Resource retrieval method, resource retrieval device, medium, product and electronic equipment

Abstract

The invention relates to the technical field of computers, and provides a resource retrieval method, a resource retrieval device, a medium, a product and electronic equipment. The method comprises the steps that in response to a first input operation on a client side, under the condition that first input content of the first input operation has a resource retrieval intention, a retrieval type indicated by the first input operation is determined, and the retrieval type comprises precise retrieval and fuzzy retrieval; and according to the retrieval type, retrieving a target resource corresponding to the first input operation in a resource library, and pushing the target resource to the client. According to the scheme, the accuracy of resource retrieval can be improved.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technologies, and in particular, to a resource retrieval method, a resource retrieval device, a computer-readable storage medium, a computer program product, and an electronic device. Background Art

[0002] This section aims to provide background or context for the embodiments of the present disclosure recited in the claims. The description herein is not admitted to be prior art merely because it is included in this section.

[0003] Resource retrieval can be understood as a process of searching for resources in a resource library. Taking music search as an example, a user can input retrieval content in a music client, such as inputting "songs that can make people happy". The music client can search for songs that meet the user's needs in the music library according to the input content of the user, and return them to the music client where the user is located.

[0004] The resource retrieval methods in the related art mainly rely on keyword matching, that is, relevant resources are matched through keywords, so as to perform retrieval and recommendation. Summary of the Invention

[0005] However, in the related art, the input of the user cannot be subjected to emotion recognition, and when the user performs fuzzy retrieval, it is difficult to match accurate relevant resources, resulting in low accuracy of resource retrieval.

[0006] Therefore, there is a great need for a resource retrieval method to improve the accuracy of the same resource retrieval.

[0007] In this context, the embodiments of the present disclosure are expected to provide a resource retrieval method, a resource retrieval device, a computer-readable storage medium, a computer program product, and an electronic device.

[0008] According to a first aspect of an embodiment of the present disclosure, there is provided a resource retrieval method, including: in response to a first input operation on a client, when there is a resource retrieval intention in a first input content of the first input operation, determining a retrieval type indicated by the first input operation, where the retrieval type includes precise retrieval and fuzzy retrieval; according to the retrieval type, retrieving a target resource corresponding to the first input operation in a resource library, and pushing the target resource to the client.

[0009] According to a second aspect of the embodiments of the present disclosure, a resource retrieval device is provided, including: a retrieval type determination module configured to, in response to a first input operation on a client, determine the retrieval type indicated by the first input operation when there is a resource retrieval intention in the first input content of the first input operation, where the retrieval type includes precise retrieval and fuzzy retrieval; a retrieval module configured to retrieve, according to the retrieval type, a target resource corresponding to the first input operation from a resource library and push the target resource to the client.

[0010] According to a third aspect of the present disclosure, a computer program product including instructions is provided, which, when running on a computer, causes the computer to execute the steps of the resource retrieval method as in the first aspect.

[0011] According to a fourth aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, it implements the resource retrieval method as described in the first aspect of the above embodiments.

[0012] According to a fifth aspect of the embodiments of the present disclosure, an electronic device is provided, including: a processor; and a storage device for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the resource retrieval method as described in the first aspect of the above embodiments.

[0013] According to the resource retrieval method, resource retrieval device, computer-readable storage medium, computer program product, and electronic device of the embodiments of the present disclosure, through the analysis of the user's retrieval intention and retrieval type, resource retrieval is assisted according to the user's retrieval intention and retrieval type. Compared with the related art, on the one hand, the present disclosure can perform different retrieval processes according to the retrieval intention and retrieval type through the identification and analysis of the retrieval intention and retrieval type, so as to provide personalized retrieval resource recommendations for users and improve the accuracy of resource retrieval; on the other hand, the present disclosure can process precise retrieval and fuzzy retrieval simultaneously through the identification and analysis of the retrieval type, thereby improving the applicability of the resource retrieval method. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features, and advantages of the exemplary embodiments of the present disclosure will become readily understandable. In the drawings, several embodiments of the present disclosure are shown in an exemplary rather than restrictive manner, where:

[0015] Figure 1 A schematic diagram showing an exemplary system architecture to which the embodiments of the present disclosure can be applied is shown;

[0016] Figure 2Schematic flowchart of a resource retrieval method in an exemplary embodiment of the present disclosure;

[0017] Figure 3 Schematic flowchart of a method for determining the above-mentioned pre-trained target language model in an exemplary embodiment of the present disclosure;

[0018] Figure 4 Schematic flowchart of a method for retrieving a target resource from a resource library in the case where the retrieval type is fuzzy retrieval in an exemplary embodiment of the present disclosure;

[0019] Figure 5 Schematic flowchart of a method for retrieving a target resource from a resource library according to first retrieval content in an exemplary embodiment of the present disclosure;

[0020] Figure 6 Schematic flowchart of a method for obtaining a first number of target resources in an exemplary embodiment of the present disclosure;

[0021] Figure 7 Schematic flowchart of a method for determining a pre-trained resource identification prediction model in an exemplary embodiment of the present disclosure;

[0022] Figure 8 Schematic flowchart of a method for resource retrieval according to an input image in an exemplary embodiment of the present disclosure;

[0023] Figure 9 Schematic diagram of the composition of a resource retrieval device in an exemplary embodiment of the present disclosure;

[0024] Figure 10 Schematic diagram of the structure of an electronic device in an exemplary embodiment of the present disclosure.

[0025] In the drawings, the same or corresponding reference numerals indicate the same or corresponding parts. Detailed implementation manners

[0026] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and then implement the present disclosure, and do not limit the scope of the present disclosure in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to be able to fully convey the scope of the present disclosure to those skilled in the art.

[0027] Those skilled in the art will appreciate that the embodiments of the present disclosure may be implemented as a system, device, equipment, medium, method or computer program product. Therefore, the present disclosure may be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0028] According to an embodiment of the present disclosure, a resource retrieval method, a resource retrieval device, a computer-readable storage medium, a computer program product, and an electronic device are provided.

[0029] In this document, any number of elements in the drawings is used for illustration rather than limitation, and any naming is used only for distinction and does not have any limiting meaning.

[0030] The principle and spirit of the present disclosure are explained in detail below with reference to several representative embodiments of the present disclosure.

[0031] Overview of Application Scenarios

[0032] It should be noted that the following application scenarios are only shown to facilitate understanding of the spirit and principle of the present disclosure, and the embodiments of the present disclosure are not limited in this respect. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario.

[0033] In an exemplary application scenario, the resource retrieval method disclosed in the present invention can be used to search for songs in a music client. For example, when a user enters "I feel a little down today, and I want to listen to some songs that can cheer me up" in a music client, the language model can refer to the user's previous search history (such as records of searching for "relaxing music") to complete the sentence to "I feel down today, and I want to listen to some songs that can cheer me up and improve my mood." The language model then determines the search intent of the completed content, identifies the search intent for songs, and determines that the song search intent is a fuzzy retrieval type. The language model then rewrites the completed sentence in natural language, for example, rewriting it to "songs with positive emotional labels and brisk rhythms." Then, the rewritten content is segmented, and each segmented word with practical meaning is matched with the label value of each label in the music library label library. According to the matching results, the search term corresponding to each segmented word is obtained. For example, the segmented word "fast rhythm" is converted into a semantic vector and matched with the semantic vector of each label in the music library label library. It is determined that the search term corresponding to the segmented word "fast rhythm" is "cheerful". Then, songs with the rhythm label of "cheerful" are searched in the music library, and songs with the label of "cheerful" are sorted according to the sorting model. The songs that best meet the user's needs are recommended to the user. At the same time, the intelligent reply technology is used to generate guiding text copy, such as "These songs are full of positive energy. I hope they can help you get rid of your low mood and cheer up again."

[0034] Exemplary System Architecture

[0035] First, refer to Figure 1 to describe the system architecture of the exemplary application environment of the present disclosure.

[0036] As Figure 1 shown, the system architecture 100 may include a terminal device 110 and a server 120. Among them, the terminal device 110 may be a terminal device such as a smart phone, a tablet computer, a desktop computer, a laptop computer, a smart wearable device, etc. The server 120 generally refers to a background system that provides services related to the resource retrieval method in this exemplary embodiment, and may be a single server or a cluster formed by multiple servers. A connection may be formed between the terminal device 110 and the server 120 through a wired or wireless communication link for data interaction.

[0037] In an exemplary embodiment, the above-mentioned resource retrieval method may be executed by the server 120. Correspondingly, a resource retrieval device may be provided in the server 120 to implement the corresponding module functions. For example, when a user inputs a first input content through a first input operation in the terminal device 110, the server 120 will perform an analysis of the resource retrieval intention on the first input content to determine whether the user wants to perform a resource retrieval. If it is determined that the user wants to perform a resource retrieval, the server 120 will continue to determine whether the user wants to perform an accurate retrieval or a fuzzy retrieval, and thus, according to different retrieval types, search for a target resource corresponding to the first input content input by the first input operation in the resource library, and return the found target resource to the terminal device 110.

[0038] It should be understood that Figure 1 the numbers of the terminal devices and servers in

[0039] are merely illustrative. According to the implementation requirements, there may be any number of terminal devices and servers. For example, the server 120 may be an independent physical server, or a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

[0040] Exemplary Method

[0041] Figure 2 shows a schematic flowchart of a resource retrieval method in an exemplary embodiment of the present disclosure. Refer to Figure 2, the method includes:

[0042] Step S210, in response to a first input operation on the client, when there is a resource retrieval intention in the first input content of the first input operation, determine the retrieval type indicated by the first input operation, where the retrieval type includes precise retrieval and fuzzy retrieval;

[0043] Step S220, according to the retrieval type, retrieve the target resource corresponding to the first input operation in the resource library, and push the target resource to the client.

[0044] Next, a detailed description will be given of the specific implementation manner of "Step S210, in response to a first input operation on the client, when there is a resource retrieval intention in the first input content of the first input operation, determine the retrieval type indicated by the first input operation".

[0045] In an exemplary implementation manner, a resource can be understood as information that can be obtained through a network application, including at least one of text information, picture information, video information, and audio information. For example, if the network application is an e-commerce application, the resources can be various merchants and various products. If the network application is a video application, the resources can be various videos. If the network application is a live broadcast application, the resources can be various live broadcast rooms. If the network application is a social application, the resources can be various news and information, friend dynamics, etc. If the network application is a music application, the resources can be various types of music, such as songs, podcasts, audiobooks, etc.

[0046] In an exemplary implementation manner, the first input content can include one or more of text input content, voice input content, and image input content. That is to say, the first input content can be input in the form of text, can be input in the form of voice, can also be input in the form of an image, or can be any combination form of text, voice, and image. This exemplary implementation manner does not make special limitations on this.

[0047] For example, a user can input the content that they want to understand through the client on various clients. Taking a music client as an example, the user can input information related to the songs they want to listen to on the music client, or can input music theory questions on the music client. The music client can search, recommend, and reply to relevant information according to the user's needs. If the user is performing a song search, the songs searched in the music library according to the content input by the user will be returned to the client where the user is located. If the user is not performing a song search, but is using the music client to understand other information, such as music theory knowledge, recent music activities of a certain musician, etc., there is no need to perform a song search, but only answer the content input by the user.

[0048] In the present disclosure, the resource retrieval intention of the user's input content can be determined first. When it is determined that the user has the resource retrieval intention, the retrieval of relevant resources is then carried out. By judging the retrieval intention, the user's intention can be accurately analyzed. Thus, when it is determined that there is a resource retrieval intention, the retrieval focus can be concentrated on resource retrieval. Taking music retrieval as an example, for instance, if the user inputs "I want to listen to the songs of Artist A. By the way, do you know what activities he has recently?", it can be judged that the first half of this input content is the intention to search for songs, and the second half is a query intention unrelated to song search. That is, there is a song search intention, and the search focus can be concentrated on song finding, thereby assisting in improving the accuracy of retrieval.

[0049] Exemplarily, Figure 2 A flowchart showing a method for determining whether there is a resource retrieval intention in an exemplary embodiment of the present disclosure is shown. Refer to Figure 2 This method may include steps S210 to S260. Among them:

[0050] In step S210, in response to a first input operation on the client, the first input content of the first input operation is input into a pre-trained target language model, and it is determined whether the first input content needs to be associated with context according to the target language model.

[0051] For example, the first input operation may include a confirmation operation after the user finishes inputting the relevant content to be searched on the client. The confirmation operation may include an operation of clicking a search control in the client, etc. This exemplary embodiment does not make special limitations on this.

[0052] After the user completes the first input operation on the client, the content of the first input operation can be input into the pre-trained target language model, and it is determined whether the first input content needs to be associated with context through the pre-trained target language model. When the target language model determines that the first input content does not need to be associated with context, the target language model can directly determine whether there is a resource retrieval intention for the first input content according to the first input content, that is, the target language model can directly judge whether there is a resource retrieval intention for the input first input content; when the target language model determines that the first input content needs to be associated with context, steps S220 to S260 can be performed.

[0053] In step S220, when the target language model determines that the first input content needs to be associated with context, the most recent historical retrieval communication content is obtained.

[0054] In an exemplary embodiment, the most recent historical retrieval communication content may include the most recent input content or the reply content to the most recent input content.

[0055] For example, in the present disclosure, a user can perform resource retrieval in the client, or can have a question-and-answer interaction with the client. When the user has a question-and-answer interaction with the client, the client generates a response content according to the content input by the user. When the user performs resource retrieval in the client, relevant resources are retrieved from the resource library according to the retrieval conditions input by the user and returned to the client where the user is located. In other words, when the content of the most recent input has a response content, that is, the content of the most recent input includes question-and-answer content, the most recent historical retrieval communication content is the response content. When the content of the most recent input does not have a response content, that is, the content of the most recent input includes retrieval conditions, the most recent historical retrieval communication content is the retrieval content of the most recent input.

[0056] Taking a music client as an example, when the user inputs "relaxing music after work" in the client, at this time, no response content will be generated, but the resource retrieval result corresponding to the content input by the user will be determined and returned to the client where the user is located. When the user then inputs "a bit tired and want to listen to music" in the client, "relaxing music after work" can be used as the most recent historical retrieval communication content for "a bit tired and want to listen to music". For another example, when the user inputs "Does singer A have any recent activities?" in the client and a response content "Singer A has a concert recently, which will be held in city B, and the concert theme is XXX" is generated for this content, and then the user inputs "a bit tired and want to listen to music" in the client, then "Singer A has a concert recently, which will be held in city B, and the concert theme is XXX" can be used as the most recent historical retrieval communication content.

[0057] In step S230, the first input content and the most recent historical retrieval communication content are input into the target language model, and based on the target language model, the first input content is complemented by the most recent historical retrieval communication content to obtain a first complemented content.

[0058] For example, the first input content and the most recent historical retrieval communication content can be concatenated to obtain a concatenated content, and then the concatenated content is input into the target language model. The target language model will complement the first input content according to the most recent historical retrieval communication content in the concatenated content, so as to obtain a first complemented content.

[0059] Taking the above example where the user inputs "a bit tired and want to listen to music" and uses "relaxing music after work" as the most recent historical retrieval communication content for "a bit tired and want to listen to music", a first complemented content "Feeling a bit tired after work and want to listen to some music that can help relax the body and mind" can be generated through the target language model.

[0060] In step S240, input the first completion content into the target language model, and determine whether the first completion content has a resource retrieval intention through the target language model. If so, go to step S250; otherwise, go to step S260.

[0061] In step S250, it is determined that the first input content has a resource retrieval intention.

[0062] In step S260, it is determined that the first input content does not have a resource retrieval intention.

[0063] For example, the first completion content can be input into the target language model, and the target language model can perform in-depth semantic analysis on the completed content to determine whether the user has a resource retrieval intention. When the target language model determines that the first completion content has a resource retrieval intention, it is determined that the first input content input by the user has a resource retrieval intention. When the target language model determines that the first completion content does not have a resource retrieval intention, it is determined that the first input content input by the user does not have a resource retrieval intention.

[0064] In the present disclosure, the powerful semantic understanding ability of the language model can be utilized, and the natural sentence input by the user can be completed in combination with the user's search history record and the current context interaction information. Through the completion operation, the potential intention of the user can be captured more comprehensively, and the accuracy of the retrieval can be improved.

[0065] In an exemplary implementation manner, when it is determined that the first input content has a resource retrieval intention, the retrieval type of the first input content can be further determined. Among them, the retrieval types include precise retrieval and fuzzy retrieval.

[0066] In an exemplary implementation manner, precise retrieval can be understood as a retrieval where each retrieval condition is a specific named entity, and fuzzy retrieval can be understood as a retrieval where the retrieval condition includes a non-specific named entity.

[0067] Taking a music client as an example, for example, if the first input content is "Play song C by singer A" or "Play song C", both singer A and song C are specific named entities, then the retrieval type of this first input content is precise retrieval; for another example, if the first input content is "Quiet songs suitable for listening to alone at night", and the input content includes a retrieval condition that is not a specific named entity, then the retrieval type of this first input content is fuzzy retrieval; for still another example, if the first input content is "I want to listen to Cantonese songs by singer A", although there is a specific named entity singer A in this first input content, there is also a retrieval condition "Cantonese songs" that is not a specific named entity, and it can also be understood as fuzzy retrieval.

[0068] In an exemplary embodiment, determining the retrieval type indicated by the first input operation includes: when the target language model determines that the first completion content has a resource retrieval intention, performing named entity analysis on the first completion content through the target language model, and determining the retrieval type of the first input operation indication according to the named entity analysis result.

[0069] For example, when training the target language model, input content of corresponding precise retrieval types can be generated according to different named entities, and the input content of the fuzzy retrieval type and the input content of the precise retrieval type are used as training data. In this way, the pre-trained target language model can perform named entity analysis on the retrieval content, determine whether all the retrieval conditions corresponding to the retrieval content are named entities, and when it is determined that all the retrieval conditions are named entities, determine that the retrieval type is precise retrieval; otherwise, determine that the retrieval type is fuzzy retrieval.

[0070] In the present disclosure, the pre-trained target language model can perform named entity recognition analysis on the input retrieval content or the completion content corresponding to the input retrieval content, and distinguish whether the user's retrieval intention is precise entity retrieval or general semantic fuzzy retrieval according to the result of the named entity recognition analysis. Through named entity recognition, the key named entities in the input statement can be accurately recognized. Taking a music client as an example, named entities such as song names and singer names can be recognized, so as to judge the retrieval type.

[0071] Exemplarily, Figure 3 The flowchart shows a method for determining the above-mentioned pre-trained target language model in an exemplary embodiment of the present disclosure. Refer to Figure 3 This method may include steps S310 to S340. Among them:

[0072] In step S310, input the preset prompt word into the existing first language model, and through the preset prompt word, enable the first language model to determine whether the second input content input into the first language model needs to be completed, the corresponding second completion content in the case of needing to be completed, whether the second completion content has a retrieval intention, the retrieval type of the second completion content in the case of having a retrieval intention, and the corresponding second retrieval content after rewriting the second completion content in the case that the retrieval type of the second completion content is fuzzy retrieval.

[0073] For example, for a specific resource retrieval field, such as song retrieval in a music client, preset prompt words can be generated. Taking the music client as an example, each preset prompt word includes the current second input content, whether the current second input content needs to be completed, the corresponding second completion content in the case of needing to be completed, whether the second completion content has a resource retrieval intention, whether the retrieval type of the second completion content is precise retrieval or fuzzy retrieval when the second completion content has a retrieval intention, the corresponding second retrieval content after rewriting the second completion content when the retrieval type of the second completion content is fuzzy retrieval, whether the second input content has a resource retrieval intention in the case of not needing to be completed, whether the retrieval type of the second input content is precise retrieval or fuzzy retrieval when the second input content has a resource retrieval intention, and the description information of the corresponding second retrieval content after rewriting the second input content when the retrieval type of the second input content is fuzzy retrieval.

[0074] In other words, through the preset prompt words, the first language model can learn how to determine whether the second input content input into the first language model needs to be completed, what the completed content is, whether the second input content or the completed content has a resource retrieval intention, whether the resource retrieval type is fuzzy retrieval or precise retrieval when there is a resource retrieval intention, and how to rewrite the second input content or the completed content into a retrieval content suitable for resource library retrieval when the resource retrieval type is fuzzy retrieval.

[0075] It should be noted that in order to enable the first language model to have the corresponding capabilities, the preset prompt words should include different types of preset prompt words such as the second input content that needs to be completed, the second input content that does not need to be completed, the second input content with a retrieval intention, the second input content without a retrieval intention, the second input content with precise retrieval, and the second input content with fuzzy retrieval.

[0076] In an exemplary implementation manner, the first language model includes an existing pre-trained large language model.

[0077] In step S320, after the first language model finishes processing the preset prompt words, a second language model is obtained according to the first language model that has processed the preset prompt words.

[0078] For example, after inputting a preset number of preset prompt words into the first language model in sequence, or when it is considered according to experience that the first language model already has the above capabilities of determining whether the input content needs to be completed, what the completed content is, etc., the first language model with such capabilities can be used as the second language model.

[0079] Taking a music client as an example, for instance, the first preset prompt is "The current input content is a song suitable for listening to while driving to work. It is the first input content, does not need to be completed in connection with the context, has a song search intention, and is of the fuzzy search type. It can be rewritten as the search content 'driving commute music'". The second preset prompt is "The current input content is more soothing in rhythm. Its input time interval from the previous input content is less than 1 minute and there is no emotional conflict. It needs to be completed in connection with the context. The completed content is music that is more soothing in rhythm and suitable for listening to while driving to work. It has a song search intention and is of the fuzzy search type. The completed content can be rewritten as the search content 'driving relaxation music'". The third preset prompt is "The current input content is not what I want to listen to at noon. Its input time interval from the previous input content is less than 1 minute and there is no emotional conflict. It needs to be completed in connection with the context. The completed content is non-Chinese music that is more soothing in rhythm and suitable for listening to while driving to work. It has a song search intention and is of the fuzzy search type. The corresponding completed content can be rewritten as the search content 'driving relaxation English songs'".

[0080] In step S330, obtain an input example for the resource library, input the input example for the resource library into the second language model, and obtain the label information corresponding to the input example according to the second language model.

[0081] For example, multiple input examples for the resource library can be pre-generated. Taking a music client as an example, the search needs of users can be simulated, and multiple input examples can be pre-generated, such as "I want to listen to songs by singer A", "I want to listen to song B by singer A", "Music for relaxation after work", "Music suitable for studying", "Music that makes people happy", "Play song D", and so on.

[0082] Exemplarily, the input example can be input into the second language model, and the label information corresponding to the input example can be obtained according to the output of the second language model.

[0083] In an exemplary implementation manner, the label information includes whether the input example needs to be completed, the third completed content corresponding to the input example in the case of needing to be completed, whether the third completed content has a search intention, in the case of not needing to be completed, whether the input example has a search intention, in the case of having a search intention, the search type of the third completed content or the input example, in the case of not having a search intention, the reply content corresponding to the third completed content or the input example, and the third search content corresponding to the rewritten third completed content or the input example in the case where the search type of the third completed content or the input example is a fuzzy search.

[0084] In step S340, the initial third - language model is trained according to the input examples and the label information, and the target language model is obtained according to the training results.

[0085] In an exemplary implementation, the number of model parameters of the first - language model is more than that of the third - language model. For example, the first - language model is a large - language model, and the third - language model can be an open - source small - language model.

[0086] The input examples can be used as training data, and the output of the second - language model for the input examples can be used as the labels of the input examples, thereby generating training data. Based on this training data, the initial third - language model is fine - tuned, and the above - mentioned pre - trained target language model is obtained according to the training results.

[0087] For example, in the present disclosure, through prompt engineering, the large - language model is enabled to have relevant capabilities, and then the large - language model with relevant capabilities is used to generate training data in a specific domain. Based on the training data in the specific domain, a small pre - trained language model is fine - tuned to obtain a target language model suitable for retrieval processing and knowledge answering in the specific domain.

[0088] In the present disclosure, through prompt engineering, the input - output structure of the language model can be streamlined, enabling the language model to better process natural language related to specific domains, such as music retrieval domain, video retrieval domain, etc., ensuring that the language model can accurately give resource search results or reply results related to the specific domain, and at the same time improving the response speed of the language model.

[0089] Next, a detailed description of the specific implementation of "step S220: According to the retrieval type, retrieve the target resource corresponding to the first input operation in the resource library and push the target resource to the client" will be given.

[0090] Related resource retrieval methods focus on resource retrieval through keywords, and their ability to parse natural language is limited, making it difficult to meet the user's need to find resources through dialogue, such as finding songs through dialogue. In the present disclosure, by identifying the retrieval type and performing different processing on different retrieval types, keyword retrieval and natural - language understanding and parsing retrieval are combined to support the user's resource retrieval needs in complex situations, thereby improving the efficiency and accuracy of resource retrieval and helping the user quickly and accurately locate the required target resource.

[0091] Next, in combination with Figure 4 An exemplary implementation of step S220 will be described. Exemplarily, Figure 4The flowchart shows a method for retrieving a target resource in a resource library when the retrieval type is fuzzy retrieval in an exemplary embodiment of the present disclosure. Refer to Figure 4 The method may include steps S410 to S420. Among them:

[0092] In step S410, when the retrieval type is the fuzzy retrieval, the first completion content is rewritten by the target language model, and the first completion content is rewritten into a first retrieval content related to the resource library.

[0093] For example, as described above, during the process of training the target language model, training and learning are carried out on how to rewrite the first completion content. Therefore, the target language model has the ability to rewrite the first completion content, and the target language model can rewrite the first completion content into a first retrieval content suitable for resource library retrieval.

[0094] For example, continuing with the music client as an example, the target language model can rewrite the input "songs that can make people happy" into "songs with a cheerful mood label". Through rewriting, the fuzzy user requirements can be converted into precise query statements that the music resource library can effectively process, thereby improving the accuracy of song recall.

[0095] Similarly, when it is determined that the first input content does not need to be completed, the target language model can directly determine whether the first input content has a resource retrieval intention. When it is determined that there is a resource retrieval intention, the target language model can directly perform named entity analysis on the first input content to determine whether the retrieval type of the first input content is fuzzy retrieval or precise retrieval. When it is determined that the first input content is fuzzy retrieval, the first input content can be rewritten into a query suitable for the resource library to perform retrieval (i.e., the above-mentioned first retrieval content, also referred to as the first query content).

[0096] In step S420, according to the first retrieval content, the target resource corresponding to the first input operation is retrieved in the resource library.

[0097] Next, an exemplary implementation of step S420 will be described in conjunction with Figure 5 An exemplary flowchart shows a method for retrieving a target resource in a resource library according to the first retrieval content in an exemplary embodiment of the present disclosure. Refer to Figure 5 The method may include steps S510 to S540. Among them: Figure 5 In step S510, the first retrieval content is segmented to obtain at least one segment.

[0098]

[0099] For example, the first retrieved content can be segmented to obtain at least one meaningful segment, ensuring the accurate basic deconstruction of the user's query statement. For example, for the query "Cantonese love songs in the 1990s", it can be split into three segments: "1990s", "Cantonese", and "love songs".

[0100] In step S520, for each segment, the segment is respectively matched with the tag values corresponding to each type of resource tag in the resource tag library corresponding to the resource library. If the match is successful, proceed to steps S530 to S550; otherwise, proceed to steps S560 to S570.

[0101] In an exemplary embodiment, the resources in the resource library have multiple tags, and these multiple tags respectively represent the attribute values of different attributes of the resources. Taking songs as an example, the tag categories of songs include music genre, language, era, song emotion, etc. For a certain song, its tag value in a certain type of attribute tag is determined. All possible tag values of all songs in a certain type of attribute tag are the tag library of songs in this type of tag, and the tag libraries of each type of all songs are the song tag library corresponding to the song library.

[0102] For example, the music genres of songs include ethnic, rock, punk, etc., so the tag values in the music genre tag library include ethnic, rock, punk, etc. The languages of songs include Chinese, Korean, Japanese, English, German, Russian, etc., so the tag values in the language tag library include Chinese, Korean, Japanese, English, German, Russian, etc. The rhythms of songs include lively, soothing, exciting, etc., so the tag values in the rhythm tag library include lively, soothing, exciting, etc. The music genre tag library, language tag library, rhythm tag library, and tag libraries of other attributes of songs together constitute the song tag library corresponding to the song library.

[0103] In other words, the resource tag library corresponding to the resource library includes sub-tag libraries corresponding to each type of resource tag. The sub-tag libraries corresponding to each type of resource tag include the attribute tag values corresponding to the attribute tags of this type. That is to say, the tag values in the resource tag library are stored classified according to the tag categories, and the tag values of each category correspond to a sub-tag library. Among them, the categories of resource tags included in the resource tag library can be custom-determined according to requirements, and this exemplary embodiment does not make special limitations on this.

[0104] For example, each segment can be matched with all the tag values in the resource tag library. For example, there are 5 types of tags in the resource tag library, and there are 3 tag values corresponding to each type of tag. That is to say, there are a total of 15 tag values. Then, each segment can be respectively matched with these 15 tag values.

[0105] Exemplarily, one matching method in step S520 may be: calculating the similarity between the first semantic vector of the word segmentation and the second semantic vectors of the label values corresponding to each category of resource labels in the resource label library, sorting all the similarities in descending order, and determining that the word segmentation and the label value corresponding to the maximum similarity match successfully when the maximum similarity is greater than the similarity threshold; otherwise, determining that the word segmentation fails to match.

[0106] Among them, the above similarity threshold can be custom-determined according to requirements. The higher the similarity threshold, the higher the reliability of successful matching. The similarity threshold should not be set too small. For example, the similarity threshold can be greater than or equal to 70%. The method of converting the word segmentation and the label value into corresponding semantic vectors can refer to existing related technologies, such as using the BRET (Bidirectional Encoder Representations from Transformers) model for conversion, etc. This exemplary embodiment does not make special limitations on this.

[0107] In the present disclosure, using semantic vectors for matching can enhance the fuzzy matching ability for the user's colloquial input content, improving the accuracy of fuzzy retrieval while facilitating the user to use natural language for retrieval.

[0108] In step S530, according to the successfully matched label value, determine the retrieval word corresponding to the word segmentation and the target resource label category corresponding to the word segmentation.

[0109] Exemplarily, one implementation manner of step S530 may include: determining the successfully matched label value as the retrieval word corresponding to the word segmentation; determining the resource label category to which the successfully matched label value belongs as the target resource label category corresponding to the word segmentation.

[0110] For example, for the user's colloquial expression, such as the word segmentation "fast rhythm", it will be converted into a semantic vector. By means of vector recall, it is determined that the label value successfully matched with "fast rhythm" is the "lively" label. Then, the "lively" label is used as the retrieval word for "fast rhythm". At the same time, the attribute category "rhythm" to which the "lively" label value belongs is used as the target resource label category corresponding to this word segmentation, that is, the retrieval word corresponding to the word segmentation "fast rhythm" is "lively", and the corresponding target resource label category is rhythm.

[0111] In step S540, for each word segmentation, from the resources in the resource library in the target resource label category, determine the resources whose label values in the target resource label category are the retrieval word, and obtain the first resource set corresponding to the word segmentation.

[0112] For example, for the above-mentioned word segment "fast rhythm", songs with the label "lively" can be retrieved from the song library, so as to obtain the first resource set corresponding to this word segment.

[0113] In the present disclosure, by performing specific retrieval and recall through resource tags, the relevance between the target resources in the retrieval results and the user requirements can be ensured, and the accuracy of the retrieved target resources can be improved. Taking the music resource tag library as an example, the music resource tag library includes detailed information of the reference resources in the music library, and these information exist in two forms: text tags and vectors. For example, when a user searches for "rock songs", songs with rock-related tags can be screened out from the music library according to the rock-related tags of the songs in the music library.

[0114] When performing retrieval for each word segment, each word segment can be regarded as a retrieval condition. Therefore, fuzzy retrieval based on word segments can be understood as multi-condition fuzzy retrieval. Multi-condition fuzzy retrieval can be divided into multi-condition fuzzy retrieval related to precise entities and pure fuzzy retrieval.

[0115] Among them, the multi-condition fuzzy retrieval related to precise entities can be understood as that among multiple word segments, there are two types of word segments: precise entities and fuzzy semantics. For example, for the "happy songs of singer A", there are two word segments: "singer A" and "happy songs". "Singer A" belongs to the word segment of the precise named entity type, and "happy songs" belongs to the word segment of the fuzzy semantics type; pure fuzzy retrieval can be understood as that each word segment in the retrieval content is of the fuzzy semantics type. For example, for the above-mentioned "Cantonese emotions in the 1990s", its word segments "1990s", "Cantonese", and "love songs" are not precise named entities, but are all of the fuzzy semantics type, and it involves the comprehensive screening and retrieval of multiple retrieval conditions such as the era, language, and theme.

[0116] For the multi-condition fuzzy retrieval related to precise entities, only the label value matching needs to be performed on the word segments with fuzzy meanings, and the word segments belonging to the precise named entities do not need to perform label value matching. Instead, directly search for the resources corresponding to the precise named entity in the resource library to obtain the first resource set corresponding to the precise named entity.

[0117] In step S550, according to the same first resources in the first resource sets corresponding to each word segment, the target resources corresponding to the first input operation are determined.

[0118] For example, the resources in the intersection of the first resource sets corresponding to each word segment are determined as the target resources corresponding to the first input operation, and then the target resources are pushed to the client.

[0119] In an exemplary embodiment, in the scenario of fuzzy retrieval, the number of target resources is the first number. When the number of the same first resources in the first resource set corresponding to each word segmentation is greater than or less than the first number, the first resources can be excluded or added according to the interest weights of the user in the client for different resource tag categories, so as to obtain the first number of target resources. The first number can be determined customarily according to requirements, such as 50, 100, etc., and this exemplary embodiment does not make special limitations on this.

[0120] Exemplarily, Figure 6 FIG. shows a schematic flowchart of a method for obtaining the first number of target resources in an exemplary embodiment of the present disclosure. Refer to Figure 6 and the method may include steps S610 to S620. Wherein:

[0121] In step S610, according to the historical retrieval data of the client, the interest weights of the client for different resource tag categories are determined.

[0122] For example, the historical retrieval content can be analyzed to determine the categories of resource tags that appear in each historical retrieval content, and then the number of occurrences of each category of resource tags in the historical retrieval content is counted. The interest weights of the client for different resource tag categories are determined according to the number of occurrences, that is, the more the number of occurrences, the greater the interest of the user in the client for the resource tags of this category, and the greater the interest weight of the resource tags of this category.

[0123] For example, taking a music client as an example, in the historical retrieval content of a certain music client in the recent 30 days, the retrieval term of the music genre tag category appears 50 times, and the retrieval term of the language tag category appears 20 times. Then the interest weight of the user of this client for the music genre tag category is greater than that of the language tag category, indicating that when the user listens to music, he pays more attention to the music genre of the song and less attention to the language. The specific value of the interest weight of each user for each resource tag category can be determined customarily according to requirements. For example, the number of occurrences can be directly determined as the interest weight, or the resource tags of different categories can be sorted according to the ascending order of the number of occurrences, and then a specific number is assigned to the interest weight of the resource tag category ranked first, such as 1, and the interest weights of other resource tag categories are increased by 1 in turn according to the sorting order, such as the interest weight of the resource tag category ranked second is 2, and so on.

[0124] Of course, it is also possible to determine the interest weights of the client for different resource tag categories according to the historical operation conditions of the resources in the client by the user of the client. For example, obtain the historical resources clicked, browsed, and consumed by the user in the client in the recent three months, and obtain the interest weights of the client for each category of resource tags according to the number of resource tags of different categories in the historical resources. This exemplary embodiment does not make special limitations on this.

[0125] In step S620, when the number of the same first resources in the first resource set corresponding to each word segment does not meet the first quantity, each first resource set is processed according to the interest weight to obtain the first quantity of target resources.

[0126] For example, when the number of the same first resources in the first resource set corresponding to each word segment is greater than the first quantity, some of the first resources can be randomly deleted first in the first resource set corresponding to the word segment indicated by the tag value with the lowest interest weight to obtain a new first resource set corresponding to the word segment. Then, the intersection of the new first resource set and the first resource sets corresponding to other word segments is taken to see if the quantity meets the first quantity. If it meets, stop. If it is still greater, continue to delete. If it is less than the first quantity, the randomly deleted part of the first resources is added back to the first resource set corresponding to the word segment, and the intersection with the first resource sets of other word segments is taken again, and so on until the quantity is the first quantity.

[0127] For example, there are a total of N word segments, and these N word segments correspond to M fuzzy retrieval conditions. When the number of the same first resources in the first resource set corresponding to each word segment is less than the first quantity, then according to the order of the interest weights corresponding to the resource tag categories to which the word segments corresponding to the M fuzzy retrieval conditions belong, from largest to smallest, first take the intersection in the first resource sets corresponding to the word segments indicated by the first M - 1 fuzzy retrieval conditions with larger interest weights to obtain the second resources. If the sum of the number of the same first resources in the first resource set corresponding to each word segment and the number of the second resources is greater than the first quantity, then randomly delete some of the second resources in the second resources so that the sum of the two quantities is the first quantity; if the sum of the number of the same first resources in the first resource set corresponding to each word segment and the number of the second resources is less than the first quantity, then continue to take the intersection in the first resource sets corresponding to the word segments indicated by the first M - 2 fuzzy retrieval conditions with larger interest weights to obtain the third resources. If the sum of the number of the same first resources in the first resource set corresponding to each word segment, the number of the second resources, and the number of the third resources is greater than the first quantity, then randomly delete some of the third resources in the third resources so that the sum of the three quantities is the first quantity; if the sum of the three quantities is less than the first quantity, then continue to take the intersection in the first resource sets corresponding to the word segments indicated by the first M - 3 fuzzy retrieval conditions with larger interest weights to obtain the fourth resources, and then determine whether to continue to complete the target resources according to the interest weights based on the number of the same first resources, the number of the second resources, the number of the third resources, and the number of the fourth resources in the first resource set corresponding to each word segment. Loop like this until the first quantity of target resources is obtained.

[0128] In another exemplary embodiment, the quantity of the target resources may not be specified, and directly take the same first resources in the first resource set corresponding to each word segment as the target resources. This exemplary embodiment does not make special limitations on this.

[0129] In an exemplary embodiment, when the word segments of each fuzzy semantics are successfully matched, it can be confirmed that the first retrieval content is successfully tagged. In the case of successful tagging, step S550 can be executed to obtain the final target resources. That is, when there are successfully matched tag values for the word segments of each fuzzy semantics type, the target resources corresponding to the first input operation can be determined according to the same first resources in the first resource set corresponding to each word segment.

[0130] In another exemplary embodiment, for word segmentation of fuzzy semantic types, in the case where the word segmentation fails to match the tag value, the word segmentation of the fuzzy semantic type can be input into the resource identifier prediction model mentioned below, and based on the resources indicated by the resource identifier output by the resource identifier prediction model, a first resource set corresponding to the word segmentation is obtained. Then, based on the resources in the intersection of the first resource sets corresponding to each word segmentation, the target resource is determined.

[0131] In yet another exemplary embodiment, in the case where any word segmentation of fuzzy semantics fails to match the tag value, it can be confirmed that the first retrieved content tagging fails. The first retrieved content can be processed based on steps S560 to S570 to obtain the finally retrieved target resource.

[0132] In step S560, the first retrieved content is input into a pre-trained resource identifier prediction model, and the predicted resource identifier of the first retrieved content is obtained according to the output of the resource identifier prediction model.

[0133] As mentioned above, the word segmentation that fails to match the tag value, that is, only the word segmentation that cannot be tagged, can be input into the pre-trained resource identifier prediction model. Based on the output of the resource identifier prediction model, a first resource set corresponding to the word segmentation is obtained, and then the target resource is obtained according to the intersection of the first resource sets of each word segmentation. It is also possible to directly input the entire first retrieved content into the resource tag prediction model to obtain the target resource in the case where there is word segmentation in the first retrieved content that fails to match the tag value.

[0134] Exemplarily, Figure 7 The flowchart shows a method for determining a pre-trained resource identifier prediction model in an exemplary embodiment of the present disclosure. Refer to Figure 7 , this method may include steps S710 to S740. Among them:

[0135] In step S710, the existing resources in the existing resource list are encoded by an identifier encoder to obtain the resource identifiers corresponding to the existing resources.

[0136] In one exemplary embodiment, the existing resource list can be determined according to the resource list created by the user, and the resource list created by the user is the resource list created and named by the user.

[0137] Taking a music client as an example, playlists that meet the requirements can be selected from the user's UGC (User Generated Content) playlists as existing playlists. The existing playlists need to meet the requirement that the playlist titles are tidy and can clearly convey the key information of the songs, so that the playlist titles can be used as important data sources for analyzing the relationship between them and the songs.

[0138] For example, the collected UGC playlists can be subjected to strict multi-level data cleaning to remove noise and invalid data, ensuring data quality and making the data better reflect the true song relationships and characteristics. For example, playlists corresponding to playlist titles with non-standard formats, such as titles containing garbled characters or excessive special characters, are removed, and playlists with incomplete or obviously incorrect content are excluded, such as playlists with too few songs or playlists where the songs in the playlist have no relation to the title. The existing playlists are obtained according to the cleaning results.

[0139] Exemplarily, the resources in the existing resource list can be encoded by an Identity document encoder to obtain the existing resource IDs in the existing resource list, that is, the resource identifiers of the existing resources.

[0140] In step S720, the description information of the existing resource list is input into the initial resource identifier prediction model to obtain a first predicted resource identifier corresponding to the description information of the existing resource list.

[0141] In an exemplary implementation, the description information of the existing resource list can be understood as the title information of the existing resource list. Taking the above-mentioned existing playlists as an example, the description information of the existing playlists can be understood as the playlist titles.

[0142] In an exemplary implementation, the description information of the existing resource list can be encoded by a text encoder to obtain a text vector, thereby representing the semantic information in the playlist title through the text vector.

[0143] For example, the text vector corresponding to the description information of the existing resource list can be used as training data, and the resource identifier corresponding to the existing resource in the existing resource list can be used as the training label data corresponding to the text vector to perform supervised iterative training on an initial resource identifier prediction model, thereby obtaining a pre-trained resource identifier prediction model.

[0144] For example, the text vector corresponding to the description information of the existing resource list is input into the initial resource identifier prediction model, and the output of the initial resource identifier prediction model is determined as the predicted resource identifier corresponding to the description information of the existing resource list, that is, the first predicted resource identifier.

[0145] In an exemplary embodiment, the initial resource identification prediction model may include any machine learning model capable of making predictions. For example, it may include a classification model based on the Transformer architecture, so as to model the mapping relationship between the playlist theme and the song ID by using the classification model based on the Transformer architecture. The Transformer architecture has a self-attention mechanism, which can effectively capture long-distance dependencies and is suitable for processing the complex semantic relationships between playlist titles and songs.

[0146] In step S730, based on the multi-label classification function, determine the degree of difference between the first predicted resource identifier and the resource identifiers corresponding to the existing resources in the existing resource list.

[0147] Taking music resources as an example, during the training process of the model, let the model learn to predict the corresponding song ID encoding from the encoding of the playlist title, so as to establish the mapping relationship between the two. During the training process, since in playlist classification, there may be some categories, for example, the number of playlists of a specific style is large, while the number of other categories is small, the multi-label classification function can be used as the training loss function, so as to obtain the degree of difference between the first predicted identifier and the training label, that is, the training loss.

[0148] In the present disclosure, using the multi-label classification function can automatically balance the number of different categories, avoid the model from over-biasing towards the categories with large numbers, ensure the balance of each category during the recall process, thereby improving the prediction accuracy of the resource identification prediction model, and further improving the accuracy of retrieval.

[0149] In step S740, iteratively train the initial resource identification prediction model according to the degree of difference until a preset condition is met, and obtain the pre-trained resource identification prediction model.

[0150] For example, the initial resource identification prediction model can be iteratively trained with the gradually decreasing degree of difference as the training goal until the degree of difference in a certain training is less than the preset value or the preset number of iterative training times is reached, then stop the training, evaluate the performance of the resource identification prediction model obtained after stopping the training, and after the performance evaluation passes, determine the resource identification prediction model obtained after stopping the training as the pre-trained resource identification prediction model.

[0151] After obtaining the pre-trained resource identification prediction model through the above steps S710 to S740, all of the first retrieval content can be input into the pre-trained resource identification prediction model, and the output of the resource identification prediction model is used as the predicted resource identifier of the first retrieval content.

[0152] In step S570, the target resource is determined according to the resource indicated by the predicted resource identifier in the resource library.

[0153] For example, after obtaining the predicted resource identifier of the first retrieval content, the resource identifier corresponding to the resource in the resource library can be matched with the predicted resource identifier, and the resource that matches successfully is used as the target resource, that is, the resource with the resource identifier in the resource library being the predicted resource identifier is the target resource retrieved for the first input operation.

[0154] For the non-tagged information in the first retrieval content, if conditional retrieval is used, since it cannot be accurately matched with the tags of the resources in the resource library, the retrieval result cannot be accurately obtained. In the present disclosure, through the resource identifier prediction model, the non-tagged information can be processed to improve the accuracy of resource retrieval.

[0155] Through the above steps S510 to S570, the fuzzy retrieval can be effectively processed to improve the accuracy of resource retrieval.

[0156] Exemplarily, in the case where the retrieval type is exact retrieval, another exemplary implementation manner of step S220 may include: retrieving, in the resource library, a target resource that matches the named entity in the first input content.

[0157] For example, in the case where the retrieval type is exact retrieval, it indicates that all the word segments in the first input content are exact named entities, and then the resource indicated by the exact named entity can be directly retrieved in the resource library. For example, if the first input content is "I want to listen to song E by singer A", then song E can be directly found and played among the songs by singer A in the resource library.

[0158] Exemplarily, in the case where the first input content has a corresponding first completion content and the retrieval type of the first completion content is exact retrieval, a target resource that matches the named entity in the first completion content is retrieved in the resource library.

[0159] Exemplarily, the resource retrieval method in the present disclosure may further include: in the case where the first input content of the first input operation does not have a resource retrieval intention, generating an answer content for the first input content according to a pre-trained target language model, and pushing the answer content to the client.

[0160] For example, a pre-trained target language model has the ability to answer knowledge questions and retrieve resources. Taking music resources as an example, when it is determined that the user has the intention to search for songs, the target language model can provide guiding copywriting and interpretable recommendation reasons for the user to help the user better select music; when it is determined that the user does not have the intention to search for songs, such as when the user asks questions about music knowledge or other life questions, it can chat with the user to answer the user's questions.

[0161] In an exemplary embodiment, the first input content in the present disclosure may include a voice type, a text type, and an image type. When the first input content is of the voice type, the text in the voice can be recognized through voice recognition technology, and then based on the recognized text, a retrieval result can be provided for the user or a response result can be generated according to the resource retrieval method in the present disclosure. When the first input content is of the text type, the first input content can be directly processed based on the resource retrieval method of the present disclosure device to obtain a retrieval result or generate a response result.

[0162] When the first input content is of the image type, the target resource can be obtained according to Figure 8 the method shown. Exemplarily, Figure 8 FIG. shows a schematic flowchart of a method for resource retrieval according to an input image in an exemplary embodiment of the present disclosure. Referring to Figure 8 , the method may include steps S810 to S830. Among them:

[0163] In step S810, when the first input content is the first input image, perform scene recognition on the first input image to determine the target scene category to which the first input image belongs.

[0164] For example, when the user's input content is an image, the scene shown in the image can be recognized according to the scene elements in the image. For example, for an image of a sunset by the sea, it can be recognized that it belongs to the seaside scene.

[0165] In step S820, input the target scene category into the pre-trained target language model, and generate the fourth retrieval content of the first input image according to the pre-trained target language model.

[0166] For example, the recognized scene elements and the target scene category can be input into a pre-trained target language model. Continuing with the example of the seaside sunset image mentioned above, the recognized scene elements, the sea and the setting sun, and the target scene category, "seaside scene", can be input into the target language model together, that is, "the sea, the setting sun, and the seaside scene" are input into the target language model. The target language model can generate the fourth retrieval content according to the input content, such as "romantic melody by the sea".

[0167] In step S830, based on the fourth retrieval content, the target resource corresponding to the fourth retrieval content is retrieved from the resource library, and the target resource is pushed to the client.

[0168] For example, the specific implementation of retrieving the target resource corresponding to the fourth retrieval content from the resource library can refer to the specific implementation of retrieving the first retrieval content mentioned above, and will not be elaborated here.

[0169] In an exemplary implementation manner, after the target scene category corresponding to the first input image is recognized, the resource retrieval method in the present disclosure may further include: inputting the target scene category into a pre-trained target language model, and generating recommended description information of the target resource according to the pre-trained target language model.

[0170] For example, the recognized scene elements and the target scene category can be input into a pre-trained target language model. Continuing with the example of the seaside sunset image mentioned above, the recognized scene elements, the sea and the setting sun, and the target scene category, "seaside scene", can be input into the target language model together, that is, "the sea, the setting sun, and the seaside scene" are input into the target language model. The target language model can generate description information such as "In the beautiful scenery of the seaside sunset, let these songs immerse you in romance and tranquility, and feel the perfect integration of the sea and music". In this way, the interpretability of the recommended target resources can be increased. Such a copywriting full of emotion and sense of scene can enable users to better understand the reasons for the recommendation and enhance the user experience and immersion when obtaining resource recommendations.

[0171] Taking the resource in the present disclosure as music as an example, the present disclosure also provides a music resource retrieval system, which may include a service layer, an algorithm application layer, a sorting model layer, a resource recall layer, a content understanding layer, and a data precipitation layer.

[0172] Among them, the service layer provides users with a variety of different music service operations, including music search, song picking, knowledge Q&A, etc. Different service modules can meet the diverse needs of users in the process of music acquisition and music community interaction, providing users with an all-round music experience. The algorithm application layer can implement two functions: searching for songs by text and knowledge Q&A based on different service requirements. In the function of searching for songs by text, it includes functions such as semantic creation, tag extraction, copywriting generation, and chat reply, providing users with music search and interaction services based on text input through natural language processing technology. The emotional dialogue module is dedicated to meeting the needs of users in music-related emotional communication through natural language interaction.

[0173] The sorting model layer establishes multiple sorting models, including a comprehensive sorting model, a user interest model, and a personalized sorting model. These models are based on user portrait data and music content features to sort the recalled music resources, ensuring that the music recommended to users is more in line with their personalized preferences, improving user satisfaction with the recommended results, and thus enhancing the accuracy of music retrieval.

[0174] The resource recall layer can process precise retrieval and fuzzy retrieval respectively to accurately retrieve music resources from different perspectives.

[0175] The content understanding layer includes two sub-modules: image understanding and text understanding. The image understanding module can process the images uploaded by users based on a multi-modal graph-text large model, extract key information in the images, such as scenes, atmospheres, etc., and obtain corresponding retrieval results based on the identified key information. The text understanding module includes functions such as retrieval intention recognition, named entity recognition, sentiment analysis, and context association, deeply analyzing the content input by users to accurately determine user intentions, thereby realizing precise music retrieval and interaction services.

[0176] The data precipitation layer includes rich music-related data, including song tags, song reviews, user playlists, etc. These data provide a solid data foundation for the upper-layer model training and algorithm application. Through continuous data accumulation and update, the models and algorithms are continuously optimized according to the updated data.

[0177] The music resource retrieval system in this disclosure can provide users with efficient and intelligent music retrieval services through the interaction between layers and the combination of their respective functional modules.

[0178] In the present disclosure, the context understanding technology of the large language model can improve the understanding accuracy of the user's natural language query, accurately capture the user's intention, and in the present disclosure, both precise retrieval and fuzzy retrieval can be effectively and accurately processed. At the same time, through multi-condition retrieval and the resource identification prediction model, different types in fuzzy retrieval (such as the fuzzy retrieval related to precise entities and pure fuzzy retrieval mentioned above) can be processed differently, effectively improving the accuracy and recall rate of the retrieval results.

[0179] In summary, the resource retrieval method in the present disclosure can accurately understand the user's intention, perform different types of retrieval processing according to the user's intention, thereby efficiently and accurately retrieving resources related to the user input, and providing personalized and creative retrieval result recommendations and guiding copywriting for the user, bringing a high-quality resource retrieval and recommendation service experience to the user.

[0180] In addition, it should be noted that the above-mentioned drawings are only schematic illustrations of the processes included in the method according to the exemplary embodiments of the present disclosure, rather than for limiting purposes. It is easy to understand that the processes shown in the above-mentioned drawings do not indicate or limit the time sequence of these processes. Additionally, it is also easy to understand that these processes can be executed synchronously or asynchronously, for example, in multiple modules.

[0181] Exemplary Device

[0182] The exemplary embodiments of the present disclosure also provide a resource retrieval device. Referring to Figure 9 As shown, the resource retrieval device 900 may include the following program modules: a retrieval type determination module 910, configured to, in response to a first input operation on the client, determine the retrieval type indicated by the first input operation when there is a resource retrieval intention in the first input content of the first input operation, where the retrieval type includes precise retrieval and fuzzy retrieval; a retrieval module 920, configured to retrieve the target resource corresponding to the first input operation from the resource library according to the retrieval type, and push the target resource to the client.

[0183] In an exemplary embodiment, the method for determining that the first input content has a resource retrieval intention includes: in response to a first input operation on the client, inputting the first input content of the first input operation into a pre-trained target language model, and determining whether the first input content needs to be associated with context according to the target language model; in the case where the target language model determines that the first input content needs to be associated with context, obtaining the most recent historical retrieval communication content; inputting the first input content and the most recent historical retrieval communication content into the target language model, and based on the target language model, complementing the first input content with the most recent historical retrieval communication content to obtain a first complemented content; inputting the first complemented content into the target language model, and determining whether the first complemented content has a resource retrieval intention through the target language model; in the case where the target language model determines that the first complemented content has a resource retrieval intention, determining that the first input content has a resource retrieval intention.

[0184] In an exemplary embodiment, the method for determining the retrieval type indicated by the first input operation includes:

[0185] In the case where the target language model determines that the first complemented content has a resource retrieval intention, performing named entity analysis on the first complemented content through the target language model, and determining the retrieval type indicated by the first input operation according to the named entity analysis result.

[0186] In an exemplary embodiment, retrieving the target resource required by the first input operation from the resource library according to the retrieval type includes: in the case where the retrieval type is fuzzy retrieval, rewriting the first complemented content through the target language model to rewrite the first complemented content into a first retrieval content related to the resource library; retrieving the target resource corresponding to the first input operation from the resource library according to the first retrieval content.

[0187] In an exemplary embodiment, the method for determining the pre-trained target language model includes: inputting a preset prompt into an existing first language model, and using the preset prompt to enable the first language model to determine whether the second input content input into the first language model needs to be completed, the corresponding second completion content in the case of needing completion, whether the second completion content has a retrieval intention, the retrieval type of the second completion content in the case of having a retrieval intention, and the corresponding second retrieval content after rewriting the second completion content when the retrieval type of the second completion content is fuzzy retrieval; after the first language model finishes processing the preset prompt, obtaining a second language model according to the first language model that has processed the preset prompt; obtaining input examples for the resource library, inputting the input examples for the resource library into the second language model, and obtaining label information corresponding to the input examples according to the second language model; training an initial third language model according to the input examples and the label information, and obtaining the target language model according to the training result; wherein, the number of model parameters of the first language model is more than that of the third language model; the label information includes whether the input example needs to be completed, the corresponding third completion content in the case of needing completion, whether the third completion content has a retrieval intention, the retrieval type of the third completion content in the case of having a retrieval intention, and the corresponding third retrieval content after rewriting the third completion content when the retrieval type of the third completion content is fuzzy retrieval.

[0188] In an exemplary embodiment, retrieving the target resource corresponding to the first input operation from the resource library according to the first retrieval content includes: performing word segmentation on the first retrieval content to obtain at least one word segment; for each word segment, respectively matching the word segment with the label values corresponding to each type of resource label in the resource label library corresponding to the resource library, and in the case of successful matching, determining the retrieval word corresponding to the word segment and the target resource label category corresponding to the word segment according to the successfully matched label value; for each word segment, determining, from the resources in the resource library in the target resource label category, the resource whose label value in the target resource label category is the retrieval word, to obtain a first resource set corresponding to the word segment; and determining the target resource corresponding to the first input operation according to the same first resources in the first resource set corresponding to each word segment.

[0189] In an exemplary embodiment, in the case of successful matching, determining the retrieval term corresponding to the word segmentation and the target resource tag category corresponding to the word segmentation according to the successfully matched tag value includes: determining the successfully matched tag value as the retrieval term corresponding to the word segmentation; and determining the resource tag category to which the successfully matched tag value belongs as the target resource tag category corresponding to the word segmentation.

[0190] In an exemplary embodiment, the number of the target resources is a first number, and determining the target resources corresponding to the first input operation according to the same first resources in the first resource set corresponding to each word segmentation includes: determining the interest weights of the client for different resource tag categories according to the historical retrieval data of the client; and in the case where the number of the same first resources in the first resource set corresponding to each word segmentation does not meet the first number, processing each first resource set according to the interest weights to obtain the first number of target resources.

[0191] In an exemplary embodiment, in the case where each word segmentation fails to match, the apparatus further includes a resource identifier prediction module, which may be configured to: input the first retrieval content into a pre-trained resource identifier prediction model, and obtain a predicted resource identifier of the first retrieval content according to the output of the resource identifier prediction model; and determine the target resources according to the resources indicated by the predicted resource identifier in the resource library.

[0192] In an exemplary embodiment, the determining method of the pre-trained resource identifier prediction model includes: encoding the existing resources in the existing resource list through an identifier encoder to obtain the resource identifiers corresponding to the existing resources; inputting the description information of the existing resource list into an initial resource identifier prediction model to obtain a first predicted resource identifier corresponding to the description information of the existing resource list; determining the difference degree between the first predicted resource identifier and the resource identifiers corresponding to the existing resources in the existing resource list based on a multi-label classification function; and iteratively training the initial resource identifier prediction model according to the difference degree until a preset condition is met to obtain the pre-trained resource identifier prediction model.

[0193] In an exemplary embodiment, retrieving the target resources corresponding to the first input operation in the resource library according to the retrieval type includes: in the case where the retrieval type is an exact retrieval, retrieving the target resources matching the named entity in the resource library according to the named entity in the first input content.

[0194] In an exemplary embodiment, the device further includes an answering module, which can be configured to: when the first input content of the first input operation does not have an intention of resource retrieval, generate an answer content for the first input content according to a pre-trained target language model, and push the answer content to the client.

[0195] In an exemplary embodiment, the device further includes an image recognition and retrieval module, which can be configured to: when the first input content is a first input image, perform scene recognition on the first input image to determine a target scene category to which the first input image belongs; input the target scene category into a pre-trained target language model, and generate a fourth retrieval content for the first input image according to the pre-trained target language model; based on the fourth retrieval content, retrieve a target resource corresponding to the fourth retrieval content in a resource library, and push the target resource to the client.

[0196] In an exemplary embodiment, the device further includes a recommended reason generation module, which can be configured to: input the target scene category into a pre-trained target language model, and generate recommended description information for the target resource according to the pre-trained target language model.

[0197] The specific details of each part in the above device have been described in detail in the corresponding method part of the above embodiments. The details not disclosed can be referred to the content of the method part of the above embodiments, and thus will not be repeated.

[0198] Exemplary Storage Medium

[0199] The storage medium of the exemplary embodiment of the present disclosure will be described below.

[0200] In this exemplary embodiment, the above method can be implemented by a program product. For example, a portable compact disc read-only memory (CD-ROM) can be adopted and includes program code, and can run on a device, such as a personal computer. However, the program product of the present disclosure is not limited thereto. In this document, a readable storage medium can be any tangible medium that contains or stores a program, and the program can be used by or in combination with an instruction execution system, device, or device.

[0201] The program product may employ any combination of one or more readable media. The readable media may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples (a non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable 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 disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0202] The computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which the readable program code is carried. Such a propagated data signal may take many forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the foregoing. The readable signal medium may also be any readable medium other than the readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device.

[0203] The program code contained on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0204] The program code for performing the operations of the present disclosure may be written in any combination of one or more programming languages. The programming languages include object-oriented programming languages such as Java, C++, etc., and also include conventional procedural programming languages such as the "C" language or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user computing device and partially on a remote computing device, or entirely on the remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., through the Internet using an Internet service provider).

[0205] Exemplary Computer Program Product

[0206] The exemplary embodiments of the present disclosure also provide a computer program product. The computer program product includes a computer program, which when executed by a processor implements the above-mentioned resource retrieval method.

[0207] In one embodiment, a computer program product may be a tangible product containing a computer program, such as a computer-readable storage medium storing the computer program. The readable storage medium may be a storage medium based on signals such as electricity, magnetism, light, electromagnetic, infrared, etc., including but not limited to: random access memory (RAM), read-only memory (ROM), magnetic tape, floppy disk, flash memory, mechanical hard disk drive (HDD), solid state drive (SSD), and so on. Exemplarily, the computer program product may be implemented as a non-volatile storage medium storing the computer program, such as read-only memory, NAND flash memory, etc.

[0208] In one embodiment, a computer program product may be an intangible product containing a computer program. Exemplarily, the computer program product may be implemented as a virtual digital product, such as an executable file storing the computer program, a digital file such as an installation package.

[0209] The code of the computer program can be written in one or more programming languages. Programming languages such as C, Java, C++, Python, etc. The program code can be executed entirely on the user's computing device, or partially on the user's computing device, or executed as an independent software package, or partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case involving a remote computing device, the remote computing device can be connected to the user's computing device through any type of network, such as a local area network (LAN), a wide area network (WAN), etc., or can be connected to an external computing device (for example, through an Internet connection provided by an operator).

[0210] The computer program can be carried or transmitted by signals such as electricity, magnetism, light, electromagnetic, infrared, etc. The electronic device can convert the signal carrying the computer program into a digital signal and then run the computer program. When the computer program runs on the electronic device, its code is used to cause the electronic device to execute (more specifically, can cause the processor of the electronic device to execute) the method steps of various exemplary embodiments of the present disclosure. For example, it can execute the above-mentioned resource retrieval method, which includes the following steps: in response to a first input operation at the client, when there is a resource retrieval intention in the first input content of the first input operation, determine the retrieval type indicated by the first input operation, where the retrieval type includes precise retrieval and fuzzy retrieval; according to the retrieval type, retrieve the target resource corresponding to the first input operation in the resource library, and push the target resource to the client.

[0211] Executing the above method steps through a computer program. On the one hand, through the recognition and analysis of the retrieval intention and retrieval type, different retrieval processes can be carried out according to the retrieval intention and retrieval type, so as to provide personalized retrieval resource recommendations for users and improve the accuracy of resource retrieval. On the other hand, through the recognition and analysis of the retrieval type, precise retrieval and fuzzy retrieval can be processed simultaneously, thereby improving the applicability of the resource retrieval method.

[0212] Exemplary Electronic Device

[0213] Reference Figure 10 An electronic device according to an exemplary embodiment of the present disclosure will be described. The electronic device is the above-mentioned terminal device 110 or server 120. The electronic device may include a processor and a memory. The memory stores executable instructions of the processor, which may be a computer program. The processor executes the method steps of various exemplary embodiments of the present disclosure by executing the executable instructions. In addition, the electronic device may further include a display for displaying a graphical user interface.

[0214] Next, with reference to Figure 10 , an electronic device will be described by way of example in the form of a general computing device. It should be understood that Figure 10 the electronic device 1000 shown is merely an example and should not impose limitations on the functions and scope of use of the embodiments of the present disclosure.

[0215] As Figure 10 shown, the electronic device 1000 may include: a processor 1010, a memory 1020, a bus 1030, an I / O (input / output) interface 1040, a network adapter 1050, and a display 1060.

[0216] The memory 1020 may include volatile memory, such as RAM 1021 and a cache unit 1022, and may also include non-volatile memory, such as ROM 1023. The memory 1020 may further include one or more program modules 1024. Such program modules 1024 include, but are not limited to: an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include the implementation of a network environment. For example, the program module 1024 may include each module in the above-mentioned device.

[0217] The processor 1010 may include one or more processing units. For example, the processor 1010 may include an AP (Application Processor), a modem processor, a GPU (Graphics Processing Unit), an ISP (Image Signal Processor), a controller, an encoder, a decoder, a DSP (Digital Signal Processor), a baseband processor, and / or an NPU (Neural-Network Processing Unit), etc.

[0218] The processor 1010 can be used to execute executable instructions stored in the memory 1020. For example, it can execute the above resource retrieval method, which includes the following steps: in response to a first input operation on the client, when there is a resource retrieval intention in the first input content of the first input operation, determine the retrieval type indicated by the first input operation, where the retrieval type includes precise retrieval and fuzzy retrieval; according to the retrieval type, retrieve the target resource corresponding to the first input operation in the resource library, and push the target resource to the client.

[0219] Implementing the above method through a computer program, on the one hand, through the recognition and analysis of the retrieval intention and retrieval type, different retrieval processes can be carried out according to the retrieval intention and retrieval type, so as to provide personalized retrieval resource recommendations for users and improve the accuracy of resource retrieval; on the other hand, through the recognition and analysis of the retrieval type, precise retrieval and fuzzy retrieval can be processed simultaneously, thus improving the applicability of the resource retrieval method.

[0220] The bus 1030 is used to realize the connection between different components of the electronic device 1000 and may include a data bus, an address bus, and a control bus.

[0221] The electronic device 1000 can communicate with one or more external devices 1100 (such as a keyboard, a mouse, an external controller, etc.) through the I / O interface 1040.

[0222] The electronic device 1000 can communicate with one or more networks through the network adapter 1050. For example, the network adapter 1050 can provide mobile communication solutions such as 3G / 4G / 5G, or provide wireless communication solutions such as wireless local area network, Bluetooth, and near field communication. The network adapter 1050 can communicate with other modules of the electronic device 1000 through the bus 1030.

[0223] The electronic device 1000 can display a graphical user interface through the display 1060, such as a graphical user interface for displaying the results of audio similarity recognition.

[0224] Although Figure 10 not shown in the figure, other hardware and / or software modules can also be provided in the electronic device 1000, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0225] In addition, the above-mentioned drawings are only schematic illustrations of the processes included in the method according to the exemplary embodiments of the present disclosure, rather than for limiting purposes. It is easy to understand that the processes shown in the above-mentioned drawings do not indicate or limit the chronological order of these processes. Additionally, it is also easy to understand that these processes can be executed synchronously or asynchronously in, for example, multiple modules.

[0226] It should be understood that the present disclosure is not limited to the specific method steps or structures that have been described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. Those skilled in the art will readily think of other embodiments based on the specific embodiments provided by the present disclosure. Therefore, the specific embodiments provided by the present disclosure are only exemplary, and the scope and spirit of the present disclosure are indicated by the claims, and should cover any variations, uses, or adaptations of the present disclosure, which follow the general principles of the present disclosure and include well-known common general knowledge or conventional technical means in the technical field not disclosed by the present disclosure.

Claims

1. A resource retrieval method, characterized in that: include: In response to a first input operation on the client, if the first input content of the first input operation contains a resource search intention, determining a search type indicated by the first input operation, the search type including precise search and fuzzy search; According to the search type, a target resource corresponding to the first input operation is searched in a resource library, and the target resource is pushed to the client.

2. The method according to claim 1, characterized in that The method of determining whether the first input content has a resource search intention includes: In response to a first input operation at the client, inputting a first input content of the first input operation into a pre-trained target language model, and determining whether the first input content needs to be associated with a context according to the target language model; When the target language model determines that the first input content needs to be associated with a context, obtaining the most recent historical search communication content; Inputting the first input content and the most recent historical search communication content into the target language model, and completing the first input content with the most recent historical search communication content based on the target language model to obtain first completed content; Inputting the first completion content into the target language model, and determining whether the first completion content has a resource retrieval intention through the target language model; In a case where the target language model determines that the first completion content has a resource retrieval intention, it is determined that the first input content has a resource retrieval intention.

3. The method according to claim 2, characterized in that The determining of the search type of the first input operation indication includes: When the target language model determines that the first completion content has a resource search intention, the target language model performs named entity analysis on the first completion content, and determines the search type indicated by the first input operation according to the named entity analysis result.

4. The method according to claim 2, characterized in that: According to the search type, searching the resource library for the target resource required for the first input operation includes: In the case where the search type is the fuzzy search, rewriting the first completion content by using the target language model to rewrite the first completion content into a first search content related to the resource library; According to the first search content, a target resource corresponding to the first input operation is retrieved from a resource library.

5. The method according to any one of claims 2 to 4, characterized in that The method for determining the pre-trained target language model includes: Input a preset prompt word into an existing first language model, and use the preset prompt word to enable the first language model to determine whether the second input content input into the first language model needs to be completed, the corresponding second completion content if it needs to be completed, whether the second completion content has a search intention, the search type of the second completion content if there is a search intention, and the second search content corresponding to the rewriting of the second completion content if the search type of the second completion content is fuzzy search; After the first language model completes processing of the preset prompt word, a second language model is obtained according to the first language model that has processed the preset prompt word; Acquire an input example for the resource library, input the input example for the resource library into the second language model, and obtain label information corresponding to the input example according to the second language model; Training an initial third language model according to the input example and the label information, and obtaining the target language model according to the training result; wherein the model parameter amount of the first language model is greater than the model parameter amount of the third language model; The label information includes whether the input example needs to be completed, the corresponding third completion content if completion is required, whether there is a search intent for the third completion content, the search type of the third completion content if there is a search intent, and the corresponding third search content after the third completion content is rewritten if the search type of the third completion content is fuzzy search.

6. The method according to claim 4, characterized in that The step of retrieving a target resource corresponding to the first input operation in a resource library according to the first search content includes: Performing word segmentation processing on the first search content to obtain at least one word segmentation; For each word segment, the word segment is matched with the tag value corresponding to each category of resource tags in the resource tag library corresponding to the resource library. If the match is successful, the search term corresponding to the word segment and the target resource tag category corresponding to the word segment are determined according to the successfully matched tag value; For each word segment, determine, from the resources in the target resource tag category in the resource library, resources whose tag values ​​of the target resource tag category are the search term, and obtain a first resource set corresponding to the word segment; According to the same first resource in the first resource set corresponding to each word segment, a target resource corresponding to the first input operation is determined.

7. A resource search device, characterized in that: include: A search type determination module is configured to, in response to a first input operation on a client, determine a search type indicated by the first input operation if the first input content of the first input operation contains a resource search intention, wherein the search type includes a precise search and a fuzzy search; The retrieval module is configured to retrieve the target resource corresponding to the first input operation in the resource library according to the retrieval type, and push the target resource to the client.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

10. An electronic device, characterized in that: include: one or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, enables the one or more processors to implement the method according to any one of claims 1 to 6.

Images