Search method, device, electronic device, and storage medium

Abstract

To provide a retrieval method, a device, an electronic apparatus, a computer readable storage medium, and a computer program product.SOLUTION: The present disclosure provides a retrieval method, a device, an electronic apparatus, and a storage medium, and relates to an artificial intelligence field, especially to a smart retrieval field. The retrieval method includes: converting a retrieval request of a user into a first request semantic vector; retrieving at least one first data semantic vector matching the first request semantic vector by a retrieval resource database, in which the retrieval resource database is constructed as a semantic vector space, different types of data are converted into the corresponding data semantic vector in the semantic vector space, and the different types of data include at least a text, a picture and a video; and generating a retrieval result on the basis of at least the one first data semantic vector.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to the technical field of artificial intelligence, in particular to the technical field of smart search, and in particular to a search method, an apparatus, an electronic device, a computer-readable storage medium and a computer program product. [Background technology]

[0002] Artificial intelligence is a discipline that studies how computers can simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, planning, etc.), and includes both hardware and software level technologies. Hardware technologies in artificial intelligence generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing, etc. Software technologies in artificial intelligence mainly include computer vision technology, voice recognition technology, natural language processing technology, machine learning / deep learning, big data processing technology, knowledge graph technology, etc.

[0003] Data search is one of the basic services of the Internet, and can provide search results that meet the needs of users based on their search requests.

[0004] The approaches described in this section are not necessarily approaches that have been previously conceived or adopted. Unless otherwise expressly stated, it should not be assumed that any approach described in this section qualifies as prior art by virtue of its inclusion in this section. Similarly, unless otherwise expressly stated, it should not be assumed that any problems addressed in this section are admitted into the prior art. Summary of the Invention

[0005] The present disclosure provides a search method, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product.

[0006] According to one aspect of the present disclosure, a search method is provided, the method including: converting a user's search request into a first request semantic vector; searching a search resource database for at least one first data semantic vector matching the first request semantic vector, the search resource database being constructed as a semantic vector space, in which different types of data are converted into corresponding data semantic vectors, the different types of data including at least text, picture, and video; and generating search results based on the at least one first data semantic vector.

[0007] According to another aspect of the present disclosure, a data search device is provided, the device including: a unit for converting a user's search request into a first request semantic vector; a unit for searching a search resource database for at least one first data semantic vector matching the first request semantic vector, the search resource database being constructed as a semantic vector space, in which different types of data are converted into corresponding data semantic vectors, the different types of data including at least text, picture and video; and a unit for generating search results based on the at least one first data semantic vector.

[0008] According to another aspect of the present disclosure, there is provided a search system based on a pre-trained language model, comprising: a recall model configured to convert a user's search request into a first request semantic vector and search for at least one first data semantic vector matching the first request semantic vector in a search resource database, the search resource database being constructed as a semantic vector space, in which different types of data are converted into corresponding data semantic vectors, the different types of data at least including text, picture and video; and a sorting model configured to sort the one or more first data by sequentially determining a similarity between the search request and one or more first data corresponding to the at least one first data semantic vector, wherein the recall model and the sorting model are cascaded to form an end-to-end deep neural network basic model.

[0009] According to another aspect of the present disclosure, there is provided an electronic device including at least one processor and a memory communicatively connected to the at least one processor, the memory storing instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the above-mentioned search method or a method for training a search system based on a pre-trained language model.

[0010] According to another aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium having stored thereon computer instructions for causing a computer to execute the above-described search method or the method for training a search system based on a pre-trained language model.

[0011] According to another aspect of the present disclosure, there is provided a computer program product comprising a computer program which, when executed by a processor, implements the above-described search method or the method for training a search system based on a pre-trained language model.

[0012] According to one or more embodiments of the present disclosure, search performance can be improved.

[0013] It should be understood that the contents described in this section are not intended to identify key or important features of the embodiments of the present disclosure, and are not intended to limit the scope of the present disclosure. Other features of the present disclosure will become more readily understood from the following description. [Brief description of the drawings]

[0014] The drawings illustratively depict the examples, constitute a part of the specification, and together with the description serve to explain exemplary embodiments of the examples. The examples shown are used for illustrative purposes only and are not intended to limit the scope of the claims. In all figures, the same drawing reference numbers refer to similar, but not necessarily identical, elements.

[0015] [Figure 1] 2 shows a flowchart of a search method according to an exemplary embodiment of the present disclosure. [Diagram 2] 2 shows a flowchart of a search method according to an exemplary embodiment of the present disclosure. [Diagram 3] 2 shows a flowchart of a search method according to an exemplary embodiment of the present disclosure. [Figure 4] 2 shows a flowchart of a search method according to an exemplary embodiment of the present disclosure. [Diagram 5] 2 shows a flowchart of a search method according to an exemplary embodiment of the present disclosure. [Figure 6] 1 illustrates a schematic diagram of structured analysis of a search request according to an exemplary embodiment of the present disclosure. [Figure 7] FIG. 2 shows a structural block diagram of a data searching device according to an exemplary embodiment of the present disclosure; [Figure 8] FIG. 2 shows a structural block diagram of a data searching device according to an exemplary embodiment of the present disclosure; [Figure 9] FIG. 2 shows a structural block diagram of a data searching device according to an exemplary embodiment of the present disclosure; [Figure 10]FIG. 2 shows a structural block diagram of a data searching device according to an exemplary embodiment of the present disclosure; [Figure 11] FIG. 2 shows a structural block diagram of a data searching device according to an exemplary embodiment of the present disclosure; [Figure 12] FIG. 1 illustrates a structural block diagram of a search system based on a pre-trained language model according to an exemplary embodiment of the present disclosure. [Figure 13] FIG. 1 illustrates a structural block diagram of a search system based on a pre-trained language model according to an exemplary embodiment of the present disclosure. [Figure 14] FIG. 1 illustrates a structural block diagram of a search system based on a pre-trained language model according to an exemplary embodiment of the present disclosure. [Figure 15] 1 shows a schematic diagram of a model structure based on a cross-encoder structure and a schematic diagram of a model structure based on a double-encoder structure according to an exemplary embodiment of the present disclosure; [Figure 16] FIG. 2 illustrates a schematic diagram of the operation process of a search system based on a pre-trained language model according to an exemplary embodiment of the present disclosure. [Figure 17] 1 shows a flowchart of a method for training a search system based on a pre-trained language model according to an exemplary embodiment of the present disclosure. [Figure 18] FIG. 1 illustrates a structural block diagram of an exemplary electronic device that can be used to implement embodiments of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] Hereinafter, exemplary embodiments of the present disclosure will be described with reference to the drawings. The following description includes various details of the embodiments of the present disclosure for ease of understanding, but such details should be regarded as merely exemplary. Therefore, as would be obvious to one skilled in the art, various changes and modifications can be made to the embodiments described herein without departing from the scope of the present disclosure. Similarly, in the following description, well-known functions and structures are omitted for clarity and simplicity.

[0017] In this disclosure, unless otherwise specified, when terms such as "first" and "second" are used to describe various elements, they are not intended to limit the location, timing, or importance of these elements, but are merely used to distinguish one element from another. In some examples, a first element and a second element can refer to the same instance of the element, and in some cases, they can refer to different instances based on the context.

[0018] In this disclosure, the terms used in the description of various examples are merely for describing specific examples and are not intended to be limiting. Unless otherwise specified in the context, when the number of elements is not specifically limited, the elements may be one or more. In addition, the term "and / or" used in this disclosure covers any one and all possible combinations of the items shown.

[0019] Data search is one of the basic services of the Internet, and can provide search results that meet the needs of users based on their search requests.

[0020] The inventor creatively proposes a method for searching data, by collecting various different forms of network resources, such as text, picture, video, table, etc., into a unified resource database with a unified vector representation, converting different types of data into corresponding data semantic vectors. Thus, by comparing the semantic vector corresponding to the user's search request with the semantic vector corresponding to at least one data, the search results for the user's search request can be recalled from at least one data. Thus, by mapping different types of data into the same semantic vector space, similarity matching can be directly performed on the semantic vectors corresponding to different types of data, and multi-modal data matching the user's search request can be obtained, which is helpful in improving the search performance, specifically, it can improve the diversity of search content and improve the user experience.

[0021] The attributes of the data include at least one of modality, language, and data structure. The modality of the data includes text, picture, and video, the language includes various kinds of languages ​​such as Chinese, English, etc., and the data structure includes structured data (e.g., table, graph) and unstructured data. Therefore, the solution of the embodiment of the present application can realize a multi-modal, multi-language, and multi-resource unified search method.

[0022] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings.

[0023] FIG. 1 shows a flowchart of a search method according to an exemplary embodiment of the present disclosure.

[0024] As shown in Fig. 1, the method may include: converting a user's search request into a first request semantic vector in step S101; searching for at least one first data semantic vector matching the first request semantic vector in a search resource database in step S102, where the search resource database is constructed as a semantic vector space, in which different types of data are converted into corresponding data semantic vectors, the different types of data at least including text, picture and video; and generating search results based on the at least one first data semantic vector in step S103, thereby realizing a unified search for different types of data, improving the diversity of search results and improving user experience.

[0025] For example, the search resource database may contain a large amount of different types of data, such as text, pictures, videos, etc., and the different types of data may be obtained in various ways, for example, by accessing a database of existing standardized data content, the data content of which is converted into corresponding data semantic vectors in the semantic vector space of the search resource database.

[0026] The user's search request at least includes search information input by the user, the search information including but not limited to text information or voice information, and the semantic features of the search information are extracted and transformed into a first request semantic vector, which is used to match the search and obtain data that meets the user's needs.

[0027] According to some embodiments, in the search resource database, the data semantic vectors converted from the different types of data have a unified specification, thereby realizing a unified search for the different types of data.

[0028] As can be understood, the first request semantic vector has the same specifications as the data semantic vector in the search resource database, for example, if the data semantic vector in the search resource database is a 1000-dimensional vector, the search request should also be converted into a 1000-dimensional vector of the same specifications.

[0029] Exemplarily, the similarity between the first request semantic vector and each data semantic vector in the search resource database can be calculated, and at least one first data semantic vector matching the first request semantic vector can be obtained based on the similarity. First data corresponding to each first data semantic vector can be obtained based on the at least one first data, and a search result can be generated and displayed to a user based on the at least one first data. The similarity between vectors can be, for example, but not limited to, cosine similarity.

[0030] According to some embodiments, in addition to text, picture and video, the different types of data at least include tables and knowledge graphs. As can be understood, the different types of data may further include other types of data, such as maps, videos, etc., and more types of data can further enrich the search resource database, thereby further improving the diversity of search results, better meeting the needs of users, and improving the user experience.

[0031] According to some embodiments, at least one of the text or video data in the different types of data is obtained by performing fine-grained partitioning on the original complete data, which can provide a deeper understanding of the data content, realize a fine-grained index, and provide search results that are more tailored to the needs of users.

[0032] Exemplarily, obtaining at least one text or video data by performing fine-grained division on the original complete data may include performing fine-grained division on the original complete data according to semantics. In some embodiments, performing fine-grained division on the original complete data may include performing semantic division on the original complete data to obtain at least one text or video data. Taking webpage text data as an example, the original complete webpage text data may include a plurality of paragraphs, and each paragraph may have different semantic features, so that the data semantic vector corresponding to the corresponding complete webpage text data cannot fully represent the different semantic features of each paragraph, and the first required semantic vector embodying the user's needs cannot be matched with the meaning of each paragraph in the search process. By dividing the original complete webpage text data, it can be divided into a plurality of segments having different semantic features, and each segment corresponds to one of the at least one text. Each segment is converted into a corresponding data semantic vector, and is matched with the first required semantic vector embodying the user's needs in the search process, so that a search result that is more in line with the user's needs can be obtained. Similarly, fine-grained segmentation can be performed based on video text data corresponding to the video, and the specific principles and processes are similar to those of web page text data.

[0033] According to some embodiments, each data semantic vector in the semantic vector space includes a dimension related to the content quality of the corresponding data, which may be, but is not limited to, a content quality score of the corresponding data, so that the content quality of the data can be taken into account during the search process to improve the quality of the search results.

[0034] According to some embodiments, each data semantic vector in the semantic vector space includes a dimension related to the delivery time of the corresponding data, which may be, but is not limited to, the delivery time of the corresponding data, so that the timeliness of data can be taken into account in the search process, and the quality of the search results can be improved.

[0035] According to some embodiments, each data semantic vector in the semantic vector space includes a dimension related to the source trustworthiness of the corresponding data, which may be, but is not limited to, the source website type of the corresponding data and the trustworthiness of the corresponding website type, so that the authority of data can be taken into account in the search process and the quality of search results can be improved.

[0036] According to some embodiments, each data semantic vector in the semantic vector space includes at least two of the following dimensions of the corresponding data: content quality, delivery time, and source credibility. In the search process, the quality of the search results can be further improved by increasing the multiple dimensions of the data.

[0037] Based on a similar principle, according to some embodiments, the first request semantic vector includes context information about the user's search, including at least one of time, location, terminal device, and the user's route-oriented search, which can further improve the accuracy of the search.

[0038] As can be seen, based on the context information of the user's search, the user's direct search request can be more accurately described. For example, if the search information input by the user is "what's the weather tomorrow", the first request semantic vector may correspondingly include the location where the user is located, for example, Beijing, so that the search results related to "tomorrow's weather in Beijing" can be provided to the user, which can more accurately meet the user's needs and improve the user experience.

[0039] According to the research of the inventors, after a user queries one search request, the user may continue to query another related search request, and the other related search request may embody the related search intent when the user queries the first search request. For example, the user may query "Who is the wife of public figure A?" and then continue to query "Where is the hometown of public figure B?" based on the search result "Public figure A's wife is public figure B". Therefore, when querying a user's search request, it is advantageous to further mine the user's related search intent to improve search performance.

[0040] Based on this, according to some embodiments, as shown in Figure 2, the search method further includes the steps of: obtaining the user's related search intent based on the user's search request in step S201; converting the related search intent into a second request semantic vector in step S202; searching the search resource database for at least one second data semantic vector that matches the second request semantic vector in step S203; and generating a recommendation result based on the at least one second data semantic vector in step S204. Thus, by responding to the user's search request, obtaining the user's related search intent and making a recommendation based on the obtained intent, the query result for the search request input by the user can further meet the user's expansion needs, thereby further improving the user experience. The search method may have one or more related expansion intents.

[0041] According to some embodiments, obtaining the user's related search intent includes obtaining the user's related search intent based on a needs graph, the needs graph including search need distribution information, search need related information, etc., obtained based on big data, so that the user's related search intent can be more accurately obtained based on the user's direct search intent.

[0042] For example, the user's related search intent may be obtained by querying popular search needs related to the user's search request based on the user's search request. In one example, if the search request input by the current user is "how to get to tourist spot A", querying the needs graph to find that the popular search needs related to "tourist spot A" are "ticket reservation methods for tourist spot A", "tourist map for tourist spot A", etc., the user's related search intent is correspondingly found to be "preparations required for sightseeing at tourist spot A", and thus content such as ticket reservations and tourist guides for tourist spot A can be recommended to the user. Therefore, the user's related search intent can also be understood as a latent search intent.

[0043] According to another embodiment, obtaining the user's related search intent includes obtaining the user's related search intent based on the user's search history in the same session process, so as to more reliably obtain the user's related search intent based on the user's search history.

[0044] For example, if the search request input by the user is "How do I get to tourist spot A?", according to the user's search history, it can be determined that the search history of the user's session process contains historical search information such as "How do I get to tourist spot B?" and "How do I get to tourist spot C?". By performing semantic analysis on the historical search information, it can be determined that the user's related search intent is "inquire about a certain tourist spot", and content related to other tourist spots in a certain area can be recommended to the user.

[0045] In the above, two specific examples of how to obtain a user's related search intent are described, so that it can be understood that the manner of obtaining a user's related search intent is not limited to the above two kinds, for example, the user's related search intent can also be obtained based on the user's relationship graph.

[0046] In addition, obtaining the user's related search intent may be realized, for example, by performing structured analysis on the user's search request. The structured analysis may be realized, for example, based on the user's intent graph. Referring to the example of FIG. 6, for example, when the search information input by the user is "hairstyle", the user's search log is correspondingly obtained, mining modeling is performed on the search log, the user's intent graph is obtained, and the result of the search request analysis is combined with comprehensive computer reasoning to obtain the user's related search intent, and based on this, content in various fields that match the related search intent, such as "barbershops with good haircuts" corresponding to the beauty and fashion field, "automatic clippers" in the goods and merchandise field, and "barber training" in the education and training field, are recommended to the user.

[0047] After determining the related search intention corresponding to the user's search request, execute steps S202 to S204, thus searching for at least one matching second data semantic vector from the search resource database based on the second request semantic vector corresponding to the related search intention, and generating recommendation results based on the at least one second data semantic vector.

[0048] According to some embodiments, each second data meaning vector in the at least one second data meaning vector respectively includes dimensions related to semantic relevance features and perceived relevance features of corresponding data, where the semantic relevance features are used to explain the direct meaning of the corresponding data, and the perceived relevance features emphasize dimensions related to the needs and interests of users, so as to further improve the accuracy of the recommendation system and better meet the potential needs of users.

[0049] Correspondingly, according to some embodiments, the second request semantic vector includes dimensions related to semantic relevance features and perceived relevance features of the search request, respectively, where the semantic relevance features are used to explain the direct meaning of the search request, and the perceived relevance features emphasize dimensions related to user needs and interests, so as to further improve the accuracy of the recommendation system and better meet the potential needs of users.

[0050] For example, the semantic relevance feature dimension in the data semantic vector corresponding to a webpage containing content introducing public figure A is used to describe the direct meaning of the webpage content. The sensed relevance feature dimension in the data semantic vector corresponding to the webpage focuses on describing the user's possible interests that can be expanded around public figure A, for example, the user may be interested in who public figure A's wife is, what works public figure A has, etc., and the sensed relevance feature dimension may include the corresponding content, which can further improve the accuracy of the recommendation system and better meet the potential needs of the user.

[0051] According to some embodiments, both the first data meaning vectors for recalling the user's search request and the second data meaning vectors for recalling the user's related search intent may be completely the same, partly the same, or completely different. That is, both the first data corresponding to the first data meaning vectors and the second data corresponding to the second data meaning vectors may be completely the same, partly the same, or completely different. One or more first data matching the user's search request are obtained by recalling from the first data, and one or more second data matching the user's related search intent are obtained by recalling from the second data.

[0052] In some embodiments, the number of the plurality of second data is smaller than the number of the plurality of first data, thereby realizing accurate recommendation, improving recommendation effect, and further improving user experience.

[0053] According to some embodiments, the search resource database includes a first sub-library and a second sub-library used only for the related search intent, and the amount of data in the second sub-library is smaller than the amount of data in the first sub-library. By providing two sub-libraries separately, the respective objectives of search and recommendation can be strengthened, the first sub-library is used to meet the mass recall requirement of related search, and the second sub-library is used to meet the exact recall requirement of meeting the user's potential requirements, thereby better improving the quality of the query result and further improving the user experience.

[0054] For example, the amount of data in the first sub-library may be on the order of 10 billion or 100 billion, which can cover many content resources and more comprehensively cover the content needs of users, and correspondingly, the amount of data in the second sub-library may be on the order of one million.

[0055] As can be appreciated, the data in the second sub-library that is applied for recommendation may be selected based on predefined quality criteria, thereby providing the user with higher quality recommended content, better meeting the user's enhanced needs and improving the user experience.

[0056] Exemplarily, a similarity between the second request semantic vector and each second data semantic vector in the search resource database can be calculated, and at least one second data semantic vector matching the second request semantic vector can be obtained based on the similarity. Second data corresponding to each second data semantic vector can be obtained based on the at least one second data, and a recommendation result can be generated and displayed to a user based on the at least one second data.

[0057] In the above technical solution, it is possible to respond to a user's search request, generate search results, determine the user's relevant search intent, and generate recommendation results, thereby accurately meeting the user's direct needs, and also expanding the user's horizons and satisfying the user's expanded needs.

[0058] According to some embodiments, as shown in Fig. 3, the step S103 of generating search results based on the at least one first data semantic vector includes the steps of: obtaining one or more first data corresponding to the at least one first data semantic vector; and sorting the one or more first data based on a first sorting logic S1032, the first sorting logic being based on at least one of the content quality, delivery time, and source reliability of the corresponding data. This allows the relevance of the search results to the search request to be more considered, and the search results to be more in line with the needs of the user can be generated, improving the user experience.

[0059] According to some embodiments, as shown in Fig. 4, the step S204 of generating recommendation results based on the at least one second data semantic vector includes the steps of: obtaining one or more second data corresponding to the at least one second data semantic vector; and sorting the one or more second data based on a second sorting logic different from the first sorting logic based on the semantic relevance features and the perception relevance features of the corresponding data in a step S2042. This makes it possible to obtain recommendation results by taking into more consideration the user's perception and interests, generate recommendation results that are more focused on perception of potential needs, and improve the user experience.

[0060] In the above two embodiments, the search results and the recommendation results are generated based on different sorting logics. The first sorting logic places more importance on the semantic relevance of data, content quality, source reliability, etc., and can obtain search results that are more relevant to the user's direct search request. Correspondingly, the second sorting logic places more importance on sensing the user's potential needs, and can obtain recommendation results that are more in line with the user's perceptions and interests. As can be seen, by generating search results and recommendation results using two different kinds of sorting logics, respectively, it is possible to better meet the user's direct search request and potential search request, and improve the user experience.

[0061] According to some embodiments, as shown in Fig. 3, the step S103 of generating search results based on the at least one first data semantic vector further includes the step S1033 of determining one or more corresponding data based on the at least one first data semantic vector, and organizing the one or more data based on an understanding of the corresponding content to generate the search results, so as to simplify the search results and more accurately meet the needs of users.

[0062] For example, if the search request input by the user is "cuisine near tourist spot A," the multiple data include "menu of cuisine B near tourist spot A," "business hours and address of cuisine B near tourist spot A," "menu of cuisine C near tourist spot A," "address of cuisine C near tourist spot A," etc., and based on understanding of the corresponding content, the multiple data can be organized into "menu, business hours and address of cuisine B near tourist spot A," and "menu and address of cuisine C near tourist spot A" and shown to the user. This makes it possible to provide the user with search results that better match the user's needs and improve the user experience.

[0063] As can be seen, the manner of organizing one or more data is not limited to the simple superposition manner in the above example, and a suitable organizing manner can be designed based on a specific application scenario, for example, if one or more data is a text segment, semantic understanding can be performed on the one or more data, and a recommendation result can be generated based on the semantic understanding result, for example, if the text segment describes the weather conditions such as temperature, humidity, and sunlight intensity in a certain area, a recommendation result such as "a certain area is suitable for outdoor activities" can be generated based on the semantic understanding of the text segment.

[0064] According to some embodiments, as shown in Fig. 5, the search method includes step S501 of performing a structured analysis on the search request to determine whether there is ambiguity, and step S502 of providing the user with content requiring clarification in response to determining that there is ambiguity. The functions and implementation manners of steps S503 to S505 are similar to steps S101 to S103 in Fig. 1, and will not be described in detail here. This can realize clarification of the user's search request and guidance to the user's needs.

[0065] A structured analysis may be performed on the search request to determine whether the search request is unambiguous and unambiguous. The structured analysis may be implemented, for example, based on a knowledge graph.

[0066] Next, referring to the example of Fig. 6, a structured analysis is performed on the search information "hairstyle" input by the user, and the information is combined with a knowledge graph based on big data to perform comprehensive computer inference to obtain content that can be provided to the user who needs clarification. The [hairstyle] input by the user may have multiple degrees of ambiguity, such as men's hairstyles and women's hairstyles, short hairstyles and long hairstyles, hairstyles for long faces and hairstyles for round faces, etc., and the content that needs clarification is provided to the user, so that the user's needs can be more clearly guided.

[0067] In the technical solutions disclosed herein, the collection, storage, use, processing, transmission, provision, disclosure and other processing of user personal information shall all comply with the provisions of relevant laws and regulations and shall not violate public order and morals.

[0068] According to another aspect of the present disclosure, a data search device is provided, referring to Fig. 7, the device includes: a first conversion unit 701 for converting a user's search request into a first request semantic vector; a first search unit 702 for searching at least one first data semantic vector matching the first request semantic vector in a search resource database, the search resource database is constructed as a semantic vector space, in the semantic vector space, different types of data are converted into corresponding data semantic vectors, the different types of data at least including text, picture and video; and a first generation unit 703 for generating search results according to the at least one first data semantic vector, thereby realizing a unified search for data of different types and different structures, improving the diversity of search results, and improving user experience.

[0069] For example, the search resource database may contain a large amount of different types of data, such as text, pictures, videos, etc., and the different types of data may be obtained in various ways, for example, by accessing a database of existing standardized data content, the data content of which is converted into corresponding data semantic vectors in the semantic vector space of the search resource database.

[0070] The user's search request at least includes search information input by the user, the search information including but not limited to text information or voice information, and the semantic features of the search information are extracted and transformed into a first request semantic vector, which is used to match the search and obtain data that meets the user's needs.

[0071] According to some embodiments, in the search resource database, the data semantic vectors converted from the different types of data have a unified specification, thereby realizing a unified search for the different types of data.

[0072] As can be understood, the first request semantic vector has the same specification as the data semantic vector in the search resource database, for example, if the data semantic vector in the search resource database is a 1000-dimensional vector, the search request should also be converted into a 1000-dimensional vector with the same specification. Exemplarily, the similarity between the first request semantic vector and each data semantic vector in the search resource database can be calculated, and at least one first data semantic vector matching with the first request semantic vector can be obtained based on the similarity. Based on the at least one first data semantic vector, first data corresponding to each first data semantic vector can be obtained, and further based on the at least one first data, a search result can be generated and displayed to a user. The similarity between vectors can be, for example, but not limited to, cosine similarity.

[0073] According to some embodiments, in addition to text, picture and video, the different types of data include at least a table and a knowledge graph. As can be understood, the different types of data may further include other types of data, such as maps, videos, etc., and more types of data can further enrich the search resource database, thereby further improving the diversity of search results, better meeting the needs of users, and improving the user experience.

[0074] According to some embodiments, at least one of the text or video data in the different types of data is obtained by performing fine-grained partitioning on the original complete data, which can provide a deeper understanding of the data content, realize a fine-grained index, and provide search results that are more tailored to the needs of users.

[0075] Exemplarily, obtaining the at least one text or video data by performing fine-grained division on the original complete data may include performing fine-grained division on the original complete data according to semantics. In some embodiments, performing fine-grained division on the original complete data may include performing semantic division on the original complete data to obtain at least one text or video data. Taking webpage text data as an example, the original complete webpage text data may include a plurality of paragraphs, and each paragraph may have different semantic features, so that the data semantic vector corresponding to the corresponding complete webpage text data cannot fully represent the different semantic features of each paragraph, and the first required semantic vector embodying the needs of the user cannot be matched with the meaning of each paragraph in the search process. By dividing the original complete webpage text data, it can be divided into a plurality of segments having different semantic features, and each segment corresponds to one of the at least one text. Each segment is converted into a corresponding data semantic vector, and is matched with the first required semantic vector embodying the needs of the user in the search process, so that a search result that is more in line with the needs of the user can be obtained. Similarly, fine-grained segmentation can be performed based on video text data corresponding to the video, and the specific principles and processes are similar to those of web page text data.

[0076] According to some embodiments, each data semantic vector in the semantic vector space includes a dimension related to the content quality of the corresponding data, which may be, but is not limited to, a content quality score of the corresponding data, so that the content quality of the data can be taken into account during the search process to improve the quality of the search results.

[0077] According to some embodiments, each data semantic vector in the semantic space vector includes a dimension related to the delivery time of the corresponding data, which may be, but is not limited to, the delivery time of the corresponding data, so that the timeliness of data can be taken into account in the search process, and the quality of the search results can be improved.

[0078] According to some embodiments, each data semantic vector in the semantic space vector includes a dimension related to the source authority of the corresponding data, which may be, but is not limited to, the source website type of the corresponding data and the authority of the corresponding website type, so that the authority of data can be taken into account in the search process and the quality of the search results can be improved.

[0079] According to some embodiments, each data semantic vector in the semantic space vector includes at least two of the following dimensions of the content quality, delivery time, and source credibility of the corresponding data. In the search process, the quality of the search results can be further improved by increasing the multiple dimensions of the data.

[0080] Based on a similar principle, according to some embodiments, the first request semantic vector includes context information about the user's search, including at least one of time, location, terminal device, and the user's route-oriented search, which can further improve the accuracy of the search.

[0081] As can be seen, based on the context information of the user's search, the user's direct search request can be more accurately described. For example, if the search information input by the user is "what's the weather tomorrow", the first request semantic vector may correspondingly include the location where the user is located, for example, Beijing, so that the search results related to "tomorrow's weather in Beijing" can be provided to the user, which can more accurately meet the user's needs and improve the user experience.

[0082] According to the research of the inventors, after a user queries one search request, the user may continue to query another related search request, and the other related search request may embody the related search intent when the user queries the first search request. For example, the user may query "Who is the wife of public figure A?" and then continue to query "Where is the hometown of public figure B?" based on the search result "Public figure A's wife is public figure B". Therefore, when querying a user's search request, it is advantageous to further mine the user's related search intent to improve search performance.

[0083] Based on this, according to some embodiments, as shown in FIG. 8, the data search device further includes an acquisition unit 804 for acquiring the user's related search intention based on the user's search request, a second conversion unit 805 for converting the related search intention into a second request semantic vector, a second search unit 806 for searching the search resource database for at least one second data semantic vector matching the second request semantic vector, and a second generation unit 807 for generating a recommendation result based on the at least one second data semantic vector. The functions and implementation manners of units 801 to 803 in FIG. 8 are similar to those of units 701 to 703 in FIG. 7, and will not be described in detail here. Thus, by responding to the user's search request, acquiring the user's related search intention, and making recommendations based on the acquired related search intention, the query result for the search request input by the user can further meet the user's expansion needs, thereby further improving the user experience. The related expansion intention may have one or more.

[0084] According to some embodiments, the unit for obtaining the user's related search intent according to the user's search request is further configured to obtain the user's related search intent according to a needs graph, the needs graph including search need distribution information, search need related information, etc. obtained based on big data, so as to more accurately obtain the user's related search intent.

[0085] For example, the user's related search intent may be obtained by querying popular search needs related to the user's search request. In one example, if the search request input by the current user is "how to get to tourist spot A", querying the needs graph to find that the popular search needs related to "tourist spot A" are "ticket reservation methods for tourist spot A", "tourist map for tourist spot A", etc., the user's related search intent is found to be "preparations required for sightseeing at tourist spot A", and thus content such as ticket reservations and tourist guides for tourist spot A can be recommended to the user.

[0086] According to another embodiment, the unit for obtaining the user's related search intent based on the user's search request is further configured to obtain the user's related search intent based on the user's search history in the same session process, so as to more reliably obtain the user's related search intent based on the user's search history.

[0087] For example, if the search request input by the user is "How do I get to tourist spot A?", according to the user's search history, it can be determined that the search history of the user's session process contains historical search information such as "How do I get to tourist spot B?" and "How do I get to tourist spot C?". By performing semantic analysis on the historical search information, it can be determined that the user's related search intent is "inquire about a certain tourist spot", and content related to other tourist spots in a certain area can be recommended to the user.

[0088] In the above, two specific examples of how to obtain a user's related search intent are described, so that it can be understood that the manner of obtaining a user's related search intent is not limited to the above two kinds, for example, the user's related search intent can also be obtained based on the user's relationship graph.

[0089] Also, for example, the acquisition of the user's related search intention may be realized by performing a structured analysis on the user's search request. The structured analysis may be realized, for example, based on the user's intention graph. Referring to the example of FIG. 6, for example, when the search information input by the user is "hairstyle", the user's search log is acquired correspondingly, mining modeling is performed on the search log, the user's intention graph is obtained, and the result of the search request analysis is combined with comprehensive computer inference to acquire the user's related search intention, and based on this, content in various fields that match the related search intention, such as "barbershops with good haircuts" corresponding to the beauty and fashion field, "automatic clippers" in the goods field, and "barber training" in the education and training field, is recommended to the user.

[0090] After determining the related search intention corresponding to the user's search request, execute steps S202 to S204, thus searching for at least one matching second data semantic vector from the search resource database based on the second request semantic vector corresponding to the related search intention, and generating recommendation results based on the at least one second data semantic vector.

[0091] According to some embodiments, each second data meaning vector in the at least one second data meaning vector respectively includes dimensions related to semantic relevance features and perceived relevance features of corresponding data, where the semantic relevance features are used to explain the direct meaning of the corresponding data, and the perceived relevance features emphasize dimensions related to the needs and interests of users, so as to further improve the accuracy of the recommendation system and better meet the potential needs of users.

[0092] Correspondingly, according to some embodiments, the second request semantic vector includes dimensions related to semantic relevance features and perceived relevance features of the search request, respectively, where the semantic relevance features are used to explain the direct meaning of the corresponding data, and the perceived relevance features emphasize dimensions related to the needs and interests of users, so as to further improve the accuracy of the recommendation system and better meet the potential needs of users.

[0093] For example, the semantic relevance feature dimension in the data semantic vector corresponding to a webpage containing content introducing public figure A is used to describe the direct meaning of the webpage content. The sensed relevance feature dimension in the data semantic vector corresponding to the webpage focuses on describing the user's possible interests that can be expanded around public figure A, for example, the user may be interested in who public figure A's wife is, what works public figure A has, etc., and the sensed relevance feature dimension may include the corresponding content, which can further improve the accuracy of the recommendation system and better meet the potential needs of the user.

[0094] According to some embodiments, both the first data meaning vectors for recalling the user's search request and the second data meaning vectors for recalling the user's related search intent may be completely the same, partly the same, or completely different. That is, both the first data corresponding to the first data meaning vectors and the second data corresponding to the second data meaning vectors may be completely the same, partly the same, or completely different. One or more first data matching the user's search request are obtained by recalling from the first data, and one or more second data matching the user's related search intent are obtained by recalling from the second data.

[0095] In some embodiments, the number of the plurality of second data is smaller than the number of the plurality of first data, thereby realizing accurate recommendation, improving recommendation effect, and further improving user experience.

[0096] According to some embodiments, the search resource database includes a first sub-library and a second sub-library used only for the related search intent, and the amount of data in the second sub-library is smaller than the amount of data in the first sub-library. By providing two sub-libraries separately, the respective objectives of search and recommendation can be strengthened, the first sub-library is used to meet the mass recall requirement of related search, and the second sub-library is used to meet the exact recall requirement of meeting the user's potential requirements, thereby better improving the quality of the query result and further improving the user experience.

[0097] For example, the amount of data in the first sub-library may be on the order of 10 billion or 100 billion, which can cover many content resources and more comprehensively cover the content needs of users, and correspondingly, the amount of data in the second sub-library may be on the order of one million.

[0098] As can be appreciated, the data in the second sub-library that is applied for recommendation may be selected based on predetermined quality criteria, thereby providing higher quality recommended content to the user, better meeting the user's enhanced needs and improving the user experience.

[0099] Exemplarily, a similarity between the second request semantic vector and each second data semantic vector in the search resource database can be calculated, and at least one second data semantic vector matching the second request semantic vector can be obtained based on the similarity. Second data corresponding to each second data semantic vector can be obtained based on the at least one second data, and a recommendation result can be generated and displayed to a user based on the at least one second data.

[0100] In the above technical solution, it is possible to respond to a user's search request, generate search results, determine the user's relevant search intent, and generate recommendation results, thereby accurately meeting the user's direct needs, and also expanding the user's horizons and satisfying the user's expanded needs.

[0101] According to some embodiments, as shown in Fig. 9, the first generating unit 703 for generating search results based on the at least one first data semantic vector further includes: a first acquiring subunit 7031 for acquiring one or more first data corresponding to the at least one first data semantic vector; and a first sorting subunit 7032 for sorting the one or more first data according to a first sorting logic based on at least one of the content quality, delivery time, and source reliability of the corresponding data. This can better consider the relevance of the search results to the search request, generate search results that are more in line with the needs of users, and improve the user experience.

[0102] According to some embodiments, as shown in Fig. 10, the second generating unit 807 for generating recommendation results based on the at least one second data semantic vector includes: a second acquiring subunit 8071 for acquiring one or more second data corresponding to the at least one second data semantic vector; and a second sorting subunit 8072 for sorting the one or more second data based on a second sorting logic different from the first sorting logic based on the semantic relevance features and the sensing relevance features of the corresponding data. This can obtain recommendation results by taking into more consideration the user's sensitivity and interests, generate recommendation results that are more focused on sensing potential needs, and improve user experience.

[0103] In the above two embodiments, the search results and the recommendation results are generated based on different sorting logics. The first sorting logic places more importance on the semantic relevance of data, content quality, source reliability, etc., and can obtain search results that are more relevant to the user's direct search request. Correspondingly, the second sorting logic places more importance on sensing the user's potential needs, and can obtain recommendation results that are more in line with the user's perceptions and interests. As can be seen, by generating search results and recommendation results using two different kinds of sorting logics, respectively, it is possible to better meet the user's direct search request and potential search request, and improve the user experience.

[0104] As can be seen, the manner of organizing one or more data is not limited to the simple superposition manner in the above example, and a suitable organizing manner can be designed based on a specific application scenario, for example, if one or more data is a text segment, semantic understanding can be performed on the one or more data, and a recommendation result can be generated based on the semantic understanding result, for example, if the text segment describes the weather conditions such as temperature, humidity, and sunlight intensity in a certain area, a recommendation result such as "a certain area is suitable for outdoor activities" can be generated based on the semantic understanding of the text segment.

[0105] According to some embodiments, as shown in Fig. 9, the first generating unit 703 for generating search results based on the at least one first data semantic vector further includes a sorting subunit 7033 for sorting the one or more first data to generate the search results based on corresponding content understanding, so as to simplify the search results and more accurately meet the needs of users.

[0106] For example, if the search request input by the user is "cuisine near tourist spot A," the multiple data include "menu of cuisine B near tourist spot A," "business hours and address of cuisine B near tourist spot A," "menu of cuisine C near tourist spot A," "address of cuisine C near tourist spot A," etc., and based on understanding of the corresponding content, the multiple data can be organized into "menu, business hours and address of cuisine B near tourist spot A," and "menu and address of cuisine C near tourist spot A" and shown to the user. This makes it possible to provide the user with search results that better match the user's needs and improve the user experience.

[0107] According to some embodiments, as shown in Fig. 11, a data search device 1100 includes an analysis subunit 1101 for performing structured analysis on the search request to determine whether there is ambiguity, and a subunit 1102 for providing a content requiring clarification to a user in response to determining that there is ambiguity. The functions and implementation manners of units 1103 to 1105 are similar to those of units 701 to 703 in Fig. 7, and will not be described in detail here. This can realize clarification of the user's search request and guidance to the user's needs.

[0108] A structured analysis may be performed on the search request to determine whether the search request is unambiguous and unambiguous. The structured analysis may be implemented, for example, based on a knowledge graph.

[0109] Next, referring to the example of Fig. 6, a structured analysis is performed on the search information "hairstyle" input by the user, and the information is combined with a knowledge graph based on big data to perform comprehensive computer inference to obtain content that can be provided to the user who needs clarification. The [hairstyle] input by the user may have multiple degrees of ambiguity, such as men's hairstyles and women's hairstyles, short hairstyles and long hairstyles, hairstyles for long faces and hairstyles for round faces, etc., and the content that needs clarification is provided to the user, so that the user's needs can be more clearly guided.

[0110] According to another aspect of the present disclosure, as shown in Fig. 12, a search system 1200 based on a pre-trained language model is further provided, the search system 1200 includes: a recall model 1201 configured to convert a user's search request into a first request semantic vector, and search at least one first data semantic vector matching the first request semantic vector in a search resource database; and a sorting model 1202 configured to sort the one or more first data by sequentially determining the similarity between the search request and one or more first data corresponding to the at least one first data semantic vector. The recall model and the sorting model are cascaded to form an end-to-end deep neural network basic model, which can realize a unified search for data of different types and different structures, improve the diversity of search results, and improve user experience.

[0111] Exemplarily, the recall model may be based on a double encoder structure, and the sorting model may be based on a cross encoder structure. FIG. 15 shows a schematic diagram of a model structure based on a cross encoder structure and a schematic diagram of a model structure based on a double encoder structure. A question sentence (i.e., a search request input by a user) and a passage sentence (i.e., data in a search resource database) are input to a recall model based on a double encoder structure, and the two sentences are transmitted to two independent encoding networks, which may be the same network. The two encoding networks respectively generate semantic vectors E corresponding to the two sentences. q (q) and E p The output of the method is (p), which allows a similarity calculation to be performed based on the semantic vector. Based on the similarity, one or more passage sentences that match the question sentence are obtained.

[0112] The question sentence and one or more passage sentences recalled by the recall model are input into a recommendation model configured based on a cross-encoder, and a similarity score between the question sentence and one or more passage sentences is output. The one or more passage sentences are sorted based on the similarity score. Here, the input recommendation model is a semantic vector corresponding to the question sentence and one or more recalled passage sentences, respectively.

[0113] As can be appreciated, the above model structures are only examples of the structures of the recall model and sort model described in this disclosure, and the recall model and sort model may be models based on other structures.

[0114] For example, the search resource database may contain a large amount of different types of data, such as text, pictures, videos, etc., and the different types of data may be obtained in various ways, for example, by accessing a database of existing standardized data content, the data content of which is converted into corresponding data semantic vectors in the semantic vector space of the search resource database.

[0115] The user's search request at least includes search information input by the user, the search information including but not limited to text information or voice information, and the semantic features of the search information are extracted and transformed into a first request semantic vector, which is used to match the search and obtain data that meets the user's needs.

[0116] According to some embodiments, in the search resource database, the data semantic vectors converted from the different types of data have a unified specification, thereby realizing a unified search for the different types of data.

[0117] As can be understood, the first request semantic vector has the same specification as the data semantic vector in the search resource database, for example, if the data semantic vector in the search resource database is a 1000-dimensional vector, the search request should also be converted into a 1000-dimensional vector with the same specification. Exemplarily, the similarity between the first request semantic vector and each data semantic vector in the search resource database can be calculated, and at least one first data semantic vector matching with the first request semantic vector can be obtained based on the similarity. Based on the at least one first data semantic vector, first data corresponding to each first data semantic vector can be obtained, and further based on the at least one first data, a search result can be generated and displayed to a user. The similarity between vectors can be, for example, but not limited to, cosine similarity.

[0118] According to some embodiments, in addition to text, picture and video, the different types of data include at least a table and a knowledge graph. As can be understood, the different types of data may further include other types of data, such as maps, videos, etc., and more types of data can further enrich the search resource database, thereby further improving the diversity of search results, better meeting the needs of users, and improving the user experience.

[0119] According to some embodiments, at least one of the text or video data in the different types of data is obtained by performing fine-grained partitioning on the original complete data, which can provide a deeper understanding of the data content, realize a fine-grained index, and provide search results that are more tailored to the needs of users.

[0120] Exemplarily, obtaining the at least one text or video data by performing fine-grained division on the original complete data may include performing fine-grained division on the original complete data according to semantics. In some embodiments, performing fine-grained division on the original complete data may include performing semantic division on the original complete data to obtain at least one text or video data. Taking webpage text data as an example, the original complete webpage text data may include a plurality of paragraphs, and each paragraph may have different semantic features, so that the data semantic vector corresponding to the corresponding complete webpage text data cannot fully represent the different semantic features of each paragraph, and the first required semantic vector embodying the needs of the user cannot be matched with the meaning of each paragraph in the search process. By dividing the original complete webpage text data, it can be divided into a plurality of segments having different semantic features, and each segment corresponds to one of the at least one text. Each segment is converted into a corresponding data semantic vector, and is matched with the first required semantic vector embodying the needs of the user in the search process, so that a search result that is more in line with the needs of the user can be obtained. Similarly, fine-grained segmentation can be performed based on video text data corresponding to the video, and the specific principles and processes are similar to those of web page text data.

[0121] According to some embodiments, each data semantic vector in the semantic vector space includes a dimension related to the content quality of the corresponding data, which may be, but is not limited to, a content quality score of the corresponding data, so that the content quality of the data can be taken into account during the search process to improve the quality of the search results.

[0122] According to some embodiments, each data semantic vector in the semantic space vector includes a dimension related to the delivery time of the corresponding data, which may be, but is not limited to, the delivery time of the corresponding data, so that the timeliness of data can be taken into account in the search process, and the quality of the search results can be improved.

[0123] According to some embodiments, each data semantic vector in the semantic space vector includes a dimension related to the source authority of the corresponding data, which may be, but is not limited to, the source website type of the corresponding data and the authority of the corresponding website type, so that the authority of data can be taken into account in the search process and the quality of the search results can be improved.

[0124] According to some embodiments, each data semantic vector in the semantic space vector includes at least two of the following dimensions of the content quality, delivery time, and source credibility of the corresponding data. In the search process, the quality of the search results can be further improved by increasing the multiple dimensions of the data.

[0125] Based on a similar principle, according to some embodiments, the first request semantic vector includes context information about the user's search, including at least one of time, location, terminal device, and the user's route-oriented search, which can further improve the accuracy of the search.

[0126] As can be seen, based on the context information of the user's search, the user's direct search request can be more accurately described. For example, if the search information input by the user is "what's the weather tomorrow", the first request semantic vector may correspondingly include the location where the user is located, for example, Beijing, so that the search results related to "tomorrow's weather in Beijing" can be provided to the user, which can more accurately meet the user's needs and improve the user experience.

[0127] According to some embodiments, as shown in FIG. 13, the search system further includes a related search intent acquisition model 1303 configured to acquire related search intent of the user based on the input search request.

[0128] According to some embodiments, in the search system shown in FIG. 13 , the recall model 1301 is further configured to convert the related search intent into a second request semantic vector and search the search resource database for at least one second data semantic vector that matches the second request semantic vector.

[0129] Thereby, in response to a user's search request, the relevant search intent of the user is obtained, and search or recommend is performed based on the obtained intent, so that the query result can further meet the user's expansion needs, thereby further improving the user experience. The potential expansion intent may be one or more.

[0130] According to some embodiments, each second data meaning vector in the at least one second data meaning vector respectively includes dimensions related to semantic relevance features and perceived relevance features of corresponding data, where the semantic relevance features are used to explain the direct meaning of the corresponding data, and the perceived relevance features emphasize dimensions related to the needs and interests of users, so as to further improve the accuracy of the recommendation system and better meet the potential needs of users.

[0131] According to some embodiments, the second request semantic vector includes dimensions related to semantic relevance features and perceived relevance features of the search request, respectively, where the semantic relevance features are used to explain the direct meaning of the corresponding data, and the perceived relevance features emphasize dimensions related to the needs and interests of users, so as to further improve the accuracy of the recommendation system and better meet the potential needs of users.

[0132] For example, the semantic relevance feature dimension in the data semantic vector corresponding to a webpage containing content introducing public figure A is used to describe the direct meaning of the webpage content. The sensed relevance feature dimension in the data semantic vector corresponding to the webpage focuses on describing the user's possible interests that can be expanded around public figure A, for example, the user may be interested in who public figure A's wife is, what works public figure A has, etc., and the sensed relevance feature dimension may include the corresponding content, which can further improve the accuracy of the recommendation system and better meet the potential needs of the user.

[0133] According to some embodiments, both the first data meaning vectors for recalling the user's search request and the second data meaning vectors for recalling the user's related search intent may be completely the same, partly the same, or completely different. That is, both the first data corresponding to the first data meaning vectors and the second data corresponding to the second data meaning vectors may be completely the same, partly the same, or completely different. One or more first data matching the user's search request are obtained by recalling from the first data, and one or more second data matching the user's related search intent are obtained by recalling from the second data.

[0134] In some embodiments, the number of the plurality of second data is smaller than the number of the plurality of first data, thereby realizing accurate recommendation, improving recommendation effect, and further improving user experience.

[0135] According to some embodiments, the search resource database includes a first sub-library and a second sub-library used only for the related search intent, and the amount of data in the second sub-library is smaller than the amount of data in the first sub-library. By providing two sub-libraries separately, the respective objectives of search and recommendation can be strengthened, the first sub-library is used to meet the mass recall requirement of related search, and the second sub-library is used to meet the exact recall requirement of meeting the user's potential requirements, thereby better improving the quality of the query result and further improving the user experience.

[0136] For example, the amount of data in the first sub-library may be on the order of 10 billion or 100 billion, which can cover many content resources and more comprehensively cover the content needs of users, and correspondingly, the amount of data in the second sub-library may be on the order of one million.

[0137] As can be appreciated, the data in the second sub-library that is applied for recommendation may be selected based on predetermined quality criteria, thereby providing higher quality recommended content to the user, better meeting the user's enhanced needs and improving the user experience.

[0138] According to some embodiments, as shown in FIG. 14, the search system 1400 based on the pre-trained language model further includes a recommendation model 1404 configured to sort the one or more second data by sequentially determining a similarity between the related search intent and one or more second data corresponding to the at least one second data semantic vector.

[0139] According to some embodiments, as shown in Figure 14, the pre-trained language model-based search system 1400 further includes a search result generation model 1405 configured to organize the one or more first data based on corresponding content understanding to generate search results, thereby simplifying the search results and more accurately meeting the needs of users.

[0140] For example, if the search request input by the user is "cuisine near tourist spot A," the multiple data include "menu of cuisine B near tourist spot A," "business hours and address of cuisine B near tourist spot A," "menu of cuisine C near tourist spot A," "address of cuisine C near tourist spot A," etc., and based on understanding of the corresponding content, the multiple data can be organized into "menu, business hours and address of cuisine B near tourist spot A," and "menu and address of cuisine C near tourist spot A" and shown to the user. This makes it possible to provide the user with search results that better match the user's needs and improve the user experience.

[0141] As can be seen, the manner of organizing one or more data is not limited to the simple superposition manner in the above example, and a suitable organizing manner can be designed based on a specific application scenario, for example, if one or more data is a text segment, semantic understanding can be performed on the one or more data, and a recommendation result can be generated based on the semantic understanding result, for example, if the text segment describes the weather conditions such as temperature, humidity, and sunlight intensity in a certain area, a recommendation result such as "a certain area is suitable for outdoor activities" can be generated based on the semantic understanding of the text segment.

[0142] According to some embodiments, as shown in Fig. 14, the search system 1400 based on the pre-trained language model further includes a recommendation result generation model 1406 configured to organize the one or more second data to generate recommendation results based on the understanding of corresponding content, so as to simplify the recommendation results and more accurately meet the needs of users.

[0143] According to some embodiments, the search system further includes a disambiguation model configured to perform a structured analysis on the search request to determine whether there is ambiguity and, in response to determining there is ambiguity, provide the user with content requiring clarification, thereby enabling clarification of the user's search request and guidance to the user's needs. Performing a structured analysis on the search request can determine whether the search request is unambiguous and unambiguous. The structured analysis may be realized, for example, based on a knowledge graph.

[0144] 16 shows a schematic diagram of the operation process of a search system based on a pre-trained language model according to an exemplary embodiment of the present disclosure. The arrows indicate the direction of the signal stream, the solid arrows indicate the operation process of online search and recommendation, and the dashed arrows indicate the operation process of building a uniform resource database.

[0145] Before describing the operation process of the search system, the construction of the database used by the search system will be described.

[0146] As shown in Fig. 16, the data in the uniform resource database may be obtained by data capture, data access, data generation, etc., where data capture may specifically be a method of automatically extracting web page data using a computer program, data access may specifically be a method of accessing an existing database, and data generation may specifically be a method of guiding the production of third-party content and accessing according to a user's search request. All the data obtained by various methods are converted into a unified data format and stored in the uniform resource database for further processing.

[0147] A unified data structure corresponding to data in a uniform resource database may be composed of, for example, a key-value tuple, where the key is an attribute of the data and the value is the value of the data. The key-value tuple may include common types and special types. The common type key-value tuples may cover attributes shared by data of different types and different sources, such as title, summary, uniform resource locator (URL), time, etc. The special type key-value tuples may cover attributes specific to data of different types and different sources, such as video type data may include a resolution attribute, etc.

[0148] Data of different types and sources are converted into a data semantic vector of unified specification by the multi-source heterogeneous data uniform semantic representation model, and then stored in the search resource database.

[0149] According to each data semantic vector, corresponding data may be obtained from the uniform resource database. For example, a unique identification number may be set for each data, and the search resource database may store a correspondence between each data semantic vector and the unique identification number of the corresponding data, and the uniform resource database may store a correspondence between each data and the corresponding unique identification number, thereby realizing the association between the uniform resource database and the search resource database, and obtaining one or more corresponding data according to at least one data semantic vector matching the search request.

[0150] According to some embodiments, a content quality score may be performed for each data in the uniform resource database by the content quality score model, and the dimension of the content quality score information of the data may be increased and stored in the uniform resource database. Correspondingly, the dimensions of information such as the delivery time, source reliability, etc. of each data may be increased and stored in the uniform resource database.

[0151] According to some embodiments, a content fine-grained segmentation model can be used to perform fine-grained segmentation on the original complete data based on semantics, and the data in the uniform resource database is the data after fine-grained segmentation. For example, a webpage text data including multiple paragraphs can be segmented into multiple text segments with different semantic features through semantic understanding, and the multiple text segments can be stored in the uniform resource database respectively. Furthermore, each text segment can be converted into a corresponding data semantic vector by a multi-source heterogeneous data uniform semantic representation model, and then stored in the search resource database.

[0152] Continuing with reference to FIG. 16, the operation process of the search system is as follows.

[0153] Step S11: In response to a search request input by a user, a structured analysis is performed on the search request using a disambiguation model to determine whether there is ambiguity, and thereby content requiring clarification is provided to the user, so as to realize clarification of the user's search request and guidance to the user's needs.

[0154] Specifically, the search request may include search information entered by the user and context information related to the search information, such as time, location, device, and user route search.

[0155] Step S12: Using the multi-source heterogeneous data uniform semantic representation model, convert the user's search request into a first request semantic vector, and using a recall model to search for at least one first data semantic vector matching the first request semantic vector in the search resource database.

[0156] Specifically, a similarity between the first request semantic vector and a data semantic vector in a search resource database is calculated to obtain at least one first data semantic vector matching the request semantic vector.

[0157] Step S13: obtaining one or more first data corresponding to the at least one first data semantic vector from a uniform resource database.

[0158] Step S14: input the one or more first data and the search request entered by the user into a sorting model, calculate a similarity score between the one or more first data and the search request, and sort the one or more first data based on the similarity score.

[0159] The one or more first data and the search request may include at least one of the following dimensions: semantic relevance, content quality, authority, etc.

[0160] Step S15: Using the search result generation model, generate search results based on one or more first data.

[0161] The search result generation model may organize the first data based on an understanding of the corresponding content to generate search results.

[0162] Step S16: Using a related search intent acquisition model, acquire the user's related search intent based on the search request.

[0163] Step S17: Using the multi-source heterogeneous data uniform semantic display model, convert the user's related search intention into a second request semantic vector, and using a recall model to search for at least one second data semantic vector matching the second request semantic vector in the search resource database.

[0164] Specifically, a similarity between the second request semantic vector and a data semantic vector in a search resource database is calculated to obtain at least one second data semantic vector matching the request semantic vector.

[0165] The recall model used for recommendations may be the same model as the recall model used for search, or it may be a different model built using different recall logic.

[0166] Step S18: obtaining one or more second data corresponding to the at least one second data semantic vector from a uniform resource database.

[0167] Step S19: input one or more second data and related search intents corresponding to the search request input by the user into a recommendation model, calculate a similarity score between the one or more second data and the related search intents, and sort the one or more second data based on the similarity score.

[0168] The sorting logic of the recommendation model may be multiple dimensions, such as semantic relevance, sensory relevance, etc. based on corresponding data, and the sensory relevance dimension places more emphasis on sensing the user's potential needs and interests, thereby generating search results that are closer to the user's needs and interests.

[0169] Step S20: generating recommendation results based on one or more second data using the recommendation result generation model.

[0170] According to another aspect of the present disclosure, there is further provided a method for training a search system based on a pre-trained language model, the search system including an end-to-end deep neural network basic model in which a recall model and a sorting model are cascaded, the recall model is based on a double-encoder structure, and the sorting model is based on a cross-encoder structure.

[0171] FIG. 17 illustrates a method for training a search system based on a pre-trained language model according to an exemplary embodiment of the present disclosure. As shown in FIG. 17, the method includes: a step S1701 of receiving a sample dataset, where the sample data in the sample dataset includes a sample search request and a first target output dataset; a step S1703 of initializing a plurality of parameters in the recall model and the sorting model, and for each sample data, converting the sample search request in the sample data into a first request semantic vector by a first encoder in the recall model; and a step S1704 of a second encoder in the recall model respectively converting a plurality of candidate data of different types into a corresponding plurality of first data semantic vectors, where the plurality of candidate data of different types at least include text, picture, and video, and the plurality of first data semantic vectors have a unified specification. The method includes a step S1702 of performing an operation including a step S1704 of selecting a first quantity of first data meaning vectors by calculating a first similarity between the first request semantic vector and the plurality of first data meaning vectors, where the first similarity between the first quantity of first data meaning vectors and the first request semantic vectors all satisfy a predetermined condition, a step S1706 of sequentially inputting candidate data corresponding to each first data meaning vector in the sample search request and the first quantity of first data meaning vectors into a cross encoder of the sorting model as a first linked input value, and sorting the candidate data corresponding to the first quantity of first data meaning vectors, a step S1707 of calculating a loss function based on the sorted candidate data and a first target output data, and a step S1708 of adjusting a plurality of parameters of the recall model and the sorting model based on the loss function. The similarity between vectors is, for example, but not limited to, a cosine similarity.

[0172] According to some embodiments, the different types of candidate data further include at least tables and knowledge graphs in addition to text, pictures and videos. As can be understood, the different types of data may further include other types of data, such as maps, videos, etc., and more types of data can further enrich the search resource database, thereby further improving the diversity of search results, better meeting user needs and improving user experience.

[0173] According to some embodiments, at least one of the text or video data in the plurality of candidate data of different types is obtained by performing fine-grained partitioning on the original complete data, which can provide a deeper understanding of the data content, realize a fine-grained index, and obtain search results that are more tailored to the needs of users.

[0174] Exemplarily, obtaining the at least one text or video data by performing fine-grained division on the original complete data may include performing fine-grained division on the original complete data according to semantics. In some embodiments, performing fine-grained division on the original complete data may include performing semantic division on the original complete data to obtain at least one text or video data. Taking webpage text data as an example, the original complete webpage text data may include a plurality of paragraphs, and each paragraph may have different semantic features, so that the data semantic vector corresponding to the corresponding complete webpage text data cannot fully represent the different semantic features of each paragraph, and the first required semantic vector embodying the needs of the user cannot be matched with the meaning of each paragraph in the search process. By dividing the original complete webpage text data, it can be divided into a plurality of segments having different semantic features, and each segment corresponds to one of the at least one text. Each segment is converted into a corresponding data semantic vector, and is matched with the first required semantic vector embodying the needs of the user in the search process, so that a search result that is more in line with the needs of the user can be obtained. Similarly, fine-grained segmentation can be performed based on video text data corresponding to the video, and the specific principles and processes are similar to those of web page text data.

[0175] According to some embodiments, each data semantic vector in the plurality of first data semantic vectors includes a dimension related to the content quality of the corresponding candidate data, which may be, but is not limited to, a content quality score of the corresponding data, so that the search system using the method training can further consider the content quality of the data and improve the quality of the search results.

[0176] According to some embodiments, each data meaning vector in the plurality of first data meaning vectors includes a dimension related to a delivery time of the corresponding candidate data. The dimension related to the delivery time of the corresponding data may be, but is not limited to, the delivery time of the corresponding data, so that the search system using the method training can further consider timeliness and improve the quality of the search results.

[0177] According to some embodiments, each data semantic vector in the plurality of first data semantic vectors includes a dimension related to the source authority of the corresponding candidate data. The dimension related to the source authority of the corresponding data may be, but is not limited to, the source website type of the corresponding data and the authority of the corresponding website type, so that the search system using the method training can further consider the authority of the data and improve the quality of the search results.

[0178] According to some embodiments, each data semantic vector in the semantic space vector includes at least two of the following dimensions of the content quality, delivery time, and source credibility of the corresponding data. In the search process, the quality of the search results can be further improved by increasing the multiple dimensions of the data.

[0179] Based on the same principle, according to some embodiments, the first request semantic vector includes context information about the user's search, the context information including at least one of time, location and the user's route search, which can further improve the search accuracy of the search system using the method training.

[0180] As can be seen, based on the context information of the user's search, the user's direct search request can be more accurately described. For example, if the search information input by the user is "what's the weather tomorrow", the first request semantic vector may correspondingly include the location where the user is located, for example, Beijing, so that the search results related to "tomorrow's weather in Beijing" can be provided to the user, which can more accurately meet the user's needs and improve the user experience. According to some embodiments, the first federated input value includes at least one of the content quality, delivery time, and source reliability of the corresponding candidate data, so that the content, such as the content quality, delivery time, and source reliability of the candidate data, can be fully taken into consideration in the sorting process, so as to obtain higher quality sorting results, and the search system trained in the method can generate search results that are more in line with the needs of users and improve the user experience.

[0181] According to some embodiments, the system further includes a recommendation model, and the sample data in the sample dataset further includes a second target output dataset, and the training method includes performing operations including: initializing a plurality of parameters in the recommendation model; for each sample data, sequentially inputting candidate data corresponding to each first data meaning vector in the sample search request and the first quantity of first data meaning vectors into a cross encoder of the recommendation model as second linked input values, thereby sorting the candidate data corresponding to each first data meaning vector of the first quantity; calculating a loss function based on the sorted candidate data and the second target output data; and adjusting a plurality of parameters in the target model and the recommendation model based on the loss function.

[0182] According to some embodiments, the second federated input value includes semantic relevance features and perceptual relevance features of the corresponding candidate data, the semantic relevance features are used to explain the direct meaning of the candidate data, and the semantic relevance features emphasize the dimension related to the user's needs and interests, so that the accuracy of the recommendation model trained by the method can be further improved and the potential needs of the user can be better met.

[0183] For example, the semantic relevance feature dimension in the data semantic vector corresponding to a webpage containing content introducing public figure A is used to describe the direct meaning of the webpage content. The sensed relevance feature dimension in the data semantic vector corresponding to the webpage focuses on describing the user's possible interests that can be expanded around public figure A, for example, the user may be interested in who public figure A's wife is, what works public figure A has, etc., and the sensed relevance feature dimension may include the corresponding content, which can further improve the accuracy of the recommendation system and better meet the potential needs of the user.

[0184] According to another aspect of the present disclosure, there is further provided an electronic device including at least one processor and a memory communicatively connected to the at least one processor, the memory storing instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the above-mentioned search method or a method for training a search system based on a pre-trained language model.

[0185] According to another aspect of the present disclosure, there is further provided a non-transitory computer-readable storage medium having stored thereon computer instructions for causing a computer to execute the above-described search method or the method for training a search system based on a pre-trained language model.

[0186] According to another aspect of the present disclosure, there is further provided a computer program product comprising a computer program which, when executed by a processor, implements the above-described method of searching or the method of training a search system based on a pre-trained language model.

[0187] Referring to FIG. 18, a structural block diagram of an electronic device 1800 usable as a server or client of the present disclosure is described, which may be an example of a hardware device applicable to each aspect of the present disclosure. The electronic device is intended to represent various types of digital electronic computing devices, such as, for example, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may represent various types of mobile devices, such as, for example, personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples, but are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0188] As shown in Fig. 18, the device 1800 may include a computing unit 1801 that can perform various suitable operations and processes according to a computer program stored in a read only memory (ROM) 1802 or loaded from a storage unit 1808 into a random access memory (RAM) 1803. The RAM 1803 may store various programs and data necessary for the operation of the device 1800. The computing unit 1801, the ROM 1802 and the RAM 1803 are connected to each other via a bus 1804. An input / output (I / O) interface 1805 may also be connected to the bus 1804.

[0189] The components of the device 1800 are connected to an I / O interface 1805, and include an input unit 1806, an output unit 18018, a storage unit 1808, and a communication unit 1809. The input unit 1806 may be any type of device capable of inputting information to the device 1800, the input unit 1806 may receive input digital or textual information and generate key signal inputs related to user settings and / or function control of the electronic device, and may include, but is not limited to, a mouse, a keyboard, a touch screen, a track pad, a track ball, a joystick, a microphone, and / or a remote control. The output unit 18018 may be any type of device capable of displaying information, and may include, but is not limited to, a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 1808 may include, but is not limited to, a magnetic disk, an optical disk. The communications unit 1809 enables the device 1800 to exchange information / data with other devices, for example, via the Internet computer network and / or various telecommunications networks, and may include, but is not limited to, a modem, a network card, an infrared communications device, a wireless communications transceiver and / or chipset, for example, a Bluetooth device, an 802.6 device, a WiFi device, a WiMax device, a cellular communications device, and / or the like.

[0190] The computing unit 1801 may be various general-purpose and / or dedicated processing components having processing and calculation capabilities. Some examples of the computing unit 1801 include, but are not limited to, a central processing unit (CPU), a graphic processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units that execute machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1801 executes various methods and processes described above, such as the above-mentioned search method or the method for training a search system based on a pre-trained language model. For example, in some embodiments, the above-mentioned search method or the method for training a search system based on a pre-trained language model may be realized as a computer software program physically included in a machine-readable medium such as the storage unit 1808. In some embodiments, some or all of the computer program may be loaded and / or installed in the device 1800 via the ROM 1802 and / or the communication unit 1809. When the computer program is loaded into the RAM 1803 and executed by the computing unit 1801, it can perform one or more steps of the above-described search method or the method for training a search system based on a pre-trained language model. Optionally, in another embodiment, the computing unit 1801 may be configured to perform the search method or the method for training a search system based on a pre-trained language model in any other suitable manner (e.g. by firmware).

[0191] Various embodiments of the systems and techniques described herein may be realized in digital electronic circuitry systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system on a chip systems (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may be embodied in one or more computer programs that may be executed and / or interpreted by a programmable system that includes at least one programmable processor, which may be a special purpose or general purpose programmable processor, and that may receive data and instructions from and transmit data and instructions to a storage system, at least one input device, and at least one output device.

[0192] The program codes for implementing the methods of the present disclosure can be programmed in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general purpose computer, a special purpose computer, or other programmable data processing apparatus, such that when the program codes are executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are performed. The program codes can be fully executed on a machine, partially executed on a machine, or as a separate software package, partly executed on a machine and partly executed on a remote machine, or entirely executed on a remote machine or server.

[0193] In the context of this disclosure, a machine-readable medium may be a tangible medium that contains or stores a program, which may be used by or in combination with an instruction execution system, device, or apparatus. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or apparatus, or any combination of the above. More specific examples of machine-readable storage media include an electrical connection by one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0194] To enable interaction with a user, the systems and techniques described herein can be implemented in a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user, and a keyboard and pointing device (e.g., a mouse or trackball) for allowing the user to provide input to the computer. Other types of devices can provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback), and input from the user can be received in any form (including acoustic input, speech input, or tactile input).

[0195] The systems and techniques described herein can be implemented in a computing system that includes a back-end component (e.g., as a data server), or a computing system that includes a middleware component (e.g., an application server), or a computing system that includes a front-end component (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with embodiments of the systems and techniques described herein), or a computing system that includes any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected via any form or medium of digital data communication (e.g., a communications network). Examples of communications networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0196] The computer system may include clients and servers. The clients and servers are typically remote from each other and interact through a communication network. The relationship of clients and servers is created by computer programs running on the corresponding computers and having a client-server relationship to each other. The server may be a cloud server, a distributed system server, or a server in combination with a blockchain.

[0197] It should be noted that, using the various forms of processes described above, steps can be rearranged, added, or removed, for example, the steps described in the present disclosure can be performed in parallel, sequentially, or in a different sequence, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved, and the present specification is not limited thereto.

[0198] It should be noted that, although the embodiments or examples of the present disclosure have been described with reference to the drawings, the above methods, systems and devices are merely illustrative embodiments or examples, and the scope of the present invention is not limited by these embodiments or examples, but is limited by the scope of the granted claims and the like. Various elements in the embodiments or examples may be omitted or replaced with other equivalent elements. Furthermore, each step may be performed according to a different order from that described in this disclosure. Furthermore, various elements in the embodiments or examples may be combined in various ways. Importantly, with the development of technology, many elements described herein may be replaced by equivalent elements appearing later in this disclosure.

Claims

1. 1. A computer-implemented search method, comprising: converting, by a processor of the computer, the user's search request into a first request semantic vector; searching, by the processor of the computer, a search resource database for at least one first data semantic vector matching the first request semantic vector, the search resource database being structured as a semantic vector space, in which different types of data are transformed into corresponding data semantic vectors, the different types of data including at least text, picture, and video, and each data semantic vector in the semantic vector space including a dimension related to a content quality of the corresponding data; generating, by the processor of the computer, search results based on the at least one first data semantic vector. How to search.

2. Obtaining a related search intent of the user based on the search request of the user; converting the related search intent into a second request semantic vector; searching said search resource database for at least one second data semantic vector matching said second request semantic vector; generating recommendation results based on the at least one second data semantic vector. The search method according to claim 1 .

3. The method of claim 2 , wherein obtaining the user's related search intent comprises obtaining the user's related search intent based on a needs graph.

4. The search method according to claim 2 , wherein obtaining the user's related search intent comprises obtaining the user's related search intent based on a search history of the user in a same session process.

5. 3. The search method of claim 2, wherein the search resource database includes a first sub-library and a second sub-library used only for the related search intent, and the amount of data in the second sub-library is smaller than the amount of data in the first sub-library.

6. 6. The method of claim 5, wherein the data in the second sub-library is selected based on predetermined quality criteria.

7. generating the search results obtaining one or more first data corresponding to the at least one first data semantic vector; Sorting the one or more first data based on a first sorting logic based on at least one of a content quality, a delivery time, and a source reliability of the corresponding data. The search method according to claim 2.

8. The generating of the recommendation result includes: obtaining one or more second data corresponding to the at least one second data semantic vector; and sorting the one or more second data based on a second sorting logic different from the first sorting logic based on semantic relevance features and perceptual relevance features of the corresponding data. The search method according to claim 7.

9. generating the search results The method of claim 7 , further comprising: organizing the one or more first data to generate the search results based on an understanding of the corresponding content.

10. performing a structured analysis of the search request to determine whether there is ambiguity; and in response to determining that there is an ambiguity, providing the content requiring clarification to the user. The search method according to claim 1 .

11. The search method according to claim 1 , wherein each data semantic vector in the semantic vector space includes a dimension relating to a delivery time of the corresponding data.

12. The method of claim 1 , wherein each data semantic vector in the semantic vector space includes a dimension related to a source authority of the corresponding data.

13. The method of claim 2 , wherein each second data meaning vector in the at least one second data meaning vector includes dimensions related to semantic relevance features and perceived relevance features of the corresponding data, respectively.

14. The method of claim 1 , wherein the first request semantic vector includes contextual information about the user's search including at least one of a time, a location, a terminal device, and a direction of the user's search.

15. The method of claim 2 , wherein the second request semantic vector includes dimensions related to semantic and perceived relevance features of the search request, respectively.

16. The method of claim 1 , wherein the different types of data further include at least a table and a knowledge graph.

17. The method of claim 1 , wherein at least one of the text or video data in the different types of data is obtained by performing fine-grained division on the original complete data.

18. The search method according to claim 1 , wherein the data semantic vectors converted from the different types of data have a uniform specification.

19. a unit for converting a user's search request into a first request semantic vector; a unit for searching at least one first data semantic vector matching the first request semantic vector in a search resource database, the search resource database being structured as a semantic vector space, in which different types of data are transformed into corresponding data semantic vectors, the different types of data at least including text, picture and video, and each data semantic vector in the semantic vector space includes a dimension related to content quality of the corresponding data; and a unit for generating search results based on the at least one first data semantic vector. Search device.

20. A unit for obtaining a related search intent of the user according to a search request of the user; a unit for converting the related search intent into a second request semantic vector; a unit for searching said search resource database for at least one second data semantic vector matching said second request semantic vector; and a unit for generating a recommendation result based on the at least one second data semantic vector.

20. The search device according to claim 19.

21. The apparatus of claim 20 , wherein the unit for obtaining a related search intent of the user based on the user's search request is further configured to obtain a related search intent of the user based on a needs graph.

22. The search device of claim 20, wherein the unit for obtaining the user's related search intent based on the user's search request is further configured to obtain the user's related search intent based on the user's search history in the same session process.

23. The search device of claim 20, wherein the search resource database includes a first sub-library and a second sub-library used only for the related search intent, and the amount of data in the second sub-library is smaller than the amount of data in the first sub-library.

24. 24. The apparatus of claim 23, wherein the data in the second sub-library is selected based on predetermined quality criteria.

25. The unit for generating search results based on the at least one first data semantic vector comprises: a unit for obtaining one or more first data corresponding to the at least one first data semantic vector; and a unit for sorting the one or more first data based on a first sorting logic based on at least one of a content quality, a delivery time, and a source reliability of the corresponding data. The search device according to claim 20.

26. The unit for generating a recommendation result based on the at least one second data semantic vector includes: a unit for obtaining one or more second data corresponding to the at least one second data semantic vector; and a unit for sorting the one or more second data based on a second sorting logic different from the first sorting logic based on semantic relevance features and perceptual relevance features of corresponding data.

26. The search device according to claim 25.

27. The unit for generating search results based on the at least one first data semantic vector comprises: and a unit for organizing the one or more first data to generate the search results based on an understanding of the corresponding content.

26. The search device according to claim 25.

28. a unit for performing a structured analysis on the search request to determine whether there is an ambiguity; and a unit for providing content requiring clarification to the user in response to determining that there is an ambiguity.

20. The search device according to claim 19.

29. The search apparatus according to claim 19 , wherein each data semantic vector in the semantic vector space includes a dimension relating to a delivery time of the corresponding data.

30. 20. The search apparatus of claim 19, wherein each data semantic vector in the semantic vector space includes a dimension related to a source authority of the corresponding data.

31. 21. The search apparatus of claim 20, wherein each second data meaning vector in the at least one second data meaning vector includes dimensions relating to semantic relevance features and perceived relevance features of the corresponding data, respectively.

32. The search apparatus of claim 20 , wherein the second request semantic vector includes dimensions related to semantic and perceived relevance features of the search request, respectively.

33. 20. The apparatus of claim 19, wherein the first request semantic vector includes contextual information about the user's search including at least one of a time, a location, a terminal device, and a direction of the user's search.

34. The search apparatus of claim 19 , wherein the different types of data further include at least a table and a knowledge graph.

35. 20. The apparatus of claim 19, wherein at least one of the text or video data in the different types of data is obtained by performing fine-grained partitioning on the original complete data.

36. The search apparatus according to claim 19 , wherein the data semantic vectors converted from the different types of data have a uniform specification.

37. A search system based on a pre-trained language model, comprising: a recall model configured to convert a user's search request into a first request semantic vector and search a search resource database for at least one first data semantic vector matching the first request semantic vector, the search resource database being structured as a semantic vector space, in which different types of data are converted into corresponding data semantic vectors, the different types of data including at least text, picture, and video, and each data semantic vector in the semantic vector space including a dimension related to content quality of the corresponding data; a sorting model configured to sort the one or more first data by sequentially determining a similarity between the search request and the one or more first data corresponding to the at least one first data semantic vector; The recall model and the sorting model are cascaded to form an end-to-end deep neural network base model. Search system.

38. 40. The search system of claim 37, further comprising a related search intent acquisition model configured to acquire related search intent of the user based on the input search request.

39. The recall model further comprises: Transforming the related search intent into a second request semantic vector; configured to search the search resource database for at least one second data semantic vector matching the second request semantic vector; 39. The search system according to claim 38.

40. 39. The search system of claim 38, wherein the search resource database includes a first sub-library and a second sub-library used only for the related search intent, and the amount of data in the second sub-library is smaller than the amount of data in the first sub-library.

41. 40. The search system of claim 39, further comprising a recommendation model configured to sort the one or more second data by sequentially determining a similarity between the related search intent and the one or more second data corresponding to the at least one second data semantic vector.

42. 40. The search system of claim 37, further comprising a search result generation model configured to organize the one or more first data to generate search results based on an understanding of the corresponding content.

43. The search system of claim 41 , further comprising a recommendation result generation model configured to organize the one or more second data to generate recommendation results based on an understanding of the corresponding content.

44. performing a structured analysis of the search request to determine whether there is any ambiguity; and a disambiguation model configured to, in response to determining that there is an ambiguity, provide content requiring clarification to the user.

38. A search system according to claim 37.

45. 38. The search system of claim 37, wherein each data semantic vector in the semantic vector space includes a dimension relating to a delivery time of the corresponding data.

46. 38. The search system of claim 37, wherein each data semantic vector in the semantic vector space includes a dimension related to a source authority of the corresponding data.

47. 40. The search system of claim 39, wherein each second data meaning vector in the at least one second data meaning vector includes dimensions relating to semantic relevance features and perceived relevance features of the corresponding data, respectively.

48. 38. The system of claim 37, wherein the first request semantic vector includes contextual information about the user's search including at least one of time, location, terminal device, and the user's itinerary.

49. 40. The search system of claim 39, wherein the second request semantic vector includes dimensions related to semantic and perceived relevance features of the search request, respectively.

50. 38. The search system of claim 37, wherein the different types of data further include at least a table and a knowledge graph.

51. 38. The search system of claim 37, wherein at least one of the text or video data in the different types of data is obtained by performing fine-grained partitioning on the original complete data.

52. 38. The search system of claim 37, wherein the data semantic vectors converted from the different types of data have a uniform specification.

53. At least one processor; a memory communicatively connected to the at least one processor; the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor such that the at least one processor can execute the search method according to any one of claims 1 to 18; electronic equipment.

54. A non-transitory computer-readable storage medium storing a computer program, the computer program being configured to enable a computer to execute a search method according to any one of claims 1 to 18.

55. A computer program product which, when executed by a processor, implements the search method according to any one of claims 1 to 18.

Images