A cross-modal retrieval method for data asset catalog tags based on vector library

Abstract

The present invention relates to the technical field of data asset retrieval, and discloses a cross-modal retrieval method for data asset catalog tags based on a vector library. By receiving data asset retrieval instructions transmitted by several user terminals during an initial retrieval period, extracting multimodal retrieval tags, performing vectorization conversion and tag vector matching, generating an initialization data asset retrieval result, and allocating the user terminals to corresponding distributed cache servers, the method optimizes and solves the cache strategy of the distributed cache server with a constraint set constructed by cache space, retrieval time and network communication status and a minimum cache communication volume as optimization targets, guides the user terminal to update cache content and retrieve data assets between the distributed cache server during the current target retrieval period, and updates the cache strategy for the next target retrieval period, thereby improving the data retrieval efficiency and retrieval scenario adaptability based on data tags, and reducing the impact of data communication pressure and reduced cache device life on retrieval.

Description

Technical Field

[0001] The present invention relates to the technical field of data asset retrieval, and in particular to a cross-modal retrieval method for data asset catalog tags based on a vector library. Background Art

[0002] Data tagging is a crucial concept in data management. In the era of big data, data tagging plays an increasingly important role. By tagging data, data retrieval, classification, and storage can be more easily performed. However, existing data management based on data tagging has the following limitations:

[0003] (1) In general data platforms, only full matching or fuzzy matching of data asset catalogs can be performed based on tag content, and data matching cannot be performed based on semantics. Alternatively, semantic retrieval can only be performed based on text tags, and tag retrieval based on images cannot be performed. This makes data retrieval based on data tags inefficient and limits the retrieval application scenarios.

[0004] (2) When performing data retrieval based on data tags (such as data asset retrieval), fast and efficient data retrieval is usually achieved by matching the retrieval tag vector after the retrieval tag is converted with the data asset tag vector of the data asset tag. In some scenarios, caching technology is applied to cache hot tags and corresponding tag vectors on the server side, which can significantly improve the overall retrieval efficiency and system performance. However, in actual applications, considering the cache space limitations of the cache server, it is necessary to reasonably select and update the cache content of the cache server to avoid data communication pressure caused by frequent updates of cache content and reduced cache device life, as well as cache space shortages caused by untimely updates of cache content that affect the overall retrieval efficiency of the system, which increases the difficulty of caching strategy planning during data asset retrieval.

[0005] Therefore, how to improve the efficiency of data retrieval based on data tags, avoid data communication pressure and reduced cache device life, and improve the adaptability of retrieval scenarios is a technical problem that needs to be solved urgently. Summary of the Invention

[0006] The present invention provides a cross-modal retrieval method for data asset catalog tags based on a vector library, aiming to solve at least one of the above technical problems.

[0007] To achieve the above objectives, the present invention provides a cross-modal retrieval method for data asset catalog tags based on a vector library, comprising the following steps:

[0008] S1: Creating a data asset directory on a search server; wherein the data asset directory includes a plurality of pieces of original data asset information including basic information of the data assets;

[0009] S2: Define several asset tags for the data asset catalog, vectorize each tag and store it in the vector database, bind corresponding asset tags to several pieces of basic data asset information in the data asset catalog, and generate actual data asset information containing basic data asset information and asset tags;

[0010] S3: Receive data asset retrieval instructions transmitted by several user terminals during the initial retrieval period, extract multimodal retrieval tags in the data asset retrieval instructions, perform vectorization conversion on the multimodal retrieval tags using a vectorization model and match them with the tag vectors to generate an initialization data asset retrieval result;

[0011] S4: Feedback the initialization data asset search results, multimodal search tags, and search tag vectors to the corresponding user terminals for storage. Based on each data asset search instruction of each user terminal during the initialization search period, the user terminals are classified for search preferences and search requirements, and a distributed cache server is allocated to each user terminal with a search preference category.

[0012] S5: Generate a cache strategy for the distributed cache server based on the cache space of each distributed cache server, the network communication status, and the search requirements of each user terminal;

[0013] S6: Receive data asset retrieval instructions transmitted by multiple user terminals during the current target retrieval period, extract the multimodal retrieval tag in the data asset retrieval instruction, perform a two-level retrieval on the distributed cache server and the retrieval server based on the caching strategy of the distributed cache server, and generate a target data asset retrieval result;

[0014] S7: Before the next target retrieval period, the cache strategy of the distributed cache server is updated according to the target data asset retrieval result.

[0015] Optionally, step S1: creating a data asset directory on the search server, specifically including:

[0016] S11: The search server obtains several pieces of data asset entry information and extracts the asset name, asset description information and several asset data items from the data asset entry information;

[0017] S12: The asset name, asset description information and several asset data items in each data asset entry information are used as data asset basic information to generate several pieces of original data asset information, and a data asset directory is constructed based on the original data asset information.

[0018] Optionally, in step S2: defining several asset tags for the data asset catalog, vectorizing each tag and storing it in a vector database, specifically includes:

[0019] S21: defining a plurality of asset tags for the data asset catalog; wherein the asset tags include text-type asset tags, image-type asset tags, and voice-type asset tags;

[0020] S22: Use the speech-to-text model to convert speech-type asset tags into text-type asset tags. Use the word2text model for text-type asset tags and the clip model for image-type asset tags. Vectorize all types of asset tags and store the obtained tag vectors in the vector database.

[0021] Optionally, in step S2: binding corresponding asset tags to several pieces of basic data asset information in the data asset catalog to generate actual data asset information containing the basic data asset information and the asset tags, specifically including:

[0022] S23: Extract asset description information from each piece of data asset basic information in the data asset target, use natural language processing tools to extract several keywords from the asset description information, solve the vector of each keyword, and generate a keyword vector for each piece of data asset basic information;

[0023] S24: Using several keyword vectors of each piece of data asset basic information and the label vector corresponding to each asset label, calculate the correlation between the data asset basic information and the asset label, bind each piece of data asset basic information to the asset label with the highest correlation, and generate actual data asset information containing the data asset basic information and the asset label.

[0024] Optionally, step S3: receiving data asset retrieval instructions transmitted by several user terminals during the initial retrieval period, extracting multimodal retrieval tags in the data asset retrieval instructions, performing vectorization conversion on the multimodal retrieval tags using a vectorization model and matching the tags with the tag vectors, and generating an initialization data asset retrieval result, specifically including:

[0025] S31: receiving data asset retrieval instructions transmitted by several user terminals during an initial retrieval period, extracting multimodal retrieval tags from the data asset retrieval instructions, and performing vectorization conversion on the multimodal retrieval tags of different types using corresponding vectorization models;

[0026] S32: Based on the retrieval tag vector obtained from the conversion result and the asset tag vector in the vector database, calculate the similarity between the multimodal retrieval tag in the data asset retrieval instruction and the asset tag, take the top N asset tags in similarity ranking as candidate asset tags, and generate the initialized data asset retrieval result.

[0027] Optionally, step S4: feeding back the initialization data asset retrieval results, multimodal retrieval tags, and retrieval tag vectors to the corresponding user terminals for storage, performing retrieval preference classification and retrieval demand generation for the user terminals based on each data asset retrieval instruction of each user terminal during the initialization retrieval period, and allocating a distributed cache server to each user terminal of a retrieval preference category, specifically including:

[0028] S41: Feedback the initialization data asset retrieval result, multimodal retrieval tag, and retrieval tag vector to the corresponding user terminal, and store them in the user terminal;

[0029] S42: Based on the candidate asset tag set and instruction reception time of each data asset retrieval instruction of each user terminal in the initial retrieval period, the user terminal is classified into retrieval preference categories and retrieval requirements are generated, and a distributed cache server is allocated to each user terminal of the retrieval preference category.

[0030] Optionally, in step S42: based on the candidate asset tag set and instruction reception time of each data asset retrieval instruction of each user terminal during the initial retrieval period, retrieval preference classification and retrieval demand generation are performed on the user terminal, specifically including:

[0031] S421: extracting a plurality of candidate asset tags from a candidate asset tag set of each data asset search instruction of each user terminal in an initial search period, and classifying the search preferences of the plurality of user terminals based on the plurality of candidate asset tags;

[0032] S422: Query the sending time of each data asset retrieval instruction of each user terminal in the initial retrieval period, convert the sending time into a sending time feature to construct a training sample, train the initially constructed convolutional neural network, and use the retrieval time prediction model obtained through training to predict the retrieval time of each user terminal in each target retrieval period and use it as the retrieval demand.

[0033] Optionally, step S5: generating a cache strategy for the distributed cache server based on the cache space of each distributed cache server, the network communication status, and the search requirements of each user terminal, specifically including:

[0034] S51: Obtaining the cache space of each distributed cache server, the network communication status between each distributed cache server and each user terminal in each minimum retrieval cycle during the target retrieval period, and the retrieval time of each user terminal during the target retrieval period;

[0035] S52: Considering the multimodal retrieval tag and retrieval tag vector received by each user terminal, a first constraint condition is that the sum of the data volume of the multimodal retrieval tag and retrieval tag vector of each user terminal cached by each distributed cache server in each minimum retrieval period of the target retrieval period is less than the cache space of the distributed cache server; a second constraint condition is that the retrieval time of each user terminal in the target retrieval period falls within the cache period determined by the first time when the distributed cache server caches the multimodal retrieval tag and retrieval tag vector of the user terminal and the second time when the distributed cache server clears the cache of the multimodal retrieval tag and retrieval tag vector of each user terminal; a third constraint condition is that the network communication status at the first time when the distributed cache server caches the multimodal retrieval tag and retrieval tag vector of each user terminal satisfies a preset cache data communication parameter threshold; an optimization goal is that the sum of the communication volume of cache data communication performed by all distributed cache servers in the target retrieval period is minimized, and the multimodal retrieval tag and retrieval tag vector of the corresponding user terminal cached by each distributed cache server in each minimum retrieval period of the target retrieval period are optimized;

[0036] A cache strategy of the distributed cache server is generated based on the multimodal retrieval tags and retrieval tag vectors of the corresponding user terminals cached by each distributed cache server in each minimum retrieval period during the target retrieval period.

[0037] Optionally, step S6: receiving data asset retrieval instructions transmitted by several user terminals during the current target retrieval period, extracting multimodal retrieval tags in the data asset retrieval instructions, and performing a two-level retrieval of the distributed cache server and the retrieval server based on the caching strategy of the distributed cache server to generate target data asset retrieval results, specifically including:

[0038] S61: receiving data asset retrieval instructions transmitted by several user terminals during the current target retrieval period, and extracting multimodal retrieval tags in the data asset retrieval instructions;

[0039] S62: Controlling each distributed cache server to perform cache update and cache clearing of the multimodal retrieval tags and retrieval tag vectors corresponding to the user terminal in each minimum retrieval period of the target retrieval period according to the caching strategy;

[0040] S63: According to the search requirements of each user terminal, control the search time of each user terminal in the target search period, use the multimodal search tag in the data asset search instruction, perform two-level search on the distributed cache server and the search server, and generate the target data asset search results.

[0041] Optionally, step S7: before the next target search period, updating the cache strategy of the distributed cache server according to the target data asset search result, specifically including:

[0042] S71: Before the next target retrieval period, determine, based on the target data asset retrieval results, whether each distributed cache server meets a cache server update condition; wherein the cache server update condition is configured as a ratio of the number of distributed cache server retrievals executed by the distributed cache server to the number of retrievals executed by the retrieval server during the target retrieval period being less than a target ratio;

[0043] S72: If yes, drive the distributed cache server to receive the multimodal search tags and search tag vectors received by all user terminals under the corresponding search preference category in the current target search period, and regenerate the cache strategy of the distributed cache server for the next target search period.

[0044] The beneficial effects of the present invention are: a cross-modal retrieval method for data asset catalog tags based on a vector library is proposed, by receiving data asset retrieval instructions transmitted by several user terminals during the initial retrieval period, extracting multimodal retrieval tags to perform vectorization conversion and tag vector matching, generating initialization data asset retrieval results, and allocating user terminals to corresponding distributed cache servers, with the constraint condition set constructed by cache space, retrieval time and network communication status and the minimum cache communication volume as optimization goals, optimizing and solving the cache strategy of the distributed cache server, guiding the user terminal to update the cache content and retrieve data assets between the current target retrieval period and the distributed cache server, and updating the cache strategy for the next target retrieval period, thereby improving the data retrieval efficiency and retrieval scenario adaptability based on data tags, reducing the impact of data communication pressure and reduced cache device life on retrieval. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Schematic diagram of the process of a cross-modal retrieval method for data asset catalog tags based on a vector library according to an embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0047] The embodiment of the present invention provides a cross-modal retrieval method for data asset catalog tags based on a vector library, referring to Figure 1 , Figure 1Schematic diagram of the process of a cross-modal retrieval method for data asset catalog tags based on a vector library according to an embodiment of the present invention.

[0048] In this embodiment, a cross-modal retrieval method for data asset catalog tags based on a vector library includes the following steps:

[0049] S1: Creating a data asset directory on a search server; wherein the data asset directory includes a plurality of pieces of original data asset information including basic information of the data assets;

[0050] S2: Define several asset tags for the data asset catalog, vectorize each tag and store it in the vector database, bind corresponding asset tags to several pieces of basic data asset information in the data asset catalog, and generate actual data asset information containing basic data asset information and asset tags;

[0051] S3: Receive data asset retrieval instructions transmitted by several user terminals during the initial retrieval period, extract multimodal retrieval tags in the data asset retrieval instructions, perform vectorization conversion on the multimodal retrieval tags using a vectorization model and match them with the tag vectors to generate an initialization data asset retrieval result;

[0052] S4: Feedback the initialization data asset search results, multimodal search tags, and search tag vectors to the corresponding user terminals for storage. Based on each data asset search instruction of each user terminal during the initialization search period, the user terminals are classified for search preferences and search requirements, and a distributed cache server is allocated to each user terminal with a search preference category.

[0053] S5: Generate a cache strategy for the distributed cache server based on the cache space of each distributed cache server, the network communication status, and the search requirements of each user terminal;

[0054] S6: Receive data asset retrieval instructions transmitted by multiple user terminals during the current target retrieval period, extract the multimodal retrieval tag in the data asset retrieval instruction, perform a two-level retrieval on the distributed cache server and the retrieval server based on the caching strategy of the distributed cache server, and generate a target data asset retrieval result;

[0055] S7: Before the next target retrieval period, the cache strategy of the distributed cache server is updated according to the target data asset retrieval result.

[0056] It should be noted that the existing data management based on data tags has the following limitations: (1) In general data platforms, only full matching or fuzzy matching of data asset catalogs can be performed based on tag content, and data matching cannot be performed based on semantics, or semantic retrieval can only be performed based on text tags, and tag retrieval cannot be performed based on images, which makes data retrieval based on data tags inefficient and limits the retrieval application scenarios. (2) When performing data retrieval based on data tags (such as data asset retrieval), fast and efficient data retrieval is usually achieved based on the matching of the retrieval tag vector after the retrieval tag conversion with the data asset tag vector. In some scenarios, caching technology is applied to cache hot tags and corresponding tag vectors on the server side, which can significantly improve the overall retrieval efficiency and system performance. However, in actual applications, considering the cache space limitations of the cache server, it is necessary to reasonably select and update the cache content of the cache server to avoid the data communication pressure caused by frequent cache content updates and the reduction of cache device life, as well as the cache space shortage caused by untimely cache content updates that affect the overall retrieval efficiency of the system, which increases the difficulty of caching strategy planning when retrieving data assets.

[0057] In order to solve the above problems, this embodiment receives data asset retrieval instructions transmitted by several user terminals during the initial retrieval period, extracts multimodal retrieval tags, performs vectorized conversion and tag vector matching, generates initialization data asset retrieval results, and assigns user terminals to corresponding distributed cache servers. The constraint condition set constructed by cache space, retrieval time and network communication status and the minimum cache communication volume are used as optimization goals to optimize and solve the cache strategy of the distributed cache server, guide the user terminal to update the cache content and retrieve data assets between the current target retrieval period and the distributed cache server, and update the cache strategy for the next target retrieval period. In this way, the data retrieval efficiency and retrieval scenario adaptability based on data tags are improved, and the impact of data communication pressure and reduced cache device life on retrieval is reduced.

[0058] In a preferred embodiment, step S1: creating a data asset directory on a search server, specifically includes:

[0059] S11: The search server obtains several pieces of data asset entry information and extracts the asset name, asset description information and several asset data items from the data asset entry information;

[0060] S12: The asset name, asset description information and several asset data items in each data asset entry information are used as data asset basic information to generate several pieces of original data asset information, and a data asset directory is constructed based on the original data asset information.

[0061] On this basis, in step S2: define several asset tags for the data asset catalog, vectorize each tag and store it in the vector database, specifically including:

[0062] S21: defining a plurality of asset tags for the data asset catalog; wherein the asset tags include text-type asset tags, image-type asset tags, and voice-type asset tags;

[0063] S22: Use the speech-to-text model to convert speech-type asset tags into text-type asset tags. Use the word2text model for text-type asset tags and the clip model for image-type asset tags. Vectorize all types of asset tags and store the obtained tag vectors in the vector database.

[0064] Furthermore, in step S2: corresponding asset tags are bound to several pieces of basic data asset information in the data asset catalog to generate actual data asset information containing the basic data asset information and the asset tags, specifically including:

[0065] S23: Extract asset description information from each piece of data asset basic information in the data asset target, use natural language processing tools to extract several keywords from the asset description information, solve the vector of each keyword, and generate a keyword vector for each piece of data asset basic information;

[0066] S24: Using several keyword vectors of each piece of data asset basic information and the label vector corresponding to each asset label, calculate the correlation between the data asset basic information and the asset label, bind each piece of data asset basic information to the asset label with the highest correlation, and generate actual data asset information containing the data asset basic information and the asset label.

[0067] In this embodiment, a data asset catalog is constructed by first acquiring several pieces of data asset input information and generating several pieces of raw data asset information containing basic data asset information. Subsequently, several asset tags are defined for the data asset catalog. Keyword vector association calculations are then used to bind corresponding asset tags to each piece of basic data asset information. Ultimately, actual data asset information is generated, containing both basic data asset information and asset tags, facilitating subsequent tag-based data asset retrieval.

[0068] In a preferred embodiment, step S3: receiving data asset retrieval instructions transmitted by several user terminals during the initial retrieval period, extracting multimodal retrieval tags in the data asset retrieval instructions, performing vectorization conversion on the multimodal retrieval tags using a vectorization model and matching the tags with the tag vectors, and generating the initialization data asset retrieval results, specifically includes:

[0069] S31: receiving data asset retrieval instructions transmitted by several user terminals during an initial retrieval period, extracting multimodal retrieval tags from the data asset retrieval instructions, and performing vectorization conversion on the multimodal retrieval tags of different types using corresponding vectorization models;

[0070] S32: Based on the retrieval tag vector obtained from the conversion result and the asset tag vector in the vector database, calculate the similarity between the multimodal retrieval tag in the data asset retrieval instruction and the asset tag, take the top N asset tags in similarity ranking as candidate asset tags, and generate the initialized data asset retrieval result.

[0071] On this basis, step S4: feeding back the initial data asset retrieval results, multimodal retrieval tags, and retrieval tag vectors to the corresponding user terminal for storage, classifying the user terminal's retrieval preferences and generating retrieval requirements based on each data asset retrieval instruction of each user terminal during the initial retrieval period, and allocating a distributed cache server to each user terminal of each retrieval preference category, specifically including:

[0072] S41: Feedback the initialization data asset retrieval result, multimodal retrieval tag, and retrieval tag vector to the corresponding user terminal, and store them in the user terminal;

[0073] S42: Based on the candidate asset tag set and instruction reception time of each data asset retrieval instruction of each user terminal in the initial retrieval period, the user terminal is classified into retrieval preference categories and retrieval requirements are generated, and a distributed cache server is allocated to each user terminal of the retrieval preference category.

[0074] Furthermore, in step S42, based on the candidate asset tag set and instruction reception time of each data asset retrieval instruction of each user terminal during the initial retrieval period, the user terminal is classified in terms of retrieval preference and retrieval requirements, specifically including:

[0075] S421: extracting a plurality of candidate asset tags from a candidate asset tag set of each data asset search instruction of each user terminal in an initial search period, and classifying the search preferences of the plurality of user terminals based on the plurality of candidate asset tags;

[0076] S422: Query the sending time of each data asset retrieval instruction of each user terminal in the initial retrieval period, convert the sending time into a sending time feature to construct a training sample, train the initially constructed convolutional neural network, and use the retrieval time prediction model obtained through training to predict the retrieval time of each user terminal in each target retrieval period and use it as the retrieval demand.

[0077] In this embodiment, during the initialization retrieval period, the retrieval server performs an initialization data asset retrieval for a multimodal retrieval tag vector and an asset tag vector based on the data asset retrieval instructions transmitted by several user terminals. After that, considering the initialization data asset retrieval results, the user terminals are classified for retrieval preferences and retrieval requirements are generated, and a distributed cache server is allocated to each user terminal of each retrieval preference category, so that each user terminal only executes the caching of multimodal retrieval tags and retrieval tag vectors in the corresponding distributed cache server, which can significantly improve the data asset retrieval speed under the same retrieval preference category.

[0078] In a preferred embodiment, step S5: generating a cache strategy for the distributed cache server based on the cache space of each distributed cache server, the network communication status, and the search requirements of each user terminal, specifically includes:

[0079] S51: Obtaining the cache space of each distributed cache server, the network communication status between each distributed cache server and each user terminal in each minimum retrieval cycle during the target retrieval period, and the retrieval time of each user terminal during the target retrieval period;

[0080] S52: Considering the multimodal retrieval tag and retrieval tag vector received by each user terminal, a first constraint condition is that the sum of the data volume of the multimodal retrieval tag and retrieval tag vector of each user terminal cached by each distributed cache server in each minimum retrieval period of the target retrieval period is less than the cache space of the distributed cache server; a second constraint condition is that the retrieval time of each user terminal in the target retrieval period falls within the cache period determined by the first time when the distributed cache server caches the multimodal retrieval tag and retrieval tag vector of the user terminal and the second time when the distributed cache server clears the cache of the multimodal retrieval tag and retrieval tag vector of each user terminal; a third constraint condition is that the network communication status at the first time when the distributed cache server caches the multimodal retrieval tag and retrieval tag vector of each user terminal satisfies a preset cache data communication parameter threshold; an optimization goal is that the sum of the communication volume of cache data communication performed by all distributed cache servers in the target retrieval period is minimized, and the multimodal retrieval tag and retrieval tag vector of the corresponding user terminal cached by each distributed cache server in each minimum retrieval period of the target retrieval period are optimized;

[0081] S53: Generate a cache strategy for the distributed cache server based on the multimodal retrieval tags and retrieval tag vectors of the corresponding user terminals cached by each distributed cache server in each minimum retrieval period during the target retrieval period.

[0082] Furthermore, step S6: receiving data asset retrieval instructions transmitted by several user terminals during the current target retrieval period, extracting the multimodal retrieval tag in the data asset retrieval instruction, performing a two-level retrieval of the distributed cache server and the retrieval server based on the caching strategy of the distributed cache server, and generating the target data asset retrieval result, specifically includes:

[0083] S61: receiving data asset retrieval instructions transmitted by several user terminals during the current target retrieval period, and extracting multimodal retrieval tags in the data asset retrieval instructions;

[0084] S62: Controlling each distributed cache server to perform cache update and cache clearing of the multimodal retrieval tags and retrieval tag vectors corresponding to the user terminal in each minimum retrieval period of the target retrieval period according to the caching strategy;

[0085] S63: According to the search requirements of each user terminal, control the search time of each user terminal in the target search period, use the multimodal search tag in the data asset search instruction, perform two-level search on the distributed cache server and the search server, and generate the target data asset search results.

[0086] In this embodiment, several user terminals are assigned to corresponding distributed cache servers. According to the cache space, network communication status of each distributed cache server and the retrieval requirements of each user terminal, the cache strategy of the distributed cache server is optimized and solved with the constraint condition set constructed by the cache space, retrieval time and network communication status and the minimum cache communication volume as the optimization target. The data asset retrieval instructions transmitted by several user terminals in the current target retrieval period are received, the multimodal retrieval tags in the data asset retrieval instructions are extracted, and based on the caching strategy of the distributed cache server, a two-level search of the distributed cache server and the retrieval server is performed to generate the target data asset retrieval result.

[0087] In a preferred embodiment, step S7: before the next target search period, updating the cache strategy of the distributed cache server according to the target data asset search result, specifically includes:

[0088] S71: Before the next target retrieval period, determine, based on the target data asset retrieval results, whether each distributed cache server meets a cache server update condition; wherein the cache server update condition is configured as a ratio of the number of distributed cache server retrievals executed by the distributed cache server to the number of retrievals executed by the retrieval server during the target retrieval period being less than a target ratio;

[0089] S72: If yes, drive the distributed cache server to receive the multimodal search tags and search tag vectors received by all user terminals under the corresponding search preference category in the current target search period, and regenerate the cache strategy of the distributed cache server for the next target search period.

[0090] In this embodiment, the cache strategy of the distributed cache server for the next target retrieval period is updated according to the target data asset retrieval results. Thus, by constructing a cross-modal retrieval based on the vector library and the data asset directory label of the cache server, the distributed cache server provides separate cache space for user terminals with different retrieval preference classifications. Then, considering the retrieval requirements of the user terminals and the communication status and cache space of the distributed cache server, the most reasonable multimodal retrieval label and the data caching and cache clearing strategy of the retrieval label vector are set, which can improve the data retrieval efficiency and retrieval scenario adaptability based on data labels and avoid the impact of data communication pressure and reduced cache device life on retrieval.

[0091] It should be understood that, in the description of this specification, reference to terms such as "one embodiment," "another embodiment," "other embodiments," or "first to Nth embodiments" means that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any appropriate manner in any one or more embodiments or examples.

[0092] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or system. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or system comprising the element.

[0093] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A cross-modal retrieval method for data asset catalog tags based on a vector library, characterized in that: The following steps are involved: S1: Creating a data asset directory on a search server; wherein the data asset directory includes a plurality of pieces of original data asset information including basic information of the data assets; S2: Define several asset tags for the data asset catalog, vectorize each tag and store it in the vector database, bind corresponding asset tags to several pieces of basic data asset information in the data asset catalog, and generate actual data asset information containing basic data asset information and asset tags; S3: Receive data asset retrieval instructions transmitted by several user terminals during the initial retrieval period, extract multimodal retrieval tags in the data asset retrieval instructions, perform vectorization conversion on the multimodal retrieval tags using a vectorization model and match them with the tag vectors to generate an initialization data asset retrieval result; S4: Feedback the initialization data asset search results, multimodal search tags, and search tag vectors to the corresponding user terminals for storage. Based on each data asset search instruction of each user terminal during the initialization search period, the user terminals are classified for search preferences and search requirements, and a distributed cache server is allocated to each user terminal with a search preference category. S5: Generate a cache strategy for the distributed cache server based on the cache space of each distributed cache server, the network communication status, and the search requirements of each user terminal; S6: Receive data asset retrieval instructions transmitted by multiple user terminals during the current target retrieval period, extract the multimodal retrieval tag in the data asset retrieval instruction, perform a two-level retrieval on the distributed cache server and the retrieval server based on the caching strategy of the distributed cache server, and generate a target data asset retrieval result; S7: Before the next target retrieval period, the cache strategy of the distributed cache server is updated according to the target data asset retrieval result.

2. The cross-modal retrieval method for data asset catalog tags based on a vector library according to claim 1 is characterized in that: Step S1: Create a data asset directory on the search server, specifically including: S11: The search server obtains several pieces of data asset entry information and extracts the asset name, asset description information and several asset data items from the data asset entry information; S12: The asset name, asset description information and several asset data items in each data asset entry information are used as data asset basic information to generate several pieces of original data asset information, and a data asset directory is constructed based on the original data asset information.

3. The cross-modal retrieval method for data asset catalog tags based on a vector library according to claim 1, characterized in that: In step S2: several asset tags of the data asset catalog are defined, and each tag is vectorized and stored in the vector database, specifically including: S21: defining a plurality of asset tags for the data asset catalog; wherein the asset tags include text-type asset tags, image-type asset tags, and voice-type asset tags; S22: Use the speech-to-text model to convert speech-type asset tags into text-type asset tags. Use the word2text model for text-type asset tags and the clip model for image-type asset tags. Vectorize all types of asset tags and store the obtained tag vectors in the vector database.

4. The cross-modal retrieval method for data asset catalog tags based on a vector library according to claim 3 is characterized in that: In step S2, corresponding asset tags are bound to several pieces of basic data asset information in the data asset catalog to generate actual data asset information containing the basic data asset information and asset tags, specifically including: S23: Extract asset description information from each piece of data asset basic information in the data asset target, use natural language processing tools to extract several keywords from the asset description information, solve the vector of each keyword, and generate a keyword vector for each piece of data asset basic information; S24: Using several keyword vectors of each piece of data asset basic information and the label vector corresponding to each asset label, calculate the correlation between the data asset basic information and the asset label, bind each piece of data asset basic information to the asset label with the highest correlation, and generate actual data asset information containing the data asset basic information and the asset label.

5. The cross-modal retrieval method for data asset catalog tags based on a vector library according to claim 1, characterized in that: Step S3: Receive data asset retrieval instructions transmitted by several user terminals during the initial retrieval period, extract multimodal retrieval tags in the data asset retrieval instructions, use the vectorization model to vectorize the multimodal retrieval tags and match them with the tag vectors, and generate the initialization data asset retrieval results, specifically including: S31: receiving data asset retrieval instructions transmitted by several user terminals during an initial retrieval period, extracting multimodal retrieval tags from the data asset retrieval instructions, and performing vectorization conversion on the multimodal retrieval tags of different types using corresponding vectorization models; S32: Based on the retrieval tag vector obtained from the conversion result and the asset tag vector in the vector database, calculate the similarity between the multimodal retrieval tag in the data asset retrieval instruction and the asset tag, take the top N asset tags in similarity ranking as candidate asset tags, and generate the initialized data asset retrieval result.

6. The cross-modal retrieval method for data asset catalog tags based on a vector library according to claim 1, characterized in that: Step S4: Feedback the initial data asset retrieval results, multimodal retrieval tags, and retrieval tag vectors to the corresponding user terminals for storage. Based on each data asset retrieval instruction of each user terminal during the initial retrieval period, the user terminals are classified for retrieval preference and retrieval requirements are generated. A distributed cache server is allocated to each user terminal of each retrieval preference category. Specifically, the following steps are performed: S41: Feedback the initialization data asset retrieval result, multimodal retrieval tag, and retrieval tag vector to the corresponding user terminal, and store them in the user terminal; S42: Based on the candidate asset tag set and instruction reception time of each data asset retrieval instruction of each user terminal in the initial retrieval period, the user terminal is classified into retrieval preference categories and retrieval requirements are generated, and a distributed cache server is allocated to each user terminal of the retrieval preference category.

7. The cross-modal retrieval method for data asset catalog tags based on a vector library according to claim 6 is characterized in that: In step S42, based on the candidate asset tag set and instruction reception time of each data asset retrieval instruction of each user terminal during the initial retrieval period, the user terminal is classified in terms of retrieval preference and retrieval requirements, specifically including: S421: extracting a plurality of candidate asset tags from a candidate asset tag set of each data asset search instruction of each user terminal in an initial search period, and classifying the search preferences of the plurality of user terminals based on the plurality of candidate asset tags; S422: Query the sending time of each data asset retrieval instruction of each user terminal in the initial retrieval period, convert the sending time into a sending time feature to construct a training sample, train the initially constructed convolutional neural network, and use the retrieval time prediction model obtained through training to predict the retrieval time of each user terminal in each target retrieval period and use it as the retrieval demand.

8. The cross-modal retrieval method for data asset catalog tags based on a vector library according to claim 1, characterized in that: Step S5: Generate a cache strategy for the distributed cache server based on the cache space of each distributed cache server, the network communication status, and the search requirements of each user terminal, specifically including: S51: Obtaining the cache space of each distributed cache server, the network communication status between each distributed cache server and each user terminal in each minimum retrieval cycle during the target retrieval period, and the retrieval time of each user terminal during the target retrieval period; S52: Considering the multimodal retrieval tag and retrieval tag vector received by each user terminal, a first constraint condition is that the sum of the data volume of the multimodal retrieval tag and retrieval tag vector of each user terminal cached by each distributed cache server in each minimum retrieval period of the target retrieval period is less than the cache space of the distributed cache server; a second constraint condition is that the retrieval time of each user terminal in the target retrieval period falls within the cache period determined by the first time when the distributed cache server caches the multimodal retrieval tag and retrieval tag vector of the user terminal and the second time when the distributed cache server clears the cache of the multimodal retrieval tag and retrieval tag vector of each user terminal; a third constraint condition is that the network communication status at the first time when the distributed cache server caches the multimodal retrieval tag and retrieval tag vector of each user terminal satisfies a preset cache data communication parameter threshold; an optimization goal is that the sum of the communication volume of cache data communication performed by all distributed cache servers in the target retrieval period is minimized, and the multimodal retrieval tag and retrieval tag vector of the corresponding user terminal cached by each distributed cache server in each minimum retrieval period of the target retrieval period are optimized; A cache strategy of the distributed cache server is generated based on the multimodal retrieval tags and retrieval tag vectors of the corresponding user terminals cached by each distributed cache server in each minimum retrieval period during the target retrieval period.

9. The cross-modal retrieval method for data asset catalog tags based on a vector library according to claim 1, characterized in that: Step S6: Receive data asset retrieval instructions transmitted by several user terminals during the current target retrieval period, extract the multimodal retrieval tag in the data asset retrieval instruction, perform a two-level retrieval of the distributed cache server and the retrieval server based on the caching strategy of the distributed cache server, and generate the target data asset retrieval results, specifically including: S61: receiving data asset retrieval instructions transmitted by several user terminals during the current target retrieval period, and extracting multimodal retrieval tags in the data asset retrieval instructions; S62: Controlling each distributed cache server to perform cache update and cache clearing of the multimodal retrieval tags and retrieval tag vectors corresponding to the user terminal in each minimum retrieval period of the target retrieval period according to the caching strategy; S63: According to the search requirements of each user terminal, control the search time of each user terminal in the target search period, use the multimodal search tag in the data asset search instruction, perform two-level search on the distributed cache server and the search server, and generate the target data asset search results.

10. The cross-modal retrieval method for data asset catalog tags based on a vector library according to claim 1, characterized in that: Step S7: Before the next target search period, the cache strategy of the distributed cache server is updated according to the target data asset search result, specifically including: S71: Before the next target retrieval period, determine, based on the target data asset retrieval results, whether each distributed cache server meets a cache server update condition; wherein the cache server update condition is configured as a ratio of the number of distributed cache server retrievals executed by the distributed cache server to the number of retrievals executed by the retrieval server during the target retrieval period being less than a target ratio; S72: If yes, drive the distributed cache server to receive the multimodal search tags and search tag vectors received by all user terminals under the corresponding search preference category in the current target search period, and regenerate the cache strategy of the distributed cache server for the next target search period.

Images