Search optimization method, device and medium based on large model

Abstract

The present application discloses a search optimization method, device and medium based on a large model, which relates to the field of data processing. The method includes: receiving query information input by a user and preprocessing the query information; performing semantic analysis on the query information based on a pre-trained deep learning model to extract the key information contained therein; using the query information and key information as input, generating a related first search result through a pre-trained large model; obtaining the business entity corresponding to the query information, determining the compressed structure information corresponding to the business entity, and optimizing the first search result based on the compressed structure information to obtain a second search result; and sorting and displaying the second search result. Accurately analyze the deep intention and context of the user's query, reduce irrelevant results caused by keyword ambiguity, and be able to quickly generate candidate results based on the reasoning ability of the large model to optimize search efficiency.

Description

Technical Field

[0001] The present application relates to the field of data processing, and specifically to a search optimization method, device and medium based on a large model. Background Art

[0002] With the development of Internet technology, Internet information is showing an explosive growth trend.

[0003] As an important tool for Internet information retrieval, the accuracy and efficiency of search engines directly affect user experience and the effective dissemination of Internet information. However, with the continuous increase in information volume and the increasing complexity of query requirements, traditional search engine technology has been unable to meet users' demand for high-quality search results.

[0004] Traditional search engines mainly rely on keyword matching and page weight ranking algorithms to provide search results. This approach is incapable of handling massive amounts of information and complex queries. On the one hand, simple keyword matching is difficult to accurately capture the deep semantics of user queries, resulting in a large number of irrelevant or low-quality results; on the other hand, traditional page weight ranking algorithms often only consider surface features such as external links and internal structures of web pages, while ignoring important factors such as content quality and user experience. Summary of the invention

[0005] In order to solve the above problems, this application proposes a search optimization method based on a large model, including: Receiving query information input by a user and preprocessing the query information; Based on a pre-trained deep learning model, semantic analysis is performed on the query information to extract key information contained therein; Taking the query information and the key information as input, generating a relevant first search result through a pre-trained large model; Acquire a business entity corresponding to the query information, determine compression structure information corresponding to the business entity, and optimize the first search result based on the compression structure information to obtain a second search result; The second search results are sorted and displayed.

[0006] In one example, preprocessing the query information specifically includes: According to the query information, obtain a corresponding character string; Performing text preprocessing on the character string; The character string is segmented and part-of-speech tagging is performed on the segmentation result.

[0007] In one example, based on a pre-trained deep learning model, semantic analysis is performed on the query information to extract key information contained therein, specifically including: Determine, based on a pre-trained deep learning model, a minimum text unit contained in a word segmentation result of the query information; Performing position encoding on the minimum text unit; Outputting a corresponding context vector for the minimum text unit; the context vector includes a first vector of aggregated semantic representation for intent classification and a second vector of fine-grained semantic features for entity recognition; Intent classification is performed according to the first vector, and key information is extracted according to the intent classification and the second vector.

[0008] In one example, obtaining the business entity corresponding to the query information and determining the compression structure information corresponding to the business entity specifically includes: Generate a corresponding first business entity according to the key information; Based on the corresponding relationship in the local database, the key information is expanded to generate a corresponding second business entity for each first business entity; According to the corresponding relationship, each first business entity and its corresponding second business entity are compressed to generate corresponding compression structure information.

[0009] In one example, optimizing the first search result based on the compression structure information to obtain the second search result specifically includes: For each compression structure information corresponding to the first business entity, select at least one second business entity as a selected entity corresponding to the compression structure information; According to the selected entities corresponding to all the compressed structure information, an entity combination is obtained; Performing a correlation analysis on the first search result according to the entity combination, and taking the first search result whose corresponding first correlation score is higher than a preset score as a sub-search result corresponding to the entity combination; For each sub-search result, several search results with the highest first relevance score are selected respectively, and the second search result is obtained by combining them after removing duplicates.

[0010] In one example, the method further includes: Determine the number of results of the second search result obtained by combining after deduplication; If the number of results is higher than a preset number range, then based on the number of corresponding second business entities in the entity combination, the corresponding sub-search results are deleted in order from low to high until the number of results meets the preset number range; If the number of results is lower than the preset number range, then based on the number of corresponding second business entities in the entity combination, corresponding sub-search results are added in descending order until the number of results meets the preset number range, or all sub-search results are added.

[0011] In one example, sorting and displaying the second search results specifically includes: Obtain access rights to the user; Filtering the second search result according to the access permission to obtain a third search result; Obtaining a second relevance score of the large model for the first search result, and obtaining a first ranking corresponding to the third search result according to the second relevance score; Obtaining access data of the third search result, and obtaining a corresponding quality score according to the access data, and adjusting the first ranking according to the quality score to obtain a second ranking; A corresponding display method is selected, and based on the display method, the third search results are sorted and displayed according to the second sorting.

[0012] In one example, the method further includes: Obtaining user operation information on the third search result; According to the operation information, obtaining the user's personalized behavior information; According to the personalized behavior information, the sorting method and display method corresponding to the user are personalized and optimized.

[0013] On the other hand, the present application also proposes a search optimization device based on a large model, comprising: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the large model-based search optimization method as described in any of the above examples.

[0014] On the other hand, the present application also proposes a non-volatile computer storage medium storing computer executable instructions, wherein the computer executable instructions are configured to be: a large model-based search optimization method as described in any of the above examples.

[0015] The search optimization method based on a large model proposed in this application can bring the following beneficial effects: 1. Through deep model + big model, deep semantic understanding can be performed to accurately analyze the deep intention and context of user queries, reduce irrelevant results caused by keyword ambiguity, and quickly generate candidate results based on the reasoning ability of the big model to optimize search efficiency.

[0016] 2. Combined with dynamic optimization of the business entity compression structure, it ensures that search results are highly matched with user needs and business rules, improves the relevance of search results, and supports automatic parsing and response of multiple conditions and implicit needs (for example, price ranges, scenario preferences), enhancing the adaptability of complex queries. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 A schematic diagram of a flow chart of a search optimization method based on a large model in an embodiment of the present application; Figure 2 A schematic diagram of a flow chart of a search optimization method based on a large model in one scenario in an embodiment of the present application; Figure 3 It is a schematic diagram of a search optimization device based on a large model in an embodiment of the present application. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0019] The technical solutions provided by various embodiments of the present application are described in detail below in conjunction with the accompanying drawings.

[0020] Search optimization can be performed based on deep learning and natural language processing technologies to improve the performance of search engines. By building deep learning models, in-depth understanding and analysis of user queries and web page content can be performed to provide more accurate and valuable search results.

[0021] However, there are still the following shortcomings: 1. The contradiction between model complexity and performance In order to capture the deep semantics of user queries and web page content, complex deep learning models need to be built. However, the increase in model complexity will lead to an increase in computational costs and training difficulty, thus limiting the performance of the model in practical applications.

[0022] 2. Data sparsity and cold start problems For newly emerged queries or web pages, it is difficult to provide accurate search results due to the lack of sufficient historical data and user feedback. This data sparsity and cold start problem limits the generalization ability and adaptability of search engines.

[0023] 3. Continuous optimization of user experience Although it can provide more accurate search results, the user experience still needs to be improved. For example, how to sort and display search results according to users' personalized needs and preferences, and how to provide users with a more intuitive and convenient search experience.

[0024] like Figure 1 and Figure 2 As shown, the embodiment of the present application provides a search optimization method based on a large model, including: S101: receiving query information input by a user, and preprocessing the query information.

[0025] When a user enters a query through a search box or other interface of a search engine, the user's query information is received to obtain a corresponding character string.

[0026] Preprocess the query information (query string), including text preprocessing, such as removing spaces, punctuation marks, special characters and other irrelevant information for subsequent processing.

[0027] Use word segmentation tools to segment the query string and break the query into several meaningful words or phrases. Based on word segmentation, further perform part-of-speech tagging to mark the part of speech for each word or phrase in order to better understand the intent of the query.

[0028] S102: Based on a pre-trained deep learning model, perform semantic analysis on the query information to extract key information contained therein.

[0029] Call pre-trained deep learning models, such as BERT, Transformer, etc., to perform deep semantic analysis on the pre-processed query information.

[0030] The deep learning model can analyze the keywords, phrases, and sentence structures in the query to understand the intent and context of the query. The model also extracts key information from the query, such as topic, category, time, etc., to provide important clues for subsequent searches.

[0031] Specifically, you can choose BERT-base (12-layer Transformer encoder, 12 attention heads, hidden layer dimension 768) as the model architecture, and set a multi-task learning head. Add two parallel task branches after the BERT output layer, namely the intent classification head (including fully connected layer + Softmax), which outputs the query intent category (for example, product search, knowledge question and answer, event query, etc.), and the sequence annotation head (including BiLSTM + CRF), which annotates the key entities in the query (for example, time, place, product name, etc.).

[0032] Based on the general BERT, according to the business scenarios corresponding to the current search function, vertical domain corpus (such as e-commerce product descriptions, corporate documents) is used as training parameters for secondary pre-training to enhance domain semantic understanding.

[0033] First, a pre-training phase is performed using general corpus and vertical domain corpus for masked language modeling (MLM) + next sentence prediction (NSP), and then a fine-tuning phase is performed using annotated search query datasets to fine-tune the model parameters.

[0034] The training parameters may be as shown in Table 1 below.

[0035] Table 1. Deep learning model training parameters

[0036] Based on this, when extracting key information, the smallest text unit contained in the word segmentation results of the query information is determined based on the pre-trained deep learning model, and the smallest text unit is positionally encoded. Position encoding includes position encoding in a single sentence and segment encoding.

[0037] Output the corresponding context vector for the smallest text unit. The context vector includes a first vector of aggregated semantic representation for intent classification and a second vector of fine-grained semantic features for entity recognition.

[0038] Taking the above text as an example, it outputs the context vector (which corresponds to 768 dimensions) of each smallest text unit (Token). The first vector (CLS vector) is used for the aggregated semantic representation of intent classification, and the second vector (Token vector) is used for the fine-grained semantic features of entity recognition.

[0039] At this time, the intent is classified according to the first vector, and the first vector is input into the fully connected layer. According to the pre-set multiple intent categories, the corresponding intent category probability is calculated. According to the intent classification and the second vector, key information is extracted, and each second vector is input into the BiLSTM-CRF layer for sequence labeling, so as to extract the corresponding entities, attributes, conditions, etc., which can be directly used as key information, or further screened based on the blacklist, whitelist, etc. in the vertical field to which it belongs, or the corresponding weights are set, so as to obtain the corresponding content according to the second vector and extract key information.

[0040] S103: Taking the query information and the key information as input, generating a relevant first search result through a pre-trained large model.

[0041] The preprocessed query information and key information are passed as input to the pre-trained big model. The big model uses its powerful capabilities to perform deep learning and reasoning on the input, predict the web pages or documents most relevant to the query, and form the first search result. The big model can output a ranked result list containing the identifiers and related information of the web pages or documents most relevant to the query, and rank the first search result and perform relevance analysis.

[0042] Among them, the large model can use the Large Language Model (LLM), which can directly use the open source large language model, or perform further training in the corresponding vertical field based on demand.

[0043] S104: Acquire a business entity corresponding to the query information, determine compression structure information corresponding to the business entity, and optimize the first search result based on the compression structure information to obtain a second search result.

[0044] Business entities refer to entities that have a certain relevance to the query information. For example, when the query information is to search for products, business entities can include the product, its supporting products, the manufacturer and seller of the product, etc. When the query information is to search for knowledge, business entities can be the knowledge itself, related knowledge, the discoverer of the knowledge, the current research status of the knowledge, etc.

[0045] The compressed structure information refers to the compressed information obtained by integrating and compressing the attributes and relationships corresponding to the business entities (this information may be stored in a database or a network).

[0046] The compressed structure information is applied to the first search result to further filter and optimize the result. For example, the search results that do not meet the conditions can be filtered out according to the attributes of the entity, or the search results can be clustered or sorted according to the relationship between entities.

[0047] Specifically, a corresponding first business entity is generated according to the key information. Based on the corresponding relationship in the local database, the key information is expanded, so as to generate a corresponding second business entity for each first business entity.

[0048] When the key information contains multiple items, the corresponding first business entities may be multiple. For example, when the query information entered by the user is: "Recommend mobile phones within 5,000 yuan that can be purchased online", the identified key information may include: 5,000 yuan, online shopping, mobile phone. At this time, for each key information, the corresponding first business entity is generated, which are: price (less than 5,000 yuan), e-commerce platform, and mobile phone.

[0049] The local database may store corresponding relationships in advance, that is, the directions in which each type of business entity can be expanded. Generally speaking, the structural information can be divided into two types: attributes and relationships. Attributes are used to describe the properties of the first business entity itself, while relationships are used to describe the content related to the first business entity. Of course, for some first business entities, only one type of expansion can be performed based on demand.

[0050] In addition, for the collected personal habits of the user, the corresponding second business entity can be selected according to the personal habits.

[0051] For example, for "price (less than 5,000 yuan)", the corresponding entity type is the price type, and the corresponding structure information may include attributes such as pricing type (including original price, promotional price, member price, etc.), promotional time period, etc., and relationships may include price range, purchase method (including retail, bulk purchase, etc.), etc. The second business entities that can be expanded include: promotional price within 5,000 yuan, member price within 5,000 yuan, current promotional time period and price within 5,000 yuan, bulk purchase price within 5,000 yuan, etc.

[0052] The entity type corresponding to the "e-commerce platform" is the network platform type, and the attributes in the corresponding structural information may include: platform name, delivery range, payment method, platform reputation, etc. Therefore, the second business entities that can be expanded include: e-commerce platforms with platform name A, e-commerce platforms whose platform delivery range meets the user's positioning, e-commerce platforms with high platform reputation, etc. Among them, user positioning and user's commonly used payment methods can be obtained in advance. Among them, the name of the e-commerce platform can be determined by the user's historical software usage records. If it cannot be determined, one or more commonly used e-commerce platforms can be selected. Other subsequent second business entities can also adopt a similar method.

[0053] For the entity type corresponding to "mobile phone" is the commodity type, and the corresponding structural information may include attributes such as mobile phone brand, mobile phone performance, mobile phone price, etc., and relationships such as matching mobile phone case price, mobile phone software ecological network, etc. The second business entities that can be expanded include: mobile phones of brand B, mobile phones with high performance scores, mobile phones of type C in mobile phone software ecological network, etc. Among them, since the price of the mobile phone has been used as the first business entity, the corresponding avoidance is performed in the second business entity.

[0054] Based on this, for each first business entity, a number of second business entities related to it are generated, and more second business entities are limited. When the number of second business entities is large, only a fixed number (such as 3) of commonly used second business entities can be selected (or personalized selection can be made based on the user's previous search records), thereby ensuring the speed of subsequent calculation processes.

[0055] According to the corresponding relationship, each first business entity and its corresponding second business entity are compressed to generate corresponding compression structure information.

[0056] At this time, when optimizing the first search result, for each compressed structure information corresponding to the first business entity, at least one second business entity can be selected as the selected entity corresponding to the compressed structure information. That is to say, at least one second business entity is selected in each compressed structure information to describe the first business entity in more detail. Of course, multiple entities can also be selected for combination as selected entities. For example, when the first business entity is a mobile phone, a mobile phone of brand B and a mobile phone with a high performance score are selected as the second business entity. The combination results in a mobile phone of brand B and a mobile phone with a high performance score, which is used as the corresponding selected entity.

[0057] According to the selected entities corresponding to all the compressed structure information, the entity combination is obtained by combining them. Here, all entity combinations can be obtained by traversing. For example, there are two first business entities, namely entity A and entity B, which correspond to the second business entities A1 and A2, and B1 and B2 respectively. At this time, the entity combination can include: (A1, B1) (A1, B2) (A2, B1) (A2, B2) (A1, A2, B1) (A1, A2, B2) (A1, B1, B2) (A2, B1, B2) (A1, A2, B1, B2). Among them, if there are conflicts between some entity combinations, for example, there is a conflict between the mobile phone of brand A and the mobile phone of brand B, and they can be deleted based on the needs. Of course, since this is for search and correlation analysis, the selection of conflicting entity combinations does not hinder the execution of corresponding work, so they can also be retained based on needs.

[0058] A correlation analysis is performed on the first search result based on the entity combination, and the first search result with a corresponding first correlation score higher than a preset score is used as the sub-search result corresponding to the entity combination. The correlation analysis can be determined by a large language model or keyword matching. The first search result contains multiple websites, documents, etc., and a correlation analysis can be performed on each entity combination. Each entity combination is sorted according to its corresponding first correlation score to obtain its corresponding sub-search result. At this time, due to the different entity combinations in each sub-search result, the selected sub-search results will usually be different.

[0059] For each sub-search result, several search results with the highest first relevance score are selected. At this time, there is usually a search result whose first relevance score with multiple entity combinations is very high and there are duplications, so the second search result is obtained by removing duplicates and combining them.

[0060] In the second search result at this time, for different entity combinations, the key information is extended in different directions, so as to perform corresponding correlation analysis, and select highly relevant content in each direction to provide users with comprehensive search results.

[0061] Of course, the number of results of the second search results obtained by combining after deduplication may also be determined to prevent the number of results from being too large or too small, thereby affecting the user experience.

[0062] A corresponding preset quantity interval is set in advance. If the number of results is higher than the preset quantity interval, that is, there are too many search results, then based on the number of corresponding second business entities in the entity combination, the corresponding sub-search results are deleted in order from low to high, that is, the search results corresponding to the second business entities with fewer numbers are deleted in a limited manner (when deleting each search result, the search results with lower first relevance scores can also be deleted first) until the number of results meets the preset quantity interval.

[0063] If the number of results is lower than the preset number range, the corresponding sub-search results are added in descending order based on the number of corresponding second business entities in the entity combination, that is, more search results that are relevant to the second business entities and contain more content are added first, until the number of results meets the preset number range, or all sub-search results are added.

[0064] S105: Sorting and displaying the second search results.

[0065] The second search results mentioned above may be sorted and displayed according to the first relevance score, the number of second business entities in the entity combination, and the like.

[0066] Specifically, the user's access rights may be obtained, and the second search results may be filtered according to the access rights to obtain the third search results.

[0067] Verify the user's access rights to ensure that the user can only see the search results that they have access to. Filter the search results based on the user's role, permissions, and other information to exclude web pages or documents that do not meet the permission requirements. The filtered results will be re-sorted to ensure that the user can only see the results that they have access to and are most relevant to the query.

[0068] The second relevance score of the large model for the first search result is obtained, and the first ranking corresponding to the third search result is obtained according to the second relevance score. When searching, the large model can give not only the first search result but also the corresponding relevance score. At this time, the third search result is still part of the first search result, so its corresponding second relevance score can also be obtained and sorted according to high and low.

[0069] Access data of the third search result is obtained, and a corresponding quality score is obtained according to the access data, and the first ranking is adjusted according to the quality score to obtain a second ranking.

[0070] On the basis of the preliminary first ranking, the ranking results are fine-tuned by comprehensively considering factors such as web page quality and user feedback to obtain the second ranking. For example, the ranking results can be adjusted according to information such as the click-through rate, browsing time, and evaluation of the web page. Search results with high click-through rate, long browsing time, and good evaluation can be promoted in the first ranking to generate the corresponding second ranking.

[0071] A corresponding display method is selected, and based on the display method, the third search results are sorted and displayed according to the second sorting. The display method may include a list, a summary, a picture, and other forms, so that the user can quickly browse and understand the search results.

[0072] 1. Through deep model + big model, deep semantic understanding can be performed to accurately analyze the deep intention and context of user queries, reduce irrelevant results caused by keyword ambiguity, and quickly generate candidate results based on the reasoning ability of the big model to optimize search efficiency.

[0073] 2. Combined with dynamic optimization of the business entity compression structure, it ensures that search results are highly matched with user needs and business rules, improves the relevance of search results, and supports automatic parsing and response of multiple conditions and implicit needs (for example, price ranges, scenario preferences), enhancing the adaptability of complex queries.

[0074] In one embodiment, the user's operation information on the third search result may also be obtained, including clicks, browsing time, evaluation and other information.

[0075] Based on the operation information, we can obtain the user's personalized behavior information, which reflects the user's search habits and needs, and understand the user's satisfaction with the search results. For example, if the user clicks more and browses for longer time on video websites, it is believed that the user prefers video websites.

[0076] According to personalized behavior information, the user's corresponding sorting method and display method are personalized and optimized. For example, when the user searches again in the future, video websites can be recommended first. Of course, the personalized behavior information can also be used to iteratively optimize large models and deep learning models to improve the accuracy of search results and user experience. Optimization methods include adjusting model parameters, increasing training data, and improving model structure.

[0077] like Figure 3 As shown, the embodiment of the present application also proposes a search optimization device based on a large model, including: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the large model-based search optimization method as described in any of the above embodiments.

[0078] The present application also proposes a non-volatile computer storage medium in an embodiment, storing computer executable instructions, wherein the computer executable instructions are configured to be: the large model-based search optimization method described in any of the above embodiments.

[0079] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0080] The devices and media provided in the embodiments of the present application correspond one-to-one to the methods. Therefore, the devices and media also have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0081] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A search optimization method based on a large model, characterized in that: include: Receiving query information input by a user and preprocessing the query information; Based on a pre-trained deep learning model, semantic analysis is performed on the query information to extract key information contained therein; Taking the query information and the key information as input, generating a relevant first search result through a pre-trained large model; Acquire a business entity corresponding to the query information, determine compression structure information corresponding to the business entity, and optimize the first search result based on the compression structure information to obtain a second search result; The second search results are sorted and displayed.

2. The search optimization method based on a large model according to claim 1, characterized in that: Preprocessing the query information specifically includes: According to the query information, obtain a corresponding character string; Performing text preprocessing on the character string; The character string is segmented and part-of-speech tagging is performed on the segmentation result.

3. The large model-based search optimization method according to claim 1, characterized in that: Based on the pre-trained deep learning model, the query information is semantically analyzed to extract the key information contained therein, including: Determine, based on a pre-trained deep learning model, a minimum text unit contained in a word segmentation result of the query information; Performing position encoding on the minimum text unit; Outputting a corresponding context vector for the minimum text unit; the context vector includes a first vector of aggregated semantic representation for intent classification and a second vector of fine-grained semantic features for entity recognition; Intent classification is performed according to the first vector, and key information is extracted according to the intent classification and the second vector.

4. The large model-based search optimization method according to claim 1, characterized in that: Acquiring the business entity corresponding to the query information and determining the compression structure information corresponding to the business entity specifically includes: Generate a corresponding first business entity according to the key information; Based on the corresponding relationship in the local database, the key information is expanded to generate a corresponding second business entity for each first business entity; According to the corresponding relationship, each first business entity and its corresponding second business entity are compressed to generate corresponding compression structure information.

5. The large model-based search optimization method according to claim 4, characterized in that: Optimizing the first search result based on the compressed structure information to obtain a second search result specifically includes: For each compression structure information corresponding to the first business entity, select at least one second business entity as a selected entity corresponding to the compression structure information; According to the selected entities corresponding to all the compressed structure information, an entity combination is obtained; Performing a correlation analysis on the first search result according to the entity combination, and taking the first search result whose corresponding first correlation score is higher than a preset score as a sub-search result corresponding to the entity combination; For each sub-search result, several search results with the highest first relevance score are selected respectively, and the second search result is obtained by combining them after removing duplicates.

6. The large model-based search optimization method according to claim 5, characterized in that: The method further comprises: Determine the number of results of the second search result obtained by combining after deduplication; If the number of results is higher than a preset number range, then based on the number of corresponding second business entities in the entity combination, the corresponding sub-search results are deleted in order from low to high until the number of results meets the preset number range; If the number of results is lower than the preset number range, then based on the number of corresponding second business entities in the entity combination, corresponding sub-search results are added in descending order until the number of results meets the preset number range, or all sub-search results are added.

7. The large model-based search optimization method according to claim 1, characterized in that: Sorting and displaying the second search results specifically includes: Obtain access rights to the user; Filtering the second search result according to the access permission to obtain a third search result; Obtaining a second relevance score of the large model for the first search result, and obtaining a first ranking corresponding to the third search result according to the second relevance score; Obtaining access data of the third search result, and obtaining a corresponding quality score according to the access data, and adjusting the first ranking according to the quality score to obtain a second ranking; A corresponding display method is selected, and based on the display method, the third search results are sorted and displayed according to the second sorting.

8. The large model-based search optimization method according to claim 7, characterized in that: The method further comprises: Obtaining user operation information on the third search result; According to the operation information, obtaining the user's personalized behavior information; According to the personalized behavior information, the sorting method and display method corresponding to the user are personalized and optimized.

9. A search optimization device based on a large model, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the large model-based search optimization method as described in any one of claims 1 to 8.

10. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions are configured as: a large model-based search optimization method as described in any one of claims 1 to 8.

Images