relevance between the message and the pushed messages corresponding to several qualified database text vectors is sorted, and the N (positive integer, less than M, the third number) pushed messages with the highest relevance are associated with the summary text corresponding to the messages and pushed to the user.

[0081] In an optional implementation provided by this embodiment, inputting each message into the large language model, and generating a summary text corresponding to each message by the large language model may include the following steps S31 to S32:

[0082] Step S31, generating prompt words based on the message.

[0083] The prompt refers to a piece of text that guides the large language model to generate specific content. It can be regarded as a way to interact with the large language model. The prompt tells the large language model what kind of content should be generated. Optionally, the prompt includes at least the following instructions: generate a summary text of the message, delete content in the message that is not related to the information to be paid attention to, and generate a comment text with a preset style for the message.

[0084] Among them, the summary text of the generated message is used to guide the large language model to generate the corresponding summary text according to the input message, and the content of the message that is irrelevant to the information to be paid attention to is deleted to guide the large language model to remove the content of the message that is irrelevant to the information to be paid attention to, such as advertising information, harmful information, etc., and the comment text of the generated message is used to guide the large language model to generate the corresponding comment text according to the input message. Optionally, by adding specific personality restrictions in the prompt words, the comment text can show a style corresponding to the set personality to increase the interest, that is, the style is determined by the prompt words.

[0085] For example, taking the information to be paid attention to as artificial intelligence (AI) as an example, the prompt words for fine-tuning the large language model can be as shown in Table 1:

[0086] Table 1 Prompt words

[0087]

[0088] Step S32, inputting the prompt word into the large language model, deleting the content in the message that is irrelevant to the information to be paid attention to through the large language model, and generating a summary text corresponding to the message with the irrelevant content deleted, and a comment text with a preset style.

[0089] Based on this, the summary text corresponding to the target message is pushed to the user, which may specifically include: the summary text corresponding to the target message and the comment text of the preset style are presented in association on the push page of the target message, and the push page is pushed to the user. Optionally, the push page carries the summary text corresponding to the target message, the related message corresponding to the target message, and the comment text corresponding to the target message. Exemplarily, the title or summary text of the related message is presented in a linkable form below the summary text corresponding to the target message, and the comment text is presented in the form of a replyable text box below the related message.

[0090] In an optional implementation provided by this embodiment, before the step of pushing the summary text corresponding to the target message to the user, the method provided by this embodiment may further include the following steps S41 to S42:

[0091] Step S41, obtaining multiple pictures included in the target message.

[0092] Step S42, obtaining the picture with the highest information content from multiple pictures, and using the picture with the highest information content as the picture to be pushed corresponding to the target message.

[0093] Taking into account that only pushing the summary text of the target message to the user will make the pure text push page lack interest, based on this, in this embodiment, the links of all the picture tags in the target message are obtained, so as to obtain all the pictures corresponding to the target message, and by determining the text information corresponding to each picture, the picture with the highest amount of information is selected as the main picture corresponding to the target message, that is, the picture to be pushed corresponding to the target message.

[0094] In this embodiment, the amount of information can be understood as the amount of text information that can be obtained from the image, and the greater the amount, the higher the amount of information of the image. Based on this, the amount of information corresponding to the image can be determined by parsing the image, extracting the text information corresponding to the image, and calculating the amount of text information.

[0095] In an optional implementation, the amount of information corresponding to the image can be determined by a multimodal large language model, which specifically includes: inputting the image into the multimodal large language model, parsing the image with the multimodal large language model, outputting text information corresponding to the image, calculating the amount of text information, and using the amount of text information as the amount of information corresponding to the image.

[0096] In another optional implementation, the text information in the image can be obtained by the OCR method, and then the amount of information corresponding to the image can be determined by calculating the amount of text information. OCR (Optical Character Recognition) enables a computer to recognize and extract text from an image or document scan. The following steps are included: image capture (using a scanner or digital camera to obtain an image. In this embodiment, the image is all the images corresponding to the acquired message); preprocessing (cleaning the image, removing noise, and may also include tilt correction, binarization and other steps); feature extraction (extracting the features of the characters from the image); recognition (using a trained algorithm or model to identify the specific characters represented by these features); post-processing (correcting errors and optimizing the output text). In a specific implementation, the more characters contained in the text information extracted from the image, the greater the amount of text information, and the higher the information content of the image.

[0097] Based on this, the summary text corresponding to the target message is pushed to the user, which may specifically include: presenting the summary text corresponding to the target message and the image to be pushed on the push page of the target message, and pushing the push page to the user. Optionally, the push page carries the summary text corresponding to the target message, the related information corresponding to the target message, the comment text corresponding to the target message, and the image to be pushed corresponding to the target message. Exemplarily, the image to be pushed is presented below the summary text corresponding to the target message, the title or summary text of the related message is presented in a linkable form below the image to be pushed, and the comment text is presented in a replyable text box below the related message, so as to push the target message with pictures and texts to the user.

[0098] In an optional implementation, the determined target message will first be pushed to the user in the form of a briefing. Specifically, the title corresponding to the target message is entered into a briefing template to form a message briefing including titles corresponding to multiple target messages; the message briefing is pushed to the user; in response to the user's viewing instruction for the first target message based on the message briefing, a summary text corresponding to the first target message is pushed to the user, wherein the first target message is any one of the multiple target messages presented in the message briefing.

[0099] In a specific implementation, in order to save the time of users viewing messages, the heat value of each target message is obtained, and the titles corresponding to the multiple target messages included in the message brief are sorted according to the heat values ​​corresponding to the multiple target messages, and the titles corresponding to the target messages whose heat values ​​are greater than the preset heat threshold are marked with heat marks, and the sorted message briefs marked with heat marks are pushed to the user. The heat threshold is an adjustable value, which is a reference standard for judging whether the title corresponding to the target message in the message brief needs to be marked with a heat mark. Specifically, when the heat value of a certain target message is greater than the heat threshold, it means that the target message is a message with a higher heat at present, and the title corresponding to the target message can be marked with a heat mark. The number of heat marks marked is related to the heat value of the target message. Specifically, the higher the heat value, the more heat marks are marked (e.g., 3 heat marks), and the lower the heat value, the fewer heat marks are marked (e.g., 1 heat mark). The heat mark can include a variety of presentation methods, such as spark icons, star icons, etc., and marking the heat mark can further increase the attractiveness of the target message with a higher heat. Optionally, the popularity value corresponding to the target message may be determined by crawling the popularity of related information in a search engine.

[0100] Figure 3 It is a schematic diagram of a message briefing provided in this embodiment.

[0101] like Figure 3 As shown, the message brief includes titles corresponding to multiple target messages determined based on the above method, and these titles are sorted according to the popularity of the messages, and the title corresponding to the target message with the highest popularity is ranked first. The popularity mark of the "spark icon" is added before the title to increase the attractiveness of the message with higher popularity.

[0102] Figure 4 It is a schematic diagram of a message push page provided in this embodiment.

[0103] like Figure 4 As shown, when the user Figure 3 When you identify a message you want to read in detail in the displayed message briefing, you will obtain the push page of the message by triggering the title corresponding to the message. Figure 4 The push page shown corresponds to Figure 3 The push page includes a title 401 corresponding to the message, a summary text 402 corresponding to the message, a main picture 403 corresponding to the message, related historical push messages 404 corresponding to the message, and a comment text 405 corresponding to the message.

[0104] The above steps provide an optional message push method. First, the method allows users to customize the information and source services they are interested in. As more and more source services are customized by users, a data flywheel effect is formed. The more sources users can choose from, the wider the types of information they can choose. Second, the method has a memory function. By vectorizing the storage of previously pushed messages and vectorizing the retrieval of new messages, it can not only help users recall relevant messages, but also filter repeated messages from different source services to avoid repeated pushes. Third, the method can remove advertisements and harmful information from messages by adjusting the instructions of the prompt words. Fourth, the method can filter out representative main images corresponding to the message by confirming the amount of information corresponding to the picture, so as to avoid the boring feeling caused to users by pure text. Fifth, the method sorts and annotates the push messages according to the search engine popularity, so that users can selectively read the push page of the message. Sixth, the method has a deep understanding of the message content based on the large language model, can summarize the core content of the message, and make stylized comments on the message.

[0105] The following is an optional message push method. Figure 5 It is a flowchart of the message push method provided in this embodiment.

[0106] like Figure 5 As shown, the message push method provided in this embodiment may also include the following steps S501 to S508:

[0107] Step S501, collect messages related to the information to be concerned preset by the user from multiple information source services at regular intervals. For example, if the information of the day is pushed to the user at 8:10 every day, then all messages related to the information to be concerned pushed from multiple information source services from 8:00 the previous day to 8:00 the current day can be collected at 8:00.

[0108] Step S502, for each message, use a large language model to perform text processing on the message, and generate a message list corresponding to each message. The message list can be understood as a collection of message-related content. Specifically, text processing may include translating the message (e.g., translating an English message into Chinese), concisely summarizing the message, etc., to convert the message into a message list. Optionally, the message list may include the title corresponding to the message, the source service link corresponding to the message, the author corresponding to the message, the summary text corresponding to the message, the main picture corresponding to the message, etc. Optionally, while the text processing is being performed on the message, the main picture corresponding to the message is determined based on the OCR method, and the main picture is recorded in the message list.

[0109] Step S503, caching the message list corresponding to each message in the text database.

[0110] Step S504, for each message, vectorize the summary text corresponding to the message to generate a text vector corresponding to each message.

[0111] Step S505, based on the vector database, a similarity search is performed on each message to delete duplicate messages and related historical push messages corresponding to each non-duplicate target message.

[0112] Step S506, caching the message list corresponding to the target message in the text database.

[0113] Step S507: for each target message, a large language model is used to perform text processing to generate interesting comments corresponding to each target message.

[0114] Step S508, generating a message push page for presentation to the user. Specifically, the message push page includes a title corresponding to the target message, a summary text corresponding to the target message, a main picture corresponding to the target message, related historical push messages corresponding to the target message, and interesting comments corresponding to the target message.

[0115] In summary, the above-mentioned first embodiment provides a message push method. On the one hand, the method is an automated message push method, which does not require human participation and does not consume labor costs. On the other hand, the method can search for messages related to the information of interest from the preset source service according to the information to be paid attention to preset by the user, while satisfying the customization of the message direction and the source of the message, and enhancing the user experience. On the third hand, the method applies a large language model to the processing of the message text, realizes the understanding of the actual content of the message, and summarizes the actual content of the message. On the fourth hand, the method realizes the screening of duplicate messages by vectorizing the summary text of the message and setting a vector database for storing the text vectors of the pushed messages. It should be noted that the example in the first embodiment is only for explaining the method described in this application, and is not intended to be a limitation for actual use. The message push method provided in this application includes but is not limited to the method described in the first embodiment.

[0116] The second embodiment of the present application provides a message push system. Figure 6 Schematic diagram of the message push system provided in this embodiment.

[0117] like Figure 6 As shown, the message push system provided in this embodiment includes: a module 601 for customizing information to be followed and information sources, a message collection module 602, a text processing module 603, a text database 604, a vector processing module 605, a vector database 606, and a push module 607.

[0118] The module 601 for customizing the information to be followed and the information source is used to provide the user with a window interface for customizing the information to be followed and the information source service.

[0119] The message collection module 602 is used to search and obtain multiple messages related to the information to be concerned from multiple customized information source services according to the information to be concerned customized by the user.

[0120] The text processing module 603 is used to input each of the messages into a large language model, and generate a summary text corresponding to each of the messages through the large language model.

[0121] The text database 604 is used to store a message list corresponding to each of the messages, and the message list at least records the summary text of the message.

[0122] The vector processing module 605 is used to perform vectorization processing on the summary text corresponding to each of the messages to obtain a text vector corresponding to each of the messages; and to retrieve and determine the similarity between the text vector corresponding to the message and the database text vector stored in the vector database 606 for each of the text vectors corresponding to the message.

[0123] The vector database 606 is used to store database text vectors, where the database text vectors are text vectors corresponding to the pushed messages.

[0124] The push module 607 is used to determine the target message from the messages based on the similarity retrieval results of the text vectors corresponding to the messages, store the text vectors corresponding to the target messages in the vector database 606, and push the summary text corresponding to the target message to the user.

[0125] The third embodiment of the present application provides a message pushing device. Figure 7 It is a structural diagram of the message pushing device provided in this embodiment.

[0126] like Figure 7 As shown, the message pushing device provided in this embodiment includes: an acquisition unit 701, a first processing unit 702, a second processing unit 703, a determination unit 704, and a pushing unit 705.

[0127] The acquisition unit 701 is used to search and acquire multiple messages related to the information to be concerned from multiple preset information source services according to the information to be concerned preset by the user.

[0128] The first processing unit 702 is used to input each of the messages into a large language model, and generate a summary text corresponding to each of the messages through the large language model.

[0129] The second processing unit 703 is used to perform vectorization processing on the summary text corresponding to each of the messages to obtain a text vector corresponding to each of the messages.

[0130] The judgment unit 704 is used to retrieve and judge the similarity between the text vector corresponding to each message and the database text vector stored in the vector database in turn, for each text vector corresponding to the message; wherein the vector database stores multiple database text vectors, and the database text vector is the text vector corresponding to the pushed message.

[0131] Optionally, before the step of pushing the summary text corresponding to the target message to the user, it is also used to:

[0132] Based on the vector database, determining whether there is a pushed message related to the target message;

[0133] The step of pushing the summary text corresponding to the target message to the user includes:

[0134] In response to the existence of the pushed message related to the target message, the pushed message is associated with the summary text corresponding to the target message and presented on a push page of the target message, and the push page is pushed to the user.

[0135] Optionally, the determining, based on the vector database, whether there is a pushed message related to the target message includes:

[0136] Calculating the similarity between the text vector corresponding to the target message and each of the database text vectors;

[0137] Obtaining a second number of the database text vectors having the highest similarity;

[0138] In response to the similarity between any one of the second number of database text vectors and the text vector corresponding to the target message being greater than or equal to a second preset similarity threshold, and the time interval between the acquisition time of the target message and the acquisition time of the pushed message corresponding to the database text vector being greater than or equal to a second preset time threshold, determining that there is the pushed message related to the target message;

[0139] In response to the existence of the pushed message related to the target message, the pushed message is associated with the summary text corresponding to the target message and presented on a push page of the target message, and the push page is pushed to the user, comprising:

[0140] sorting the pushed messages related to the target message by relevance;

[0141] The pushed messages of the third number having the highest relevance are associated with the summary text corresponding to the target message and presented on a push page of the target message, and the push page is pushed to the user.

[0142] The push unit 705 is used to determine the target message from the messages based on the similarity retrieval results of the text vectors corresponding to the messages, store the text vectors corresponding to the target messages in the vector database, and push the summary text corresponding to the target message to the user.

[0143] Optionally, the determining a target message from the messages based on the similarity search result of the text vector corresponding to the messages includes:

[0144] In response to the absence of a database text vector similar to the text vector corresponding to the message in the vector database, the message is determined to be a target message.

[0145] Optionally, also include:

[0146] In response to the presence of a database text vector in the vector database that is similar to the text vector corresponding to the message, but the time interval between the acquisition time of the message and the acquisition time of the pushed message corresponding to the database text vector is greater than or equal to the first preset time threshold, the message is determined to be a target message.

[0147] Optionally, before the step of determining the message as the target message in response to the absence of a database text vector similar to the text vector corresponding to the message in the vector database, the method is further configured to:

[0148] Based on the similarity search results of the text vectors corresponding to the message, obtaining a first number of the database text vectors with the highest similarity;

[0149] In response to the similarity between each of the first number of database text vectors and the text vector corresponding to the message being less than a first preset similarity threshold, it is determined that there is no database text vector in the vector database that is similar to the text vector corresponding to the message.

[0150] Optionally, before the step of determining the message as the target message in response to the presence of a database text vector similar to the text vector corresponding to the message in the vector database, but the time interval between the acquisition time of the message and the acquisition time of the pushed message corresponding to the database text vector is greater than or equal to a first preset time threshold, the method is further configured to:

[0151] Based on the similarity search results of the text vectors corresponding to the message, obtaining a first number of the database text vectors with the highest similarity;

[0152] In response to the similarity between any one of the first number of database text vectors and the text vector corresponding to the message being greater than or equal to a first preset similarity threshold, it is determined that there is a database text vector in the vector database that is similar to the text vector corresponding to the message.

[0153] Optionally, inputting each of the messages into a large language model, and generating a summary text corresponding to each of the messages through the large language model, includes:

[0154] Generate a prompt word based on the message, the prompt word at least including the following instructions:

[0155] Generate a summary text of the message, delete the content in the message that is irrelevant to the information to be paid attention to, and generate a comment text with a preset style for the message; wherein the preset style is determined by the prompt word;

[0156] Input the prompt word into the large language model, delete the content in the message that is irrelevant to the information to be concerned through the large language model, and generate the summary text corresponding to the message with the irrelevant content deleted, and the comment text with a preset style;

[0157] The step of pushing the summary text corresponding to the target message to the user includes:

[0158] The summary text corresponding to the target message and the review text of a preset style are associated and presented on a push page of the target message, and the push page is pushed to the user.

[0159] Optionally, the device further comprises: a picture determination unit;

[0160] Before the step of pushing the summary text corresponding to the target message to the user, the picture determination unit is used to:

[0161] Acquire multiple pictures contained in the target message;

[0162] Acquire the picture with the highest information content from the multiple pictures, and use the picture with the highest information content as the picture to be pushed corresponding to the target message;

[0163] The step of pushing the summary text corresponding to the target message to the user includes:

[0164] The summary text corresponding to the target message and the picture to be pushed are associated and presented on a push page of the target message, and the push page is pushed to the user.

[0165] Optionally, the pushing unit 705 is further used for:

[0166] Entering the title corresponding to the target message into a briefing template to form a message briefing including the titles corresponding to the multiple target messages;

[0167] Pushing the message brief to the user;

[0168] The step of pushing the summary text corresponding to the target message to the user includes:

[0169] In response to the user's viewing instruction for a first target message based on the message brief, the summary text corresponding to the first target message is pushed to the user, and the first target message is any one of the multiple target messages.

[0170] Optionally, the device further comprises: a message heat determination unit;

[0171] After the step of forming a message brief including titles corresponding to a plurality of target messages, the message popularity determination unit is used to:

[0172] Obtaining the heat value of each of the target messages;

[0173] Sorting the titles corresponding to the multiple target messages included in the message brief according to the heat values ​​corresponding to the multiple target messages, and marking heat marks on the titles corresponding to the target messages whose heat values ​​are greater than a preset heat threshold, wherein the number of heat marks marked is related to the heat value of the target message;

[0174] The step of pushing the message brief to the user includes:

[0175] The sorted message briefs marked with the heat mark are pushed to the user.

[0176] A fourth embodiment of the present application provides an electronic device, Figure 8 It is a schematic diagram of the structure of the electronic device provided in this embodiment.

[0177] like Figure 8As shown, the electronic device provided in this embodiment includes: a memory 801 and a processor 802;

[0178] The memory 801 is used to store computer instructions for executing the message push method;

[0179] The processor 802 is configured to execute computer instructions stored in the memory 801 to perform the following operations:

[0180] According to the information to be concerned preset by the user, searching and obtaining multiple messages related to the information to be concerned from multiple preset information source services;

[0181] Inputting each of the messages into a large language model, and generating a summary text corresponding to each of the messages through the large language model;

[0182] Performing vectorization processing on the summary text corresponding to each of the messages to obtain a text vector corresponding to each of the messages;

[0183] For each of the text vectors corresponding to the message, the similarity between the text vector corresponding to the message and the database text vector stored in the vector database is retrieved and determined in the vector database in turn; wherein the vector database stores a plurality of database text vectors, and the database text vector is the text vector corresponding to the pushed message;

[0184] Based on the similarity retrieval results of the text vectors corresponding to the messages, a target message is determined from the messages, the text vectors corresponding to the target message are stored in the vector database, and the summary text corresponding to the target message is pushed to the user.

[0185] Optionally, the determining a target message from the messages based on the similarity search result of the text vector corresponding to the messages includes:

[0186] In response to the absence of a database text vector similar to the text vector corresponding to the message in the vector database, the message is determined to be a target message.

[0187] Optionally, also include:

[0188] In response to the presence of a database text vector in the vector database that is similar to the text vector corresponding to the message, but the time interval between the acquisition time of the message and the acquisition time of the pushed message corresponding to the database text vector is greater than or equal to a first preset time threshold, the message is determined to be a target message.

[0189] Optionally, before the step of determining the message as the target message in response to the absence of a database text vector similar to the text vector corresponding to the message in the vector database, the following steps are further performed:

[0190] Based on the similarity search results of the text vectors corresponding to the message, obtaining a first number of the database text vectors with the highest similarity;

[0191] In response to the similarity between each of the first number of database text vectors and the text vector corresponding to the message being less than a first preset similarity threshold, it is determined that there is no database text vector in the vector database that is similar to the text vector corresponding to the message.

[0192] Optionally, before the step of determining the message as the target message in response to the presence in the vector database of a database text vector similar to the text vector corresponding to the message, but the time interval between the acquisition time of the message and the acquisition time of the pushed message corresponding to the database text vector is greater than or equal to a first preset time threshold, further performing:

[0193] Based on the similarity search results of the text vectors corresponding to the message, obtaining a first number of the database text vectors with the highest similarity;

[0194] In response to the similarity between any one of the first number of database text vectors and the text vector corresponding to the message being greater than or equal to a first preset similarity threshold, it is determined that there is a database text vector in the vector database that is similar to the text vector corresponding to the message.

[0195] Optionally, before the step of pushing the summary text corresponding to the target message to the user, the following steps are further performed:

[0196] Based on the vector database, determining whether there is a pushed message related to the target message;

[0197] The step of pushing the summary text corresponding to the target message to the user includes:

[0198] In response to the existence of the pushed message related to the target message, the pushed message is associated with the summary text corresponding to the target message and presented on a push page of the target message, and the push page is pushed to the user.

[0199] Optionally, the determining, based on the vector database, whether there is a pushed message related to the target message includes:

[0200] Calculating the similarity between the text vector corresponding to the target message and each of the database text vectors;

[0201] Obtaining a second number of the database text vectors having the highest similarity;

[0202] In response to the similarity between any one of the second number of database text vectors and the text vector corresponding to the target message being greater than or equal to a second preset similarity threshold, and the time interval between the acquisition time of the target message and the acquisition time of the pushed message corresponding to the database text vector being greater than or equal to a second preset time threshold, determining that there is the pushed message related to the target message;

[0203] In response to the existence of the pushed message related to the target message, the pushed message is associated with the summary text corresponding to the target message and presented on a push page of the target message, and the push page is pushed to the user, comprising:

[0204] sorting the pushed messages related to the target message by relevance;

[0205] The pushed messages of the third number having the highest relevance are associated with the summary text corresponding to the target message and presented on a push page of the target message, and the push page is pushed to the user.

[0206] Optionally, inputting each of the messages into a large language model, and generating a summary text corresponding to each of the messages through the large language model, includes:

[0207] Generate a prompt word based on the message, the prompt word at least including the following instructions:

[0208] Generate a summary text of the message, delete the content in the message that is irrelevant to the information to be paid attention to, and generate a comment text with a preset style for the message; wherein the preset style is determined by the prompt word;

[0209] Input the prompt word into the large language model, delete the content in the message that is irrelevant to the information to be concerned through the large language model, and generate the summary text corresponding to the message with the irrelevant content deleted, and the comment text with a preset style;

[0210] The step of pushing the summary text corresponding to the target message to the user includes:

[0211] The summary text corresponding to the target message and the review text of a preset style are associated and presented on a push page of the target message, and the push page is pushed to the user.

[0212] Optionally, before the step of pushing the summary text corresponding to the target message to the user, the following steps are further performed:

[0213] Acquire multiple pictures contained in the target message;

[0214] Acquire the picture with the highest information content from the multiple pictures, and use the picture with the highest information content as the picture to be pushed corresponding to the target message;

[0215] The step of pushing the summary text corresponding to the target message to the user includes:

[0216] The summary text corresponding to the target message and the picture to be pushed are associated and presented on a push page of the target message, and the push page is pushed to the user.

[0217] Optionally, before the step of pushing the summary text corresponding to the target message to the user, the following steps are further performed:

[0218] Entering the title corresponding to the target message into a briefing template to form a message briefing including the titles corresponding to the multiple target messages;

[0219] Pushing the message brief to the user;

[0220] The step of pushing the summary text corresponding to the target message to the user includes:

[0221] In response to the user's viewing instruction for a first target message based on the message brief, the summary text corresponding to the first target message is pushed to the user, and the first target message is any one of the multiple target messages.

[0222] Optionally, after the step of forming a message briefing including titles corresponding to a plurality of target messages, the following steps are further performed:

[0223] Obtaining the heat value of each of the target messages;

[0224] Sorting the titles corresponding to the multiple target messages included in the message brief according to the heat values ​​corresponding to the multiple target messages, and marking heat marks on the titles corresponding to the target messages whose heat values ​​are greater than a preset heat threshold, wherein the number of heat marks marked is related to the heat value of the target message;

[0225] The step of pushing the message brief to the user includes:

[0226] The sorted message briefs marked with the heat mark are pushed to the user.

[0227] The fifth embodiment of the present application provides a computer-readable storage medium, which includes computer instructions. When the computer instructions are executed by a processor, they are used to implement the methods described in the embodiments of the present application.

[0228] It should be noted that the relational terms such as "first" and "second" in this article are only used to distinguish one entity or operation from another entity or operation, and do not require or imply any actual relationship or order between these entities or operations. In addition, the words "include", "have", "include" and "includes" and other similar forms are the same in meaning, and the end of any one or more items after any of the above words is open-ended, and any of the above nouns does not mean that the one or more items have been exhaustively listed or are limited to the one or more items listed.

[0229] As used herein, unless expressly stated otherwise, the term "or" includes all possible combinations, except those that are not feasible. For example, if it is expressed that a database may include A or B, then unless otherwise specifically stated or not feasible, it may include database A, or B, or A and B. As a second example, if it is expressed that a database may include A, B, or C, then unless otherwise specifically stated or not feasible, the database may include database A, or B, or C, or A and B, or A and C, or B and C, or A, B, and C.

[0230] It is worth noting that the above embodiments can be implemented by hardware or software (program code), or a combination of hardware and software. If implemented by software, it can be stored in the above-mentioned computer-readable medium. When the software is executed by a processor, it can execute the above-mentioned disclosed method. The computing unit and other functional units described in the present disclosure can be implemented by hardware or software, or a combination of hardware and software. Those of ordinary skill in the art will also understand that the above-mentioned multiple modules / units can be combined into one module / unit, and each of the above-mentioned modules / units can be further divided into multiple sub-modules / sub-units.

[0231] In the above detailed description, the embodiments have been described with reference to many specific details, which may vary from implementation to implementation. Certain adaptations and modifications may be made to the embodiments. For those skilled in the art, other embodiments may be readily apparent from the specific embodiments disclosed in the present application. This specification and examples are for exemplary purposes only, and the true scope and essence of the present application are described by the claims. The sequence of steps shown in the diagrams is also for the purpose of explanation only and is not intended to be limited to any particular steps or sequence. Therefore, those skilled in the art will appreciate that these steps may be performed in different orders when implementing the same method.

[0232] In the drawings and detailed description of the present application, exemplary embodiments are disclosed. However, many variations and modifications may be made to these embodiments. Accordingly, although specific terms are used, these terms are only general and descriptive and not for the purpose of limitation.

Claims

1. A message push method, characterized in that: The method comprises: According to the information to be concerned preset by the user, searching and obtaining multiple messages related to the information to be concerned from multiple preset information source services; Inputting each of the messages into a large language model, and generating a summary text corresponding to each of the messages through the large language model; Performing vectorization processing on the summary text corresponding to each of the messages to obtain a text vector corresponding to each of the messages; For each of the text vectors corresponding to the message, the similarity between the text vector corresponding to the message and the database text vector stored in the vector database is retrieved and determined in the vector database in turn; wherein the vector database stores a plurality of database text vectors, and the database text vector is the text vector corresponding to the pushed message; Based on the similarity retrieval results of the text vectors corresponding to the messages, a target message is determined from the messages, the text vectors corresponding to the target message are stored in the vector database, and the summary text corresponding to the target message is pushed to the user.

2. The method according to claim 1, characterized in that The step of determining a target message from the messages based on the similarity search results of the text vectors corresponding to the messages comprises: In response to the absence of a database text vector similar to the text vector corresponding to the message in the vector database, the message is determined to be a target message.

3. The method according to claim 2, characterized in that The method further comprises: In response to the presence of a database text vector in the vector database that is similar to the text vector corresponding to the message, but the time interval between the acquisition time of the message and the acquisition time of the pushed message corresponding to the database text vector is greater than or equal to a first preset time threshold, the message is determined to be a target message.

4. The method according to claim 2, characterized in that: Before the step of determining the message as a target message in response to the absence of a database text vector similar to the text vector corresponding to the message in the vector database, the method further includes: Based on the similarity search results of the text vectors corresponding to the message, obtaining a first number of the database text vectors with the highest similarity; In response to the similarity between each of the first number of database text vectors and the text vector corresponding to the message being less than a first preset similarity threshold, it is determined that there is no database text vector in the vector database that is similar to the text vector corresponding to the message.

5. The method according to claim 3, characterized in that: Before the step of determining the message as a target message in response to the presence of a database text vector similar to the text vector corresponding to the message in the vector database, but the time interval between the acquisition time of the message and the acquisition time of the pushed message corresponding to the database text vector is greater than or equal to the first preset time threshold, the method further includes: Based on the similarity search results of the text vectors corresponding to the message, obtaining a first number of the database text vectors with the highest similarity; In response to the similarity between any one of the first number of database text vectors and the text vector corresponding to the message being greater than or equal to a first preset similarity threshold, it is determined that there is a database text vector in the vector database that is similar to the text vector corresponding to the message.

6. The method according to claim 1, characterized in that Before the step of pushing the summary text corresponding to the target message to the user, the method further includes: Based on the vector database, determining whether there is a pushed message related to the target message; The step of pushing the summary text corresponding to the target message to the user includes: In response to the existence of the pushed message related to the target message, the pushed message is associated with the summary text corresponding to the target message and presented on a push page of the target message, and the push page is pushed to the user.

7. The method according to claim 6, characterized in that The determining, based on the vector database, whether there is a pushed message related to the target message includes: Calculating the similarity between the text vector corresponding to the target message and each of the database text vectors; Obtaining a second number of the database text vectors having the highest similarity; In response to the similarity between any one of the second number of database text vectors and the text vector corresponding to the target message being greater than or equal to a second preset similarity threshold, and the time interval between the acquisition time of the target message and the acquisition time of the pushed message corresponding to the database text vector being greater than or equal to a second preset time threshold, determining that there is the pushed message related to the target message; In response to the existence of the pushed message related to the target message, the pushed message is associated with the summary text corresponding to the target message and presented on a push page of the target message, and the push page is pushed to the user, comprising: sorting the pushed messages related to the target message by relevance; The pushed messages of the third number having the highest relevance are associated with the summary text corresponding to the target message and presented on a push page of the target message, and the push page is pushed to the user.

8. The method according to claim 1, characterized in that The step of inputting each of the messages into a large language model and generating a summary text corresponding to each of the messages through the large language model includes: Generate a prompt word based on the message, the prompt word at least including the following instructions: Generate a summary text of the message, delete the content in the message that is irrelevant to the information to be paid attention to, and generate a comment text with a preset style for the message; wherein the preset style is determined by the prompt word; Input the prompt word into the large language model, delete the content in the message that is irrelevant to the information to be concerned through the large language model, and generate the summary text corresponding to the message with the irrelevant content deleted, and the comment text with a preset style; The step of pushing the summary text corresponding to the target message to the user includes: The summary text corresponding to the target message and the review text of a preset style are associated and presented on a push page of the target message, and the push page is pushed to the user.

9. The method according to claim 1, characterized in that: Before the step of pushing the summary text corresponding to the target message to the user, the method further includes: Acquire multiple pictures contained in the target message; Acquire the picture with the highest information content from the multiple pictures, and use the picture with the highest information content as the picture to be pushed corresponding to the target message; The step of pushing the summary text corresponding to the target message to the user includes: The summary text corresponding to the target message and the picture to be pushed are associated and presented on a push page of the target message, and the push page is pushed to the user.

10. The method according to claim 1, characterized in that Before the step of pushing the summary text corresponding to the target message to the user, the method further includes: Entering the title corresponding to the target message into a briefing template to form a message briefing including the titles corresponding to the multiple target messages; Pushing the message brief to the user; The step of pushing the summary text corresponding to the target message to the user includes: In response to the user's viewing instruction for a first target message based on the message brief, the summary text corresponding to the first target message is pushed to the user, and the first target message is any one of the multiple target messages.

11. The method according to claim 10, characterized in that After the step of forming a message brief including titles corresponding to a plurality of target messages, the method further comprises: Obtaining the heat value of each of the target messages; Sorting the titles corresponding to the multiple target messages included in the message brief according to the heat values ​​corresponding to the multiple target messages, and marking heat marks on the titles corresponding to the target messages whose heat values ​​are greater than a preset heat threshold, wherein the number of heat marks marked is related to the heat value of the target message; The step of pushing the message brief to the user includes: The sorted message briefs marked with the heat mark are pushed to the user.

12. A message push system, characterized in that: The system includes: a module for customizing information to be followed and information sources, a message collection module, a text processing module, a text database, a vector processing module, a vector database, and a push module; The module for customizing information to be followed and information source is used to provide a window interface for users to customize information to be followed and information source services; The message collection module is used to search and obtain multiple messages related to the information to be concerned from multiple customized information source services according to the information to be concerned customized by the user; The text processing module is used to input each of the messages into a large language model, and generate a summary text corresponding to each of the messages through the large language model; The text database is used to store a message list corresponding to each of the messages, wherein the message list at least records a summary text of the message; The vector processing module is used to perform vectorization processing on the summary text corresponding to each of the messages to obtain a text vector corresponding to each of the messages; and sequentially search and determine the similarity between the text vector corresponding to each of the messages and the database text vector stored in the vector database; The vector database is used to store database text vectors, where the database text vectors are text vectors corresponding to the pushed messages; The push module is used to determine the target message from the messages based on the similarity retrieval results of the text vectors corresponding to the messages, store the text vectors corresponding to the target messages in the vector database, and push the summary text corresponding to the target message to the user.

13. A message push device, characterized in that: The device comprises: an acquisition unit, a first processing unit, a second processing unit, a judgment unit, and a push unit; The acquisition unit is used to search and acquire multiple messages related to the information to be concerned from multiple preset information source services according to the information to be concerned preset by the user; The first processing unit is used to input each of the messages into a large language model, and generate a summary text corresponding to each of the messages through the large language model; The second processing unit is used to perform vectorization processing on the summary text corresponding to each of the messages to obtain a text vector corresponding to each of the messages; The judgment unit is used to retrieve and judge the similarity between the text vector corresponding to each message and the database text vector stored in the vector database in turn for the text vector corresponding to each message; wherein the vector database stores a plurality of database text vectors, and the database text vector is the text vector corresponding to the pushed message; The push unit is used to determine a target message from the messages based on a similarity search result of the text vector corresponding to the message, store the text vector corresponding to the target message in the vector database, and push the summary text corresponding to the target message to the user.

14. An electronic device, characterized in that: include: Memory, processor; The memory is used to store one or more computer instructions; The processor is used to execute the one or more computer instructions to implement the method according to any one of claims 1-11.

15. A computer-readable storage medium having one or more computer instructions stored thereon, characterized in that: When the instruction is executed by a processor, the method according to any one of claims 1 to 11 is performed.

Images