Large model-based summary generation method, model training method, device and medium

Abstract

This disclosure provides a summarization method, model training method, apparatus, and medium based on a large model, relating to the field of artificial intelligence technology, particularly to the fields of data search, large language models, intelligent agents, and intelligent proxies. The implementation scheme is as follows: In response to receiving a search request, a data search is performed based on the search request to obtain at least one original search result; for each of the at least one original search result, a large summarization model is used to generate a target summary of the original search result that satisfies the search intent of the search request, based on the search request and the original search results; and a search result display page is displayed in response to the search request to show the search result entry and target summary content corresponding to each of the at least one original search result.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, particularly to the fields of data search, large language models, intelligent agents, and intelligent proxies. Specifically, it relates to a summarization method based on a large model, a training method for a large summarization model, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product. Background Technology

[0002] Artificial intelligence (AI) is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. AI software technologies mainly include computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.

[0003] In globalized data search and content consumption scenarios, users often need to obtain core information from content in a non-native language. For example, a user might use a Chinese search query to find an English scientific document or foreign news report and expect to directly obtain a Chinese summary to quickly understand the content. Related technologies typically employ a cascaded pipeline of "retrieval-translation-summarization," meaning the full text of the foreign language is first obtained, then translated into Chinese, and finally a summary is extracted from the Chinese text.

[0004] The methods described in this section are not necessarily methods that had been previously conceived or adopted. Unless otherwise specified, no method described in this section should be assumed to be prior art simply because it is included in this section. Similarly, unless otherwise specified, the issues mentioned in this section should not be considered to be accepted in any prior art. Summary of the Invention

[0005] This disclosure provides a method for generating summaries based on large models, a method for training large models for generating summaries, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product.

[0006] According to one aspect of this disclosure, a summary generation method based on a large model is provided, comprising: in response to receiving a search request, performing a data search based on the search request to obtain at least one original search result; for each of the at least one original search result, using a summary generation large model, generating a target summary content of the original search result that satisfies the search intent of the search request based on the search request and the original search result; and displaying a search result display page for responding to the search request, the search result display page displaying the search result entry and the target summary content corresponding to each of the at least one original search result.

[0007] According to another aspect of this disclosure, a training method for a large-scale summary generation model is provided, comprising: acquiring first sample data, the first sample data including a first sample search request, a first sample search result corresponding to the first sample search request, and first sample summary content of the first sample search result that can satisfy the search intent of the first sample search request; based on the first sample search request and the first sample search result, using a preset first prompt word template, acquiring a first prompt word corresponding to the first sample data, wherein the first prompt word template includes prompt information for guiding the large-scale model to generate summary content of the first sample search result that can satisfy the search intent of the first sample search request based on the first sample search request and the first sample search result; based on the first prompt word, using an initial large-scale model to perform a summary content generation operation for the first sample search result to obtain a first predicted summary content corresponding to the first sample search result; and training the initial large-scale model based on the first sample summary content and the first predicted summary content to obtain a large-scale summary generation model.

[0008] According to another aspect of this disclosure, a summary generation apparatus based on a large model is provided, comprising: a search unit configured to perform a data search based on the search request in response to receiving a search request to obtain at least one original search result; a generation unit configured to generate a target summary content of the original search results that satisfies the search intent of the search request for each of the at least one original search results using a summary generation large model; and a display unit configured to display a search result display page for responding to the search request, wherein the search result display page displays the search result entry and the target summary content corresponding to each of the at least one original search results.

[0009] According to another aspect of this disclosure, a training apparatus for a large-scale summary generation model is provided, comprising: a first acquisition unit configured to acquire first sample data, the first sample data including a first sample search request, a first sample search result corresponding to the first sample search request, and first sample summary content of the first sample search result that can satisfy the search intent of the first sample search request; a second acquisition unit configured to acquire a first prompt word corresponding to the first sample data based on the first sample search request and the first sample search result, using a preset first prompt word template, wherein the first prompt word template includes prompt information for guiding the large-scale model to generate summary content of the first sample search result that can satisfy the search intent of the first sample search request based on the first sample search request and the first sample search result; a first execution unit configured to perform a summary content generation operation for the first sample search result using an initial large-scale model based on the first prompt word, to obtain a first predicted summary content corresponding to the first sample search result; and a first training unit configured to train the initial large-scale model based on the first sample summary content and the first predicted summary content, to obtain a large-scale summary generation model.

[0010] According to another aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the summary generation method based on a large model or the training method for generating a large model of summary according to this disclosure.

[0011] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause a computer to execute the large model-based summarization method or the training method for summarization generation large model of this disclosure.

[0012] According to another aspect of this disclosure, a computer program product is provided, including a computer program, wherein the computer program, when executed by a processor, implements the large model-based summarization method or the training method for summarization generation of a large model of this disclosure.

[0013] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0014] The accompanying drawings exemplify embodiments and form part of the specification, serving together with the textual description to explain exemplary implementations of the embodiments. The illustrated embodiments are for illustrative purposes only and do not limit the scope of the claims. Throughout the drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0015] Figure 1 A schematic diagram of an exemplary system in which the various methods described herein may be implemented according to embodiments of the present disclosure is shown; Figure 2 A flowchart of a large-model-based summary generation method according to an embodiment of the present disclosure is shown; Figure 3 A schematic diagram illustrating an application scenario according to an exemplary embodiment of the present disclosure is shown; Figure 4 A flowchart illustrating a training method for generating a large model from a summary according to embodiments of the present disclosure is shown; Figure 5 A flowchart illustrating the acquisition of an initial large model according to an embodiment of the present disclosure is shown; Figure 6 A flowchart illustrating a training method for generating a large model from a summary according to an exemplary embodiment of the present disclosure is shown. Figure 7 A structural block diagram of a large-model-based summary generation apparatus according to an embodiment of the present disclosure is shown; Figure 8 A structural block diagram of a training apparatus for generating large models according to embodiments of the present disclosure is shown. Figure 9 A structural block diagram of an exemplary electronic device that can be used to implement embodiments of the present disclosure is shown. Detailed Implementation

[0016] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0017] In this disclosure, unless otherwise stated, the use of terms such as "first," "second," etc., to describe various elements is not intended to limit the positional, temporal, or importance relationships of these elements; such terms are merely used to distinguish one element from another. In some examples, the first element and the second element may refer to the same instance of that element, while in other cases, based on the context, they may refer to different instances.

[0018] The terminology used in the description of the various examples described in this disclosure is for the purpose of describing particular examples only and is not intended to be limiting. Unless the context explicitly indicates otherwise, an element may be one or more unless the number of elements is specifically limited. Furthermore, the term "and / or" as used in this disclosure covers any one of the listed items and all possible combinations thereof.

[0019] In related technologies, data search scenarios typically employ a cascaded pipeline of "retrieval-translation-summarization," which involves first obtaining the full text of the foreign language, then translating it into Chinese, and finally extracting a summary from the Chinese text. In this process, the summary generation is completely independent of the user's search request, resulting in summaries that are merely general and fail to accurately address the specific information needs expressed in the user's search query.

[0020] This disclosure provides a summary generation method based on a large model. After obtaining the search results, a summary generation large model is introduced for each original search result. The user's search request and the original search results are used as input to the large model, thereby directly generating target summary content that accurately meets the user's search intent and displaying it on the page. This makes the summary content of the search results displayed on the search results page more targeted to the search request, improving the efficiency of users obtaining key information directly on the search results page and the overall search experience.

[0021] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0022] Figure 1 A schematic diagram of an exemplary system 100 in which the various methods and apparatus described herein can be implemented according to embodiments of this disclosure is shown. Reference Figure 1 The system 100 includes one or more client devices 101, 102, 103, 104, 105 and 106, a server 120, and one or more communication networks 110 coupling the one or more client devices to the server 120. The client devices 101, 102, 103, 104, 105 and 106 can be configured to execute one or more applications.

[0023] In embodiments of this disclosure, server 120 may run one or more services or software applications that enable the execution of the large model-based summarization method or the training method for summarizing large models of this disclosure.

[0024] In some embodiments, server 120 may also provide other services or software applications, which may include non-virtual and virtual environments. In some embodiments, these services may be provided as web-based services or cloud services, such as to users of client devices 101, 102, 103, 104, 105 and / or 106 under a Software as a Service (SaaS) model.

[0025] exist Figure 1 In the configuration shown, server 120 may include one or more components that implement the functions performed by server 120. These components may include software components, hardware components, or combinations thereof that can be executed by one or more processors. Users operating client devices 101, 102, 103, 104, 105, and / or 106 can sequentially interact with server 120 using one or more client applications to utilize the services provided by these components. It should be understood that various different system configurations are possible and may differ from system 100. Therefore, Figure 1 This is an example of a system used to implement the various methods described herein, and is not intended to be limiting.

[0026] Users can use client devices 101, 102, 103, 104, 105, and / or 106 to input and send search requests. The client devices can provide an interface that allows users to interact with them. The client devices can also output information to the user through this interface. Although... Figure 1 Only six client devices are described, but those skilled in the art will understand that this disclosure can support any number of client devices.

[0027] Client devices 101, 102, 103, 104, 105, and / or 106 may include various types of computer devices, such as portable handheld devices, general-purpose computers (such as personal computers and laptops), workstation computers, wearable devices, smart screen devices, self-service terminal devices, service robots, gaming systems, thin clients, various messaging devices, sensors, or other sensing devices. These computer devices can run various types and versions of software applications and operating systems, such as Microsoft Windows, Apple iOS, UNIX-like operating systems, Linux or Linux-like operating systems (such as Google Chrome OS); or include various mobile operating systems, such as Microsoft Windows Mobile OS, iOS, Windows Phone, and Android. Portable handheld devices may include cellular phones, smartphones, tablets, personal digital assistants (PDAs), etc. Wearable devices may include head-mounted displays (such as smart glasses) and other devices. Gaming systems may include various handheld gaming devices, internet-enabled gaming devices, etc. Client devices are capable of executing various applications, such as various internet-related applications, communication applications (such as email applications), short message service (SMS) applications, and can use various communication protocols.

[0028] Network 110 can be any type of network well known to those skilled in the art, and can support data communication using any of a variety of available protocols (including but not limited to TCP / IP, SNA, IPX, etc.). By way of example only, one or more networks 110 can be a local area network (LAN), an Ethernet-based network, a token ring network, a wide area network (WAN), the Internet, a virtual network, a virtual private network (VPN), an intranet, an extranet, a blockchain network, a public switched telephone network (PSTN), an infrared network, a wireless network (e.g., Bluetooth, WIFI), and / or any combination of these and / or other networks.

[0029] Server 120 may include one or more general-purpose computers, special-purpose server computers (e.g., PC (personal computer) servers, UNIX servers, mid-range servers), blade servers, mainframe computers, server clusters, or any other suitable arrangement and / or combination. Server 120 may include one or more virtual machines running a virtual operating system, or other computing architectures involving virtualization (e.g., one or more flexible pools of logical storage devices that can be virtualized to maintain virtual storage devices for servers). In various embodiments, server 120 may run one or more services or software applications that provide the functionality described below.

[0030] The computing unit in server 120 can run one or more operating systems, including any of the aforementioned operating systems and any commercially available server operating system. Server 120 can also run any of a variety of additional server applications and / or middleware applications, including HTTP servers, FTP servers, CGI servers, JAVA servers, database servers, etc.

[0031] In some implementations, server 120 may include one or more applications to analyze and merge data feeds and / or event updates received from users of client devices 101, 102, 103, 104, 105 and / or 106. Server 120 may also include one or more applications to display data feeds and / or real-time events via one or more display devices of client devices 101, 102, 103, 104, 105 and / or 106.

[0032] In some implementations, server 120 can be a server for a distributed system or a server integrated with blockchain. Server 120 can also be a cloud server, or an intelligent cloud computing server or intelligent cloud host with artificial intelligence technology. A cloud server is a host product in the cloud computing service system, designed to address the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0033] System 100 may also include one or more databases 130. In some embodiments, these databases may be used to store data and other information. For example, one or more of the databases 130 may be used to store information such as audio files and video files. Databases 130 may reside in various locations. For example, a database used by server 120 may be local to server 120, or it may be located away from server 120 and may communicate with server 120 via a network-based or dedicated connection. Databases 130 may be of different types. In some embodiments, the database used by server 120 may be, for example, a relational database. One or more of these databases may store, update, and retrieve data from and from the databases in response to commands.

[0034] In some embodiments, one or more of the databases 130 may also be used by an application to store application data. The databases used by the application may be of different types, such as key-value stores, object stores, or regular stores supported by a file system.

[0035] Figure 1The system 100 can be configured and operated in various ways to enable the application of the various methods and apparatus described in this disclosure.

[0036] According to embodiments of this disclosure, such as Figure 2 As shown, a summary generation method based on a large model is provided, including: step S201, in response to receiving a search request, performing a data search based on the search request to obtain at least one original search result; step S202, for each of the at least one original search results, using a summary generation large model, generating a target summary content of the original search results that can satisfy the search intent of the search request based on the search request and the original search results; and step S203, displaying a search results display page for responding to the search request, wherein the search results display page displays the search result entries and target summary content corresponding to each of the at least one original search results.

[0037] Therefore, by introducing a summary generation model for each original search result after obtaining the search results, and using the user's search request and the original search results as input to the big model, it is possible to directly generate target summary content that accurately meets the user's search intent and display it on the page. This makes the summary content of the search results displayed on the search results page more targeted to the search request, improving the efficiency of users obtaining key information directly on the search results page and the overall search experience.

[0038] In some embodiments, a search request can be a query instruction entered by a user into the search system to obtain specific information. In practical applications, search requests in the form of text, voice, etc., entered by the user can be obtained through the search interface provided to the user on the terminal device (such as the search box or voice input interface on the front-end page).

[0039] After receiving the search request, the system will perform a data search based on the request to obtain at least one original search result. Specifically, the search request can be submitted to a backend search engine, which will use its internal inverted index or similarity matching algorithm to retrieve web pages or documents from a massive database and recall web pages or documents highly relevant to the search request. This recalled data will then be returned as the original search result.

[0040] In some embodiments, for each of the at least one original search results, a large-scale summary generation model can be used to generate a target summary of the original search result that satisfies the search intent of the search request. Specifically, the search request, the corresponding original search result, and a preset prompt word template (e.g., a first prompt word template) can be combined to form prompt words, which are then input into a trained large-scale summary generation model. Guided by these prompt words, the large-scale summary generation model comprehensively understands the search request and the search result content, automatically performs a summarization operation, and generates the corresponding target summary content.

[0041] In some embodiments, the search intent can be the true information need or retrieval purpose hidden within a user's search query. For example, a user might be seeking an answer to a specific question, understanding the operational steps of a technology, or finding background information on an event. Correspondingly, the target summary content of the original search results that satisfies the search intent can be text information extracted, refined, or summarized by a summary generation model from the corresponding original search results for that search intent. This text information directly answers the user's question or directly responds to the user's search request. Unlike traditional search engines that mechanically extract content from fixed locations on web pages, this target summary content is highly customized and has a clear answering nature, allowing users to directly obtain the core information they truly care about simply by reading the summary on the search results display page.

[0042] In some embodiments, the above-described abstract generation method can be applied, but is not limited to, cross-language search engines, multilingual academic literature retrieval platforms, and global news aggregation applications. Figure 3 A schematic diagram illustrating an application scenario according to an exemplary embodiment of this disclosure is shown. For example... Figure 3 As shown, the search results display page used to respond to the search request can be used to centrally present to the user the various contents retrieved and processed by the system based on their search request, such as the search results page of a search engine.

[0043] For example, the page may include a search input area for the user to view or modify the search request, and within the main display area of ​​the page, information blocks corresponding to each of the at least one original search result are displayed sequentially downwards. Through this search results display page, the user can intuitively browse the summary of various results returned by the system in response to their input search request.

[0044] In some embodiments, on the search results display page, the search result entry corresponding to each original search result is used to identify the source or topic of that original search result. In the actual page display format, this search result entry can typically be displayed as the title text of the complete webpage or document containing the original search result, for example... Figure 3 The search result entries shown are "Search Results 1 Title", "Search Results 2 Title", and "Search Results 3 Title". These search result entries can be configured as interactive hyperlinks, allowing users to navigate to the original page corresponding to the original search result by clicking the title text when they need complete information.

[0045] In some embodiments, on the search results display page, for each of the at least one original search results, the system displays the target summary content generated by the summary generation model in an associated position, such as below or beside its corresponding search result entry. Figure 3 The target summary content 301 shown is presented as a text paragraph directly following the search result entry. This display structure visually constitutes a complete search result display unit with the search result entry, allowing users to directly obtain textual information that satisfies their search intent by reading the target summary content without clicking on the search result entry to access the original page.

[0046] In some embodiments, generating a target summary of the original search results that satisfies the search intent of the search request based on the search request and the original search results may include: analyzing the search intent corresponding to the search request based on the search request and the original search results; and generating a target summary of the original search results that satisfies the search intent based on the search intent and the original search results.

[0047] Therefore, by analyzing the search intent corresponding to the search request based on the search request and the original search results, and generating target summary content based on the search intent and the original search results, it is possible to ensure that the final generated target summary content accurately matches the actual search intent, thereby further improving the accuracy of the target summary content.

[0048] In some embodiments, the aforementioned preset prompt word template (e.g., the first prompt word template) may be a chain prompt word template, used to guide the large model to execute each stage task in a specific order.

[0049] In some examples, the chained prompt template can include prompts for two progressively guiding stages: the first stage can be an intent analysis stage, which guides the large model to complete the operation of analyzing the search intent corresponding to the search request based on the search request and the current original search results through specific prompts (such as "Please combine the search request and the current original search results to analyze the deep intent of the search request"); the second stage can be a summary generation stage, which guides the large model to complete the operation of generating target summary content of the original search results that can satisfy the search intent based on the search intent and the original search results through further prompts (such as "Please extract or summarize the core information that satisfies the intent from the original search results based on the above intent analysis results and organize it into summary content output").

[0050] This breakdown of intermediate steps based on the thought chain can effectively reduce the probability of large models generating "illusions" or deviating from user needs, thereby significantly improving the matching degree between the generated target summary content and search intent.

[0051] In related technologies, a cascaded pipeline approach is typically used in cross-language search and summary generation scenarios. The specific implementation process is as follows: First, in the text translation stage, the system uses an independent machine translation model to translate the retrieved source language web page documents (original search results in the original language) into Chinese. Second, in the summary generation stage, the system inputs the translated Chinese text into an independent Chinese summary generation model to obtain a Chinese summary; this model typically employs extraction or generative methods. Finally, in the result integration stage, the system integrates the generated Chinese summary with the search results and presents it to the user.

[0052] However, the aforementioned technologies have significant technical problems: First, this cascaded scheme is prone to error accumulation. Since text translation and summary generation are two independent models, semantic errors or information loss generated by the translation model during full-text translation will be directly passed downstream, severely affecting the accuracy of subsequent summary generation and resulting in poor final summary quality. Second, the scheme is inefficient. The serial calling of the two independent models (translation model and summary generation model) significantly increases the overall computation time and resource consumption of the system, making it difficult to meet the stringent business requirements of modern search engines for high concurrency, high performance, and real-time response.

[0053] In some embodiments, generating target summary content that satisfies the search intent based on the search intent and the original search results may include: generating target summary content in the target language based on the search intent and the original search results in the original language.

[0054] Therefore, by generating target summary content in the target language based on search intent and original search results in the original language, cross-language, end-to-end summary generation is achieved, avoiding the problems of low generation efficiency and low accuracy of cascaded pipeline summary generation schemes, and improving the generation efficiency and accuracy of target summary content.

[0055] Understandably, the source language and target language mentioned above can be the same or different, and no restrictions are imposed here.

[0056] In some embodiments, generating target summary content of the original search results that can satisfy the search intent of the search request based on the search request and the original search results may further include: determining core answer content in the target summary content that can answer the search intent based on the search intent, wherein the number of characters in the core answer content is less than the number of characters in the target summary content; and wherein the display style of the core answer content corresponding to each original search result in at least one original search result displayed on the search results display page may be different from the display style of the corresponding target summary content.

[0057] Therefore, by identifying the core answer content that answers the search intent within the target summary content based on the search intent, and by making the display style of the core answer content on the search results page different from the display style of the corresponding target summary content, the core answer content can be highlighted on the search results page. This reduces the time users spend locating and identifying key information that answers their search intent on the page, further improving the user's search experience.

[0058] In some embodiments, the core answer content may be a key text fragment extracted from the target summary content that directly answers the search intent. The core answer content is part of the target summary content and is contained within its text sequence. While the target summary content provides coherent and contextualized information about the original search results, the core answer content is a concise text that further focuses on the specific answer; therefore, the core answer content has fewer characters than the target summary content.

[0059] On the search results display page, the system ensures that the display style of the core answer content corresponding to each original search result in at least one original search result differs from the display style of the corresponding target summary content. See also... Figure 3In the information block containing the target summary content 301, the system can visually distinguish the text corresponding to the core answer content 302 from other regular text in the target summary content by bolding or increasing its font size. Besides the bolding method shown in the image, the system can also differentiate it by highlighting the background of the core answer content or using a different font color than the rest of the target summary content. Through these differentiated display style configurations, the core answer content can create a visual contrast with other text content on the search results page, thus being intuitively distinguished.

[0060] In some embodiments, the aforementioned preset prompt word template (e.g., the first prompt word template) may further include prompt information for the third stage (i.e., the core answer content extraction stage). Through specific prompt information (such as "Please extract the shortest core answer that can be used to directly answer the user's search request from the target summary content"), the large model is guided to complete the operation of determining the core answer content that can answer the search intent in the target summary content based on the search intent.

[0061] In some embodiments, generating target summary content of the original search results that can satisfy the search intent of the search request based on the search request and the original search results may further include: detecting the target summary content to determine whether the target summary content satisfies the search intent; and in response to determining that the target summary content does not satisfy the search intent, re-executing the generation of the target summary content until it is determined that the regenerated target summary content satisfies the search intent or the number of re-executing exceeds a preset number.

[0062] Therefore, by detecting the target summary content to determine whether it meets the search intent, and re-executing the generation of the target summary content when it does not meet the intent, until the search intent is met or the number of repeated executions exceeds the preset number, the accuracy of the generated target summary content can be further guaranteed through the feedback and retry mechanism. At the same time, the preset number of times condition is used to avoid system response timeouts caused by infinite loop generation.

[0063] In some exemplary embodiments, the aforementioned preset prompt word template (e.g., the first prompt word template) may be a chain prompt word template designed to include multiple progressive guiding steps, so as to guide the summary generation model to sequentially perform complex logical tasks including intent analysis, target summary content generation, core answer content extraction, and self-testing.

[0064] In some exemplary embodiments, the chained suggestion template may include suggestions for four stages. The first stage may be an intent analysis stage, where specific suggestions guide the large model to combine the input search request and the original search results to analyze the deep search intent corresponding to the search request.

[0065] The second step can be the target summary content generation step. Through further prompts, the large model is guided to extract or summarize the summary content that meets the search intent from the original search results in the original language based on the search intent obtained in the previous step, and directly output the target summary content in the target language (e.g., Chinese). This realizes the operation of generating target summary content in the target language based on the search intent and the original search results in the original language.

[0066] The third step is the core answer content extraction stage. Guided by relevant prompts, the large model extracts content from the generated target summary that directly and concisely answers the search intent. This core answer content is then used as the core answer content, where the character count is less than the target summary content. In this way, the large model can identify the core answer content within the target summary content that responds to the search intent, allowing for different display styles on the search results page.

[0067] The fourth stage can be a reflection and improvement stage. Specific reflection prompts (such as "Please reflect on and improve the target summary content to ensure that the target summary content meets the needs of the search request") guide the large model to detect the generated target summary content in order to determine whether the target summary content fully and accurately meets the search intent.

[0068] After receiving the chained prompt templates, search request, and original search results (comprising the four stages mentioned above), the summary generation model will sequentially perform calculations according to the above steps during a single inference process. If, in the final reflection and improvement stage, the model determines that the currently generated target summary content does not meet the search intent, it will trigger an internal retry mechanism based on the prompt template instructions to re-execute the generation of the target summary content until it determines that the regenerated target summary content meets the search intent, or that the number of repetitions exceeds a preset number.

[0069] By integrating a chain of prompts with multi-step guidance, the large model can break down the complex end-to-end generation task into controllable intermediate steps and perform self-verification and improvement before output, thereby ensuring the extremely high accuracy and reliability of the final generated target summary content.

[0070] In some embodiments, such as Figure 4As shown, a training method for a large-scale summary generation model is also provided, including: step S401, obtaining first sample data, the first sample data including a first sample search request, a first sample search result corresponding to the first sample search request, and first sample summary content of the first sample search result that can satisfy the search intent of the first sample search request; step S402, based on the first sample search request and the first sample search result, using a preset first prompt word template, obtaining a first prompt word corresponding to the first sample data, wherein the first prompt word template includes prompt information used to guide the large-scale model to generate summary content of the first sample search result that can satisfy the search intent of the first sample search request based on the first sample search request and the first sample search result; step S403, based on the first prompt word, using an initial large-scale model to perform a summary content generation operation for the first sample search result, to obtain a first predicted summary content corresponding to the first sample search result; and step S404, based on the first sample summary content and the first predicted summary content, training the initial large-scale model to obtain a large-scale summary generation model.

[0071] Therefore, by acquiring first sample data containing a first sample search request, a first sample search result, and a first sample summary, and using a first prompt word template containing relevant prompt information to obtain a first prompt word, and training an initial large model based on the first sample summary content and the first predicted summary content generated by the initial large model according to the first prompt word, it is possible to use automatically constructed sample data to specifically train the initial large model, thus obtaining a summary generation large model capable of generating search result summaries tailored to user search requests. Using this summary generation large model, target summary content that accurately meets the user's search intent can be directly generated and displayed on the page, making the summary content of the search results displayed on the search results page more targeted to the search request, improving the efficiency of users directly obtaining key information on the search results page and enhancing the overall search experience.

[0072] In some embodiments, the large-scale summarization model used in the above-described large-scale model-based summarization method can be trained using the training method for the large-scale summarization model described above.

[0073] In some embodiments, the summary content generation operation may include: analyzing the search intent corresponding to the first sample search request based on the first sample search request and the first sample search result; and generating a first predicted summary content of the first sample search result that can satisfy the search intent based on the search intent and the first sample search result.

[0074] Therefore, by first analyzing the search intent corresponding to the first sample search request in the summary content generation operation, and then generating the first predicted summary content based on the search intent and the first sample search result, the model can be guided to learn the logical relationship between intent analysis and summary generation during the training process, thereby improving the final summary generation model's understanding of search intent and content response capabilities.

[0075] In some embodiments, generating a first predicted summary content that satisfies the first sample search result based on the search intent may include: generating a first predicted summary content in the target language based on the search intent and the original search result in the original language.

[0076] Therefore, by generating the first predicted summary content in the target language based on the search intent and the original search results in the original language during the training process, the large model can achieve cross-language, end-to-end summary generation, avoiding the problems of low generation efficiency and low accuracy of cascade pipeline summary generation schemes, and improving the generation efficiency and accuracy of summary content.

[0077] In some embodiments, the first sample data may further include sample core answer content that can respond to the search intent in the first sample summary content, wherein the number of characters in the sample core answer content is less than the number of characters in the first sample summary content. The first prompt word template may further include prompt information for guiding the large model to determine the core answer content that can respond to the search intent in the first predicted summary content based on the search intent. The summary content generation operation may further include: determining the predicted core answer content for responding to the search intent in the first predicted summary content based on the search intent; and wherein training an initial large model based on the first sample summary content and the first predicted summary content to obtain a summary generation large model may include: training an initial large model based on the first sample summary content, the sample core answer content, the first predicted summary content, and the predicted core answer content to obtain a summary generation large model.

[0078] Therefore, by including prompts in the first prompt word template to guide the large model in determining the core answer content, and by jointly training the initial large model based on the first sample summary content, the sample core answer content, the first predicted summary content, and the predicted core answer content, the summary generation large model is able to extract and locate the core answer content while learning to generate summary content.

[0079] In some embodiments, the first prompt word template may further include prompt information for guiding the large model to reflect on the first predicted summary content, and the summary content generation operation may further include: detecting the first predicted summary content to determine whether the first predicted summary content satisfies the search intent; and in response to determining that the first predicted summary content does not satisfy the search intent, re-executing the generation of the first predicted summary content until it is determined that the regenerated first predicted summary content satisfies the search intent or the number of re-executions exceeds a preset number.

[0080] Therefore, by including prompts in the first prompt word template to guide the large model to reflect on the first predicted summary content, and by detecting and conditionally re-generating the first predicted summary content, the model can autonomously optimize the generation results during training. This enables the trained summary generation model to have the ability to reflect and retry, further improving the accuracy of summary generation.

[0081] In some embodiments, obtaining the first sample data may include: based on the first prompt word, performing a summary content generation operation on the first sample search result using a first general large model to obtain the first sample summary content corresponding to the first sample search result, wherein the parameter size of the first general large model is larger than the parameter size of the initial large model.

[0082] Therefore, by using a first general large model with a parameter scale larger than the initial large model to perform the summary content generation operation to obtain the first sample summary content, it is possible to automatically construct more accurate first sample data by leveraging the complex logic processing capabilities of the high parameter scale model, thereby improving the efficiency of obtaining the first sample data.

[0083] In some embodiments, the first prompt word template used above may be the same as the prompt word template used in the above-described large model-based summary generation method, and will not be elaborated here.

[0084] In some embodiments, a loss can be calculated based on the difference between the first sample summary content and the first predicted summary content, and one or more parameters in the initial large model can be adjusted based on the loss to train the initial large model.

[0085] In some embodiments, a loss can be calculated based on the difference between the first sample summary content and the first predicted summary content, the difference between the sample core answer content and the predicted core answer content, and one or more parameters in the initial large model can be adjusted based on the loss to train the initial large model.

[0086] The training process of the aforementioned initial large model is essentially a knowledge distillation process based on supervised fine-tuning. To optimize the user experience in practical applications, particularly to reduce the inference time during real-time interaction, the embodiments of this disclosure do not directly use the large-scale first general-purpose model as the final online service model. Instead, they adopt a "teacher-student" training model. Specifically, the aforementioned first general-purpose model acts as the "teacher model," leveraging its larger parameter scale and stronger logical reasoning capabilities to produce highly accurate first sample data.

[0087] Subsequently, a larger model with a smaller parameter size can be selected as the "student model" (initial large model). Here, the parameter size of the first general large model is larger than that of the second general large model. Supervised fine-tuning of the initial large model is performed using the aforementioned first sample data, forcing the small model to fit the distribution of high-quality data generated by the large model. This allows the small model to mimic the target summary content generation of the large model. In this way, the rich generalization knowledge contained in the large model is effectively transferred to the small model, thereby obtaining the summary generation large model.

[0088] In some exemplary embodiments, the aforementioned first general large model can be a general large language model with a parameter scale of 8 billion (8B) (e.g., Qwen3-8B), while the aforementioned initial large model can be a lightweight large language model with a parameter scale of only 1.7 billion (1.7B) (e.g., Qwen3-1.7B). The summary generation large model after SFT data distillation (i.e., the fine-tuned Qwen3-1.7B) not only inherits the teacher model's good format compliance ability and high functional recommendation accuracy, but also significantly reduces its inference cost and improves its response speed due to the substantial reduction in its parameter count, thus better meeting the low latency requirements of mobile deployment or high-concurrency scenarios.

[0089] In some embodiments, such as Figure 5 As shown, obtaining the initial large model may include: step S501, obtaining second sample data, which includes second sample search results in the original language and second sample summary content in the target language corresponding to the second sample search results; step S502, based on the second sample search results, using a preset second prompt word template, obtaining second prompt words corresponding to the second sample data, wherein the second prompt word template includes prompt information used to guide the large model to generate summary content in the target language based on the second sample search results; step S503, based on the second prompt words, using a second general large model, generating second predicted summary content in the target language corresponding to the second sample search results; and step S504, based on the second sample summary content and the second predicted summary content, training the second general large model to obtain the initial large model.

[0090] Therefore, by acquiring second sample data based solely on the second sample search results in the original language and the corresponding second sample summary content in the target language, and using this data to train a second general large model to obtain an initial large model, a progressive pre-training stage is constructed. This allows the model to prioritize learning cross-language mapping and basic content summary rules, providing a more stable initial large model for subsequent training in conjunction with search requests.

[0091] In some embodiments, obtaining the second sample data may include: generating a second sample summary content corresponding to the second sample search result based on the second prompt word and using a third general large model, wherein the parameter size of the third general large model is larger than the parameter size of the second general large model.

[0092] Therefore, by using a third general large model with a parameter size larger than the second general large model to generate the second sample summary content, the automated construction of high-quality sample data in the basic training stage is realized, ensuring the reliability of the training data required to obtain the initial large model.

[0093] In some embodiments, the third general large model described above may be the same as the first general large model described above, or other high-performance large models with the same parameter scale may be applied, which are different from the first general large model described above. No restrictions are imposed here.

[0094] In some embodiments, training a second general large model based on the second sample summary content and the second predicted summary content to obtain an initial large model may include calculating a loss based on the difference between the second sample summary content and the second predicted summary content, and adjusting one or more parameters of the second general large model based on the loss to obtain the initial large model.

[0095] In some embodiments, the number of first sample data can be multiple, and the training method of the above-mentioned large-scale summary generation model can further include: filtering first sample data that meets preset conditions from multiple first sample data to serve as third sample data, wherein the preset conditions include at least one of the following: the number of characters in the first sample summary content is within a preset number range; and the first sample summary content does not contain redundant content, the redundant content including at least one of the content used to describe the first sample search request and the title of the first sample search result; and training the large-scale summary generation model based on the third sample data to update the large-scale summary generation model.

[0096] Therefore, by utilizing preset conditions such as the number of characters being within a preset range and the absence of redundant content, third sample data is selected from multiple first sample data. Based on the third sample data, the summary generation model is further trained, thereby further optimizing the structural standardization and information density of the generated content and improving the accuracy of summary generation.

[0097] In some embodiments, the first sample data may further include sample core answer content that can answer the search intent in the first sample summary content, wherein the number of characters in the sample core answer content is less than the number of characters in the first sample summary content, and the preset conditions may further include at least one of the following: the sample core answer content is located within the first preset number of characters of the first sample summary content; and the sample core answer content does not have any duplicate content with other content in the first sample summary content besides the sample core answer content.

[0098] Therefore, by using the premise that the core answer content of the sample is located within the first preset number of characters of the first sample summary content and that there is no duplicate content as the preset conditions for further filtering of the first sample data, samples with answers located at the end or with redundant information can be eliminated. This guides the model to learn to generate summary content with answers at the beginning and information concise, improving the fit of the model output content with online display standards and further optimizing the user's search experience.

[0099] In some embodiments, the first sample data can be filtered by rules. For example, the first sample data can be filtered by rules to select the first sample data whose number of characters in the first sample summary content is within a preset range, whose sample core answer content is within the first preset number of characters in the first sample summary content, or whose sample core answer content does not have any duplicate content with other content in the first sample summary content except for the sample core answer content.

[0100] In some embodiments, the first sample data can be filtered using a large model. For example, the large model can be used to determine whether the first sample summary content contains at least one of the following: content describing the first sample search request (the first sample search request itself or content with similar semantics, such as "the user query is XXX") and the title of the first sample search result.

[0101] In some embodiments, the process of training a large model based on third sample data can be similar to the process of training an initial large model based on first sample data, and will not be elaborated here.

[0102] Figure 6 A flowchart illustrating a training method for generating a large model from a summary according to an exemplary embodiment of the present disclosure is shown.

[0103] In some exemplary embodiments, such as Figure 6 As shown, a method for training a large model to generate summaries is provided. This method can be generally divided into an automated sample data generation stage and a three-stage progressive model training stage.

[0104] First, the system can collect massive amounts of raw search requests and their corresponding raw search content as a data source for subsequent automated acquisition of sample data required for each training stage. To guide models with a smaller parameter set to master the joint generation capabilities of cross-language processing and perceiving search intent, embodiments of this disclosure employ a three-stage progressive training strategy to gradually train the model.

[0105] In the first stage of the three-stage progressive model training, the main focus is on improving the basic summarization capabilities of the large model. In this stage, the system first acquires multiple second sample data, each of which includes a second sample search result in the original language and a second sample summary in the target language corresponding to that search result.

[0106] In practical implementation, the system can obtain second prompt words corresponding to the second sample data based on the second sample search results using a preset second prompt word template. This second prompt word template includes prompt information to guide the large model in generating summary content in the target language based on the second sample search results. Subsequently, based on these second prompt words, the system uses a third general-purpose large model with a high parameter scale to generate the second sample summary content corresponding to the second sample search results, thereby automatically constructing the second sample data.

[0107] After acquiring the second sample data, the system can train a second general large model with a smaller parameter scale based on the second sample data until the model converges, thus obtaining the initial large model. This stage aims to establish the basic ability of the model to accurately translate and extract content in cross-language scenarios, providing a more stable initial large model for subsequent training combined with search requests.

[0108] Next, the system enters the second stage of model training. In this stage, the system can first acquire multiple first sample data. Each first sample data can include a first sample search request, a first sample search result corresponding to the first sample search request, and a first sample summary of the first sample search result that can satisfy the search intent of the first sample search request. At the same time, the first sample data also includes sample core answer content that can answer the search intent from the first sample summary content. The number of characters in the sample core answer content is less than the number of characters in the first sample summary content.

[0109] In some exemplary embodiments, the system can construct a first prompt word using a preset first prompt word template, a sample search request, and a first sample search result corresponding to the first sample search request. It can then use a first general large model with a parameter size larger than the initial large model to perform a summary content generation operation for the first sample search result to obtain the first sample summary content and the sample core answer content therein, thereby constructing the first sample data.

[0110] The aforementioned first prompt word template can be a chain of prompt word templates, which can specifically include prompt information for guiding the large model to analyze the search intent corresponding to the first sample search request, prompt information for guiding the large model to generate summary content in the target language based on the search intent, prompt information for guiding the large model to determine the core answer content that can answer the search intent in the summary content, and prompt information for guiding the large model to reflect on the predicted summary content.

[0111] After obtaining the first sample data, the system can train the initial large model obtained in the first stage based on the first sample data until the model converges, thus initially obtaining a large model for summarizing. This stage improves the matching degree between the target summary content and the search intent, building upon the model's existing summarizing capabilities.

[0112] The system then proceeds to the third stage of model training. In this stage, the system can first select first sample data that meets preset conditions from multiple first sample data sets to serve as third sample data.

[0113] To meet the requirements of online display specifications, the preset conditions may specifically include: the number of characters in the first sample summary content is within a preset range; the first sample summary content does not contain redundant content, which includes at least one of the content used to describe the first sample search request and the title of the first sample search result; the core answer content is within the first preset number of characters of the first sample summary content; and the core answer content does not overlap with other content in the first sample summary content except for the core answer content.

[0114] After selecting high-quality third-party sample data using the aforementioned preset conditions, the system can further train the summary generation model based on this third-party sample data until the model converges, thus obtaining an updated summary generation model. This stage enables the updated summary generation model to stably output high-quality summary content that meets length requirements, does not contain redundant content, and places the core answer content at the beginning.

[0115] In some embodiments, the updated summary can be used to generate a large model as the final large model to be deployed online.

[0116] In some embodiments, after obtaining the final updated summary generation model through the above three-stage progressive training, the summary generation model can be further processed using an inference acceleration framework and model quantization techniques before deployment. By applying model quantization and acceleration techniques, the computational resource overhead of the large model can be effectively reduced and inference efficiency improved, enabling high-throughput online inference of the summary generation model, thereby meeting the real-time summary generation needs of a large number of users in actual search engine application scenarios.

[0117] In some embodiments, such as Figure 7 As shown, a summary generation apparatus 700 based on a large model is provided. The apparatus 700 includes: a search unit 710 configured to perform a data search based on a search request in response to receiving a search request to obtain at least one original search result; a generation unit 720 configured to generate a target summary content of the original search results that can satisfy the search intent of the search request for each of the at least one original search results using a summary generation large model; and a display unit 730 configured to display a search result display page in response to the search request, wherein the search result display page displays the search result entry and target summary content corresponding to each of the at least one original search results.

[0118] The operations performed by units 710 to 730 in the large model-based summary generation device 700 and the effects they can achieve are similar to steps S201 to S203 in the large model-based summary generation method, and will not be described in detail here.

[0119] In some embodiments, the generation unit may be further configured to: analyze the search intent corresponding to the search request based on the search request and the original search results; and generate a target summary of the original search results that satisfies the search intent based on the search intent and the original search results.

[0120] In some embodiments, generating target summary content that satisfies the search intent based on the search intent and the original search results includes: generating target summary content in the target language based on the search intent and the original search results in the original language.

[0121] In some embodiments, the generation unit may be further configured to: determine, based on the search intent, core answer content that can answer the search intent in the target summary content, wherein the number of characters in the core answer content is less than the number of characters in the target summary content; and wherein the display style of the core answer content corresponding to each original search result in at least one original search result displayed on the search results display page is different from the display style of the corresponding target summary content.

[0122] In some embodiments, the generation unit may be further configured to: detect the target summary content to determine whether the target summary content satisfies the search intent; and in response to determining that the target summary content does not satisfy the search intent, re-execute the generation of the target summary content until it is determined that the regenerated target summary content satisfies the search intent or the number of re-executions exceeds a preset number.

[0123] In some embodiments, such as Figure 8 As shown, a training apparatus 800 for a large-scale summary generation model is provided. The apparatus 800 includes: a first acquisition unit 810 configured to acquire first sample data, the first sample data including a first sample search request, a first sample search result corresponding to the first sample search request, and first sample summary content of the first sample search result that can satisfy the search intent of the first sample search request; a second acquisition unit 820 configured to acquire first prompt words corresponding to the first sample data based on the first sample search request and the first sample search result, using a preset first prompt word template, wherein the first prompt word template includes prompt information for guiding the large-scale model to generate summary content of the first sample search result that can satisfy the search intent of the first sample search request based on the first sample search request and the first sample search result; a first execution unit 830 configured to perform a summary content generation operation for the first sample search result using an initial large-scale model based on the first prompt word, to obtain first predicted summary content corresponding to the first sample search result; and a first training unit 840 configured to train the initial large-scale model based on the first sample summary content and the first predicted summary content, to obtain a large-scale summary generation model.

[0124] The operations performed by units 810 to 840 in the training device 800 for the above-mentioned large-scale summary generation model and the effects they can achieve are similar to steps S401 to S404 in the training method for the above-mentioned large-scale summary generation model, and will not be described in detail here.

[0125] In some embodiments, the above-mentioned summary content generation operation may include: analyzing the search intent corresponding to the first sample search request based on the first sample search request and the first sample search result; and generating a first predicted summary content of the first sample search result that can satisfy the search intent based on the search intent and the first sample search result.

[0126] In some embodiments, generating a first predicted summary content that satisfies the first sample search result based on the search intent may include: generating a first predicted summary content in the target language based on the search intent and the original search result in the original language.

[0127] In some embodiments, the first sample data may further include sample core answer content that can respond to the search intent in the first sample summary content, wherein the number of characters in the sample core answer content is less than the number of characters in the first sample summary content, and the first prompt word template may further include prompt information for guiding the large model to determine the core answer content that can respond to the search intent in the first predicted summary content based on the search intent, and the summary content generation operation may further include: determining the predicted core answer content for responding to the search intent in the first predicted summary content based on the search intent; and wherein the first training unit may be further configured to: train an initial large model based on the first sample summary content, the sample core answer content, the first predicted summary content, and the predicted core answer content to obtain a summary generation large model.

[0128] In some embodiments, the first prompt word template may further include prompt information for guiding the large model to reflect on the first predicted summary content, and the summary content generation operation may further include: detecting the first predicted summary content to determine whether the first predicted summary content satisfies the search intent; and in response to determining that the first predicted summary content does not satisfy the search intent, re-executing the generation of the first predicted summary content until it is determined that the regenerated first predicted summary content satisfies the search intent or the number of re-executions exceeds a preset number.

[0129] In some embodiments, the first acquisition unit may be further configured to: based on the first prompt word, use a first general large model to perform a summary content generation operation for the first sample search result, so as to obtain the first sample summary content corresponding to the first sample search result, wherein the parameter size of the first general large model is larger than the parameter size of the initial large model.

[0130] In some embodiments, obtaining the initial large model may include: obtaining second sample data, which includes second sample search results in the original language and second sample summary content in the target language corresponding to the second sample search results; obtaining second prompt words corresponding to the second sample data based on the second sample search results using a preset second prompt word template, wherein the second prompt word template includes prompt information for guiding the large model to generate summary content in the target language based on the second sample search results; generating second predicted summary content in the target language corresponding to the second sample search results using a second general large model based on the second prompt words; and training a second general large model based on the second sample summary content and the second predicted summary content to obtain the initial large model.

[0131] In some embodiments, obtaining the second sample data may include: generating a second sample summary content corresponding to the second sample search result based on the second prompt word and using a third general large model, wherein the parameter size of the third general large model is larger than the parameter size of the second general large model.

[0132] In some embodiments, the number of first sample data is multiple, and the training apparatus for the above-mentioned large-scale summary generation model may further include: a filtering unit configured to filter first sample data that meets preset conditions from the multiple first sample data to serve as third sample data, wherein the preset conditions may include at least one of the following: the number of characters in the first sample summary content is within a preset number range; and the first sample summary content does not contain redundant content, the redundant content including at least one of the content used to describe the first sample search request and the title of the first sample search result; and a second training unit configured to train the large-scale summary generation model based on the third sample data to update the large-scale summary generation model.

[0133] In some embodiments, the first sample data may further include sample core answer content that can answer the search intent in the first sample summary content, wherein the number of characters in the sample core answer content is less than the number of characters in the first sample summary content, and the preset conditions may further include at least one of the following: the sample core answer content is located within the first preset number of characters of the first sample summary content; and the sample core answer content does not have any duplicate content with other content in the first sample summary content besides the sample core answer content.

[0134] The collection, storage, use, processing, transmission, provision, and disclosure of any type of information, such as user personal information, in this technical solution comply with relevant laws and regulations and do not violate public order and good morals.

[0135] According to embodiments of this disclosure, an electronic device, a readable storage medium, and a computer program product are also provided.

[0136] refer to Figure 9 The present invention describes a structural block diagram of an electronic device 900 that can serve as a server or client of the present disclosure, which is an example of a hardware device that can be applied to various aspects of the present disclosure. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0137] like Figure 9As shown, the electronic device 900 includes a computing unit 901, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 902 or a computer program loaded from a storage unit 908 into a random access memory (RAM) 903. The RAM 903 may also store various programs and data required for the operation of the electronic device 900. The computing unit 901, ROM 902, and RAM 903 are interconnected via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0138] Multiple components in electronic device 900 are connected to I / O interface 905, including: input unit 906, output unit 907, storage unit 908, and communication unit 909. Input unit 906 can be any type of device capable of inputting information to electronic device 900. Input unit 906 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of electronic device, and can include, but is not limited to, a mouse, keyboard, touchscreen, trackpad, trackball, joystick, microphone, and / or remote control. Output unit 907 can be any type of device capable of presenting information, and can include, but is not limited to, a monitor, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 908 can include, but is not limited to, hard disk and optical disk. Communication unit 909 allows electronic device 900 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and can include, but is not limited to, modems, network cards, infrared communication devices, wireless communication transceivers, and / or chipsets, such as Bluetooth devices, 802.11 devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0139] The computing unit 901 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 901 performs the various methods and processes described above, such as the large model-based summarization method or the large model training method for summarization generation of the present disclosure. For example, in some embodiments, the large model-based summarization method or the large model training method for summarization generation of the present disclosure can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as storage unit 908. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 900 via ROM 902 and / or communication unit 909. When the computer program is loaded into RAM 903 and executed by the computing unit 901, one or more steps of the large model-based summarization method or the large model training method for summarization generation of the present disclosure described above can be performed. Alternatively, in other embodiments, the computing unit 901 may be configured by any other suitable means (e.g., by means of firmware) to perform the large model-based summarization method or the training method for summarizing large models of this disclosure.

[0140] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0141] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0142] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0143] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0144] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0145] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0146] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0147] While embodiments or examples of this disclosure have been described with reference to the accompanying drawings, it should be understood that the methods, systems, and devices described above are merely exemplary embodiments or examples, and the scope of the invention is not limited by these embodiments or examples, but only by the granted claims and their equivalents. Various elements in the embodiments or examples may be omitted or replaced by their equivalents. Furthermore, the steps may be performed in a different order than that described in this disclosure. Further, various elements in the embodiments or examples may be combined in various ways. Importantly, as the technology evolves, many elements described herein can be replaced by equivalents that appear after this disclosure.

Claims

1. A summarization generation method based on a large model, the method comprising: In response to receiving a search request, perform a data search based on the search request to obtain at least one original search result; For each of the at least one original search results, a summary generation model is used to generate a target summary of the original search results that satisfies the search intent of the search request, based on the search request and the original search results. as well as A search results display page is displayed in response to the search request, showing the search results entry and target summary content corresponding to each of the at least one original search results.

2. The method of claim 1, wherein, The target summary content of the original search results, which is generated based on the search request and the original search results and can satisfy the search intent of the search request, includes: Based on the search request and the original search results, analyze the search intent corresponding to the search request; and Based on the search intent and the original search results, a target summary content is generated that satisfies the search intent and corresponds to the original search results.

3. The method of claim 2, wherein, The process of generating the target summary content that satisfies the search intent based on the search intent and the original search results includes: Based on the search intent and the original search results in the original language, the target summary content in the target language is generated.

4. The method of any one of claims 1 to 3, wherein, The process of generating a target summary of the original search results that satisfies the search intent of the search request, based on the search request and the original search results, further includes: Based on the search intent, core answer content that can respond to the search intent is determined from the target summary content, wherein the number of characters in the core answer content is less than the number of characters in the target summary content; and wherein, The display style of the core answer content corresponding to each of the at least one original search results displayed on the search results display page is different from the display style of the corresponding target summary content.

5. The method of any one of claims 1 to 4, wherein, The process of generating a target summary of the original search results that satisfies the search intent of the search request, based on the search request and the original search results, further includes: The target summary content is inspected to determine whether it satisfies the search intent; and In response to the determination that the target summary content does not satisfy the search intent, the generation of the target summary content is re-executed until the regenerated target summary content satisfies the search intent or the number of re-executions exceeds a preset number.

6. A method for training a large model for summarizing data, the method comprising: Obtain first sample data, which includes a first sample search request, a first sample search result corresponding to the first sample search request, and a first sample summary of the first sample search result that can satisfy the search intent of the first sample search request. Based on the first sample search request and the first sample search result, a first prompt word corresponding to the first sample data is obtained using a preset first prompt word template. The first prompt word template includes prompt information for guiding the large model to generate a summary of the first sample search result that can satisfy the search intent of the first sample search request based on the first sample search request and the first sample search result. Based on the first prompt word, a summary content generation operation is performed on the first sample search result using the initial large model to obtain the first predicted summary content corresponding to the first sample search result; and Based on the first sample summary content and the first predicted summary content, the initial large model is trained to obtain the summary generation large model.

7. The method of claim 6, wherein, The summary content generation operation includes: Based on the first sample search request and the first sample search result, analyze the search intent corresponding to the first sample search request; and Based on the search intent and the first sample search results, a first predicted summary content is generated that satisfies the search intent of the first sample search results.

8. The method of claim 7, wherein, The first predicted summary content, which generates the first sample search result that satisfies the search intent based on the search intent and the first sample search result, includes: Based on the search intent and the original search results in the original language, the first predicted summary content in the target language is generated.

9. The method of any one of claims 6-8, wherein, The first sample data also includes sample core answer content that can respond to the search intent in the first sample summary content. The number of characters in the sample core answer content is less than the number of characters in the first sample summary content. The first prompt word template also includes prompt information to guide the large model to determine the core answer content that can respond to the search intent in the first predicted summary content based on the search intent. The summary content generation operation also includes: Based on the search intent, predictive core answer content for responding to the search intent is determined from the first predicted summary content; and wherein... The step of training the initial large model based on the first sample summary content and the first predicted summary content to obtain the summary generation large model includes: Based on the first sample summary content, the sample core answer content, the first predicted summary content, and the predicted core answer content, the initial large model is trained to obtain the summary generation large model.

10. The method of any one of claims 6-9, wherein, The first prompt word template also includes prompting information to guide the large model to reflect on the first predicted summary content, and the summary content generation operation further includes: The first predicted summary content is examined to determine whether it satisfies the search intent; and In response to the determination that the first predicted summary content does not satisfy the search intent, the generation of the first predicted summary content is re-executed until the regenerated first predicted summary content satisfies the search intent or the number of re-executions exceeds a preset number.

11. The method of any one of claims 6 to 10, wherein, The acquisition of the first sample data includes: Based on the first prompt word, the first general large model is used to perform the summary content generation operation for the first sample search result to obtain the first sample summary content corresponding to the first sample search result, wherein the parameter size of the first general large model is larger than the parameter size of the initial large model.

12. The method of any one of claims 6-11, wherein, The acquisition of the initial large model includes: Obtain second sample data, which includes second sample search results in the original language and second sample summary content in the target language corresponding to the second sample search results; Based on the second sample search results, a second prompt word corresponding to the second sample data is obtained using a preset second prompt word template. The second prompt word template includes prompt information for guiding the large model to generate summary content in the target language based on the second sample search results. Based on the second prompt word, using the second general large model, a second predicted summary content for the target language corresponding to the second sample search results is generated; and Based on the second sample summary content and the second predicted summary content, the second general large model is trained to obtain the initial large model.

13. The method of claim 12, wherein, The acquisition of the second sample data includes: Based on the second prompt word, the third general large model is used to generate the second sample summary content corresponding to the second sample search result, wherein the parameter size of the third general large model is larger than the parameter size of the second general large model.

14. The method of any one of claims 6-13, wherein, The number of the first sample data is multiple, and the method further includes: Select first sample data that meets preset conditions from multiple first sample data sets to serve as third sample data, wherein the preset conditions include at least one of the following: The number of characters in the first sample summary content is within a preset range; and The first sample summary content does not contain redundant content, which includes at least one of the content describing the first sample search request and the title of the first sample search result; and Based on the third sample data, the summary generation model is trained to update the summary generation model.

15. The method of claim 14, wherein, The first sample data also includes sample core answer content that can answer the search intent from the first sample summary content, wherein the number of characters in the sample core answer content is less than the number of characters in the first sample summary content, and the preset condition also includes at least one of the following: The core answer content of the sample is located within the first preset number of characters of the first sample summary content; and The core answer content of the sample does not overlap with any other content in the first sample summary, except for the core answer content of the sample.

16. A summary generation apparatus based on a large model, the apparatus comprising: The search unit is configured to perform a data search based on a received search request in response to obtaining at least one original search result; The generation unit is configured to, for each of the at least one original search results, utilize a summary generation model to generate a target summary content of the original search results that satisfies the search intent of the search request, based on the search request and the original search results. as well as The display unit is configured to display a search results display page in response to the search request, wherein the search results display page displays the search results entry and target summary content corresponding to each of the at least one original search results.

17. A training apparatus for generating large models of summaries, the apparatus comprising: The first acquisition unit is configured to acquire first sample data, which includes a first sample search request, a first sample search result corresponding to the first sample search request, and a first sample summary content of the first sample search result that can satisfy the search intent of the first sample search request. The second acquisition unit is configured to acquire a first prompt word corresponding to the first sample data based on the first sample search request and the first sample search result, using a preset first prompt word template. The first prompt word template includes prompt information for guiding the large model to generate a summary of the first sample search result that can satisfy the search intent of the first sample search request based on the first sample search request and the first sample search result. The first execution unit is configured to, based on the first prompt word, use an initial large model to perform a summary content generation operation for the first sample search result, so as to obtain the first predicted summary content corresponding to the first sample search result; and The first training unit is configured to train the initial large model based on the first sample summary content and the first predicted summary content to obtain the summary generation large model.

18. An electronic device comprising: At least one processor; as well as A memory that is communicatively connected to the at least one processor; in The memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-15.

19. A non-transitory computer readable storage medium having stored thereon computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-15.

20. A computer program product comprising a computer program, wherein, The computer program, when executed by a processor, implements the method of any one of claims 1-15.

Images