Model training method, retrieval method, device and electronic equipment

Abstract

The present disclosure provides a model training method, a retrieval method, a device and an electronic device, relates to the technical field of computers, in particular to the technical field of artificial intelligence, machine learning, large models, natural language processing and the like, and can be applied to application scenarios such as content retrieval, generative question answering, search engines, intelligent assistants and the like. The specific implementation scheme comprises: obtaining a first retrieval sentence; constructing a first feature representation of the first retrieval sentence from a plurality of feature representation dimensions; wherein the plurality of feature representation dimensions comprise at least one of a semantic feature dimension, a meta-information feature dimension, a risk feature dimension and a user behavior feature dimension; determining a first domain name set related to the first retrieval sentence based on the first feature representation by using a domain name access model; and updating parameters of the domain name access model based on the first domain name set to obtain a trained domain name access model. The present disclosure can improve the reliability and security of the retrieval result.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, particularly to the fields of artificial intelligence, machine learning, large models, and natural language processing, and can be applied to application scenarios such as content retrieval, generative question answering, search engines, and intelligent assistants. Specifically, it relates to a model training method, a retrieval method, a device, and an electronic device. Background Technology

[0002] Currently, retrieval technology mainly involves: obtaining the semantic relevance between the retrieval statement and each document content in multiple document contents, selecting the first target number of document contents with the highest semantic relevance from the multiple document contents, and then obtaining the retrieval results for the retrieval statement based on this. In other words, the acquisition of retrieval results depends only on the semantic relevance between the retrieval statement and the document content. Summary of the Invention

[0003] This disclosure provides a model training method, a retrieval method, an apparatus, and an electronic device.

[0004] According to a first aspect of this disclosure, a model training method is provided, comprising: Retrieve the first search statement; A first feature representation of the first retrieval statement is constructed from multiple feature representation dimensions; wherein, the multiple feature representation dimensions include at least one of semantic feature dimension, meta-information feature dimension, risk feature dimension, and user behavior feature dimension; Using the domain name admission model, based on the first feature representation, determine the first set of domain names related to the first search statement; Based on the first set of domain names, the parameters of the domain name admission model are updated to obtain the trained domain name admission model.

[0005] According to a second aspect of this disclosure, a retrieval method is provided, comprising: Obtain the second search statement; A second feature representation of the second retrieval statement is constructed from multiple feature representation dimensions; wherein, the multiple feature representation dimensions include at least one of semantic feature dimension, meta-information feature dimension, risk feature dimension, and user behavior feature dimension; Using the target domain admission model, based on the second feature representation, determine the set of second domains related to the second search statement; Based on the second set of domain names, the second search results are obtained for the second search statement.

[0006] According to a third aspect of this disclosure, a model training apparatus is provided, comprising: The first statement acquisition unit is used to acquire the first search statement; The first feature construction unit is used to construct a first feature representation of the first retrieval statement from multiple feature representation dimensions; wherein, the multiple feature representation dimensions include at least one of semantic feature dimension, meta-information feature dimension, risk feature dimension and user behavior feature dimension; The first domain name determination unit is used to determine the first set of domain names related to the first search statement based on the first feature representation using the domain name admission model; The model training unit is used to update the parameters of the domain admission model based on the first set of domain names, so as to obtain the trained domain admission model.

[0007] According to a fourth aspect of this disclosure, a retrieval device is provided, comprising: The second statement acquisition unit is used to acquire the second search statement; The second feature construction unit is used to construct a second feature representation of the second retrieval statement from multiple feature representation dimensions; wherein, the multiple feature representation dimensions include at least one of semantic feature dimension, meta-information feature dimension, risk feature dimension and user behavior feature dimension; The second domain name determination unit is used to determine the set of second domain names related to the second search statement based on the second feature representation using the target domain name admission model. The search result acquisition unit is used to obtain the second search result for the second search statement based on the second domain name set.

[0008] According to a fifth aspect of this disclosure, an electronic device is provided, comprising: At least one processor; Memory that is communicatively connected to at least one processor; The memory stores instructions that can be executed by at least one processor to enable the at least one processor to perform the methods provided in the first and / or second aspects of this disclosure.

[0009] According to a sixth aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions; wherein the computer instructions are used to cause a computer to perform the methods provided in the first and / or second aspects of this disclosure.

[0010] According to a seventh aspect of this disclosure, a computer program product is provided, including a computer program; wherein, when executed by a processor, the computer program is capable of implementing the methods provided in the first and / or second aspects of this disclosure.

[0011] Using this disclosure can improve the reliability and security of search results.

[0012] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0013] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 A schematic flowchart of a model training method provided in an embodiment of this disclosure; Figure 2 A flowchart illustrating a retrieval method provided in an embodiment of this disclosure; Figure 3 This is a schematic diagram illustrating an application scenario of a model training method and a retrieval method provided in an embodiment of this disclosure; Figure 4 A schematic structural block diagram of a model training device provided in an embodiment of this disclosure; Figure 5 A schematic structural block diagram of a retrieval device provided in an embodiment of this disclosure; Figure 6 This is a schematic structural block diagram of an electronic device provided in an embodiment of the present disclosure. Detailed Implementation

[0014] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0015] As mentioned earlier, traditional retrieval techniques primarily involve: obtaining the semantic relevance between the search query and each document content in a set of documents; selecting a first set of a target number of documents with the highest semantic relevance from the multiple document contents; and then obtaining the retrieval results for the search query based on this. In other words, the retrieval of results depends solely on the semantic relevance between the search query and the document content. However, this selection method, which relies solely on semantic relevance, typically makes it difficult to ensure the reliability and security of the retrieval results.

[0016] To address the above problems, this disclosure provides a model training method that can be applied to electronic devices. The electronic device can be a server, workbench, mainframe computer, conventional computer (e.g., desktop computer, laptop computer, tablet computer, etc.) or other similar computing devices. The following will be combined with... Figure 1The flowchart shown illustrates a model training method provided in an embodiment of this disclosure. It should be noted that, although in Figure 1 The flowchart shown illustrates the logical order; however, in some cases, the steps shown or described in the flowchart may be performed in a different order.

[0017] Step S101: Obtain the first search statement.

[0018] The first search statement can be a word, phrase, statement, or paragraph used to express the search intent.

[0019] Step S102: Construct the first feature representation of the first retrieval statement from multiple feature representation dimensions.

[0020] Among them, multiple feature representation dimensions may include at least one of semantic feature dimension, meta-information feature dimension, risk feature dimension, and user behavior feature dimension.

[0021] Based on this, in this embodiment of the disclosure, when multiple feature representation dimensions include semantic feature dimension, meta-information feature dimension, risk feature dimension, and user behavior feature dimension, the first individual feature of the first search statement under the semantic feature dimension, the first individual feature of the first search statement under the meta-information feature dimension, the first individual feature of the first search statement under the risk feature dimension, and the first individual feature of the first search statement under the user behavior feature dimension can be obtained, and based on these first individual features, the first feature representation of the first search statement is obtained. Specifically, the first individual feature of the first search statement under the semantic feature dimension can represent the first overall semantics of the first search statement; the first individual feature of the first search statement under the meta-information feature dimension can represent multiple first meta-information of the first search statement; the first individual feature of the first search statement under the risk feature dimension can represent the first overall risk level of the first search statement; and the first individual feature of the first search statement under the user behavior feature dimension can represent the first behavioral feature of the first user who initiated the first search statement.

[0022] Step S103: Using the domain name admission model, based on the first feature representation, determine the first set of domain names related to the first search statement.

[0023] The domain name admission model can be a neural network model, specifically a neural network model that includes an encoder. Here, the encoder can be a multilayer perceptron (MLP), a long short-term memory network (LSTM), or a large model (e.g., a neural network model with a Transformer architecture).

[0024] In this embodiment of the disclosure, the domain name admission model can determine multiple first preliminary domain names related to the first search statement from the full domain name set based on the first feature representation, thereby forming a first domain name set. The full domain name set can include all internet domain names that the target platform has access to. Here, the target platform can be a content retrieval platform, a generative question-answering platform, a search engine, an intelligent assistant, etc., capable of applying the target domain name admission model, and the target domain name admission model can be a trained domain name admission model.

[0025] Step S104: Based on the first set of domain names, update the parameters of the domain name admission model to obtain the trained domain name admission model.

[0026] In this embodiment of the disclosure, an overall reward objective can be constructed based on the first set of domain names, and the parameters of the domain name admission model can be updated by optimizing the overall reward objective to achieve reinforcement learning (RL) of the domain name admission model and obtain a trained domain name admission model.

[0027] Furthermore, it should be noted that in this embodiment, steps S101, S102, S103, and S104 can be executed cyclically until the number of iterations reaches a threshold, or the trained domain name admission model meets the application requirements, and the trained domain name admission model is used as the target domain name admission model. The threshold for the number of iterations and the application requirements can be set according to the application needs of the target domain name admission model, and this embodiment does not impose any restrictions on them.

[0028] The model training method provided in this disclosure allows for the acquisition of a first search statement, the construction of a second feature representation of the second search statement from multiple feature representation dimensions, the determination of a first set of domain names related to the first search statement using a domain name admission model based on the first feature representation, and finally, the parameter update of the domain name admission model based on the first set of domain names to obtain a trained domain name admission model. The multiple feature representation dimensions may include at least one of semantic feature dimensions, meta-information feature dimensions, risk feature dimensions, and user behavior feature dimensions. In other words, in this disclosure, when the final target domain name admission model is applied to a target platform for content retrieval, it no longer relies solely on semantic relevance selection as in traditional solutions. Instead, it constructs a second feature representation of the second search statement from multiple feature representation dimensions (e.g., semantic feature dimensions, meta-information feature dimensions, risk feature dimensions, and user behavior feature dimensions), determines a second set of domain names related to the second search statement using the target domain name admission model based on the second feature representation, and then obtains a second search result for the second search statement based on the second set of domain names. This multi-dimensional comprehensive feature representation method, while ensuring that the second search result has a high semantic relevance to the second search statement, also enables the target domain name admission model to more comprehensively and deeply understand the reliability requirements and security risks implied in the second search statement. Thus, it can proactively filter out potentially untrusted domain names and high-risk domain names at the domain name selection stage, control the source quality of the second search result from the source, and effectively improve the reliability and security of the second search result.

[0029] In this embodiment of the disclosure, after executing step S101 to obtain the first search statement, the first search statement can be classified to determine the risk category label of the first search statement. The risk category label of the first search statement can be one of a high-risk label, a medium-risk label, and a low-risk label. Specifically: High-risk tags can indicate that the search query belongs to high-risk topics such as medical diagnosis, legal interpretation, and financial advice; The medium-risk label can indicate that the search query belongs to medium-risk topics such as vocational education and health consultation; Low-risk tags can indicate that the search query belongs to low-risk topics such as daily life, entertainment, and light knowledge Q&A.

[0030] In one example, when classifying the first search statement to determine its risk category label, a high-risk confidence level can be obtained to characterize the probability that the first search statement belongs to a high-risk topic. Subsequently, if the first search statement matches a set of high-risk topics and its high-risk confidence level is greater than or equal to a high-risk threshold, the risk category label of the first search statement can be determined as a high-risk label; if it matches a set of high-risk topics and its high-risk confidence level is less than a high-risk threshold, the risk category label of the first search statement can be determined as a medium-risk label; or, if it matches a set of medium-risk topics, the risk category label of the first search statement can be determined as a medium-risk label; in all other cases, the risk category label of the first search statement can be determined as a low-risk label. The high-risk topic set and the medium-risk topic set can be obtained manually, and this embodiment does not impose any restrictions on this; the high-risk threshold can be set according to the application requirements of the target domain name admission model, and this embodiment also does not impose any restrictions on this.

[0031] The above process can be characterized as follows: in, Used to characterize the first search statement; Used to characterize high-risk thematic sets; Used to characterize the high-risk confidence level of the first search statement; Used to characterize high-risk thresholds; The risk category label used to characterize the first search statement is the high-risk label; Used to characterize a set of medium-risk topics; The risk category label used to characterize the first search statement is the medium risk label; The risk category label used to characterize the first search query is the low-risk label.

[0032] Furthermore, it should be noted that in this embodiment of the disclosure, the high-risk confidence level of the first search statement can be obtained by a confidence model. The confidence model can be a neural network model, specifically including a Bidirectional Encoder Representations from Transformer (BERT) model, and a classifier (e.g., a Softmax classifier) ​​connected to the output of BERT.

[0033] It should also be noted that, in this embodiment of the disclosure, the confidence model may be trained. In some optional implementations, the training process of the confidence model may include: Obtain a sample of the first search statement; Using a confidence model, we obtain the high-risk confidence prediction results for the first search statement sample; Obtain the confidence loss of the high-risk confidence prediction result relative to the true high-risk confidence value of the first search statement sample; The confidence model is trained based on the confidence loss to obtain the trained confidence model.

[0034] The true values ​​of high-risk confidence can be obtained through manual annotation, and this disclosure does not impose any restrictions on this method.

[0035] Based on the above, in this embodiment of the disclosure, step S102, namely, "constructing a first feature representation of the first retrieval statement from multiple feature representation dimensions", may include: For each feature representation dimension among multiple feature representation dimensions, construct the first single feature of the first search statement under the feature representation dimension; Based on multiple first single features that correspond one-to-one with multiple feature representation dimensions, the first feature representation of the first retrieval statement is obtained.

[0036] As previously stated, in this embodiment of the disclosure, the multiple feature representation dimensions may include at least one of semantic feature dimensions, meta-information feature dimensions, risk feature dimensions, and user behavior feature dimensions.

[0037] Based on this, in the embodiments of this disclosure, the implementation of "constructing a first single feature of the first retrieval statement under each of the multiple feature representation dimensions" may include at least one of the following: (1) Multiple feature representation dimensions include semantic feature dimensions In this embodiment of the disclosure, when multiple feature representation dimensions include semantic feature dimensions, "constructing a first single feature of the first retrieval statement under each of the multiple feature representation dimensions" may include: The first search statement is vectorized to obtain its semantics. The semantics of the first statement are used as the first single feature of the first search statement under the semantic feature dimension.

[0038] The semantics of the first statement can represent the position of the first search statement in the semantic space.

[0039] In one example, when vectorizing the first search statement to obtain its semantics, a pre-trained semantic vector model (BAAI General Embedding, BGE) can be used to vectorize the first search statement to obtain its semantics, so that the semantics of the first statement can be used as the first single feature of the first search statement under the semantic feature dimension.

[0040] The above process can be characterized as follows: in, Used to characterize the first search statement; This is used to represent the vectorization of the first search statement using BGE; Used to characterize the semantics of the first statement, it will serve as the first single feature of the first search statement under the semantic feature dimension; Used for characterization 3D eigenvectors; This is used to characterize the semantics of the first statement obtained by vectorizing the first search statement using BGE. A dimensional eigenvector. Here, It can be a positive integer, and its specific value can be set according to BGE. This disclosure does not limit this.

[0041] In this embodiment of the disclosure, when multiple feature representation dimensions include semantic feature dimensions, "constructing a first single feature of the first search statement under each feature representation dimension" may also include: The risk category labels of the first search statement are vectorized to obtain the semantics of the first category. The first category of semantics is used as the first single feature of the first search statement under the semantic feature dimension.

[0042] Among them, the semantics of the first statement can characterize the position of the risk category label of the first search statement in the semantic space.

[0043] In one example, when vectorizing the risk category label of the first search statement to obtain the first category semantics, an embedding model can be used to vectorize the risk category label of the first search statement to obtain the first category semantics, so that the first category semantics can be used as the first single feature of the first search statement under the semantic feature dimension.

[0044] The above process can be characterized as follows: in, Risk category labels used to characterize the first search statement; This is used to represent the vectorization of the risk category labels of the first search statement using an embedding model; Used to characterize the first category of semantics, it will serve as the first single feature of the first search statement under the semantic feature dimension; Used for characterization 3D eigenvectors; The semantic classification of the first category obtained by vectorizing the risk category label of the first search statement using an embedding model is used to characterize the first category. A dimensional eigenvector. Here, It can be a positive integer, and its specific value can be determined according to... The specific value setting is not limited in the embodiments disclosed herein.

[0045] Based on the above, it can be understood that in this embodiment of the disclosure, the first overall semantics of the first search statement may include the first statement semantics and the first category semantics of the first search statement.

[0046] Through the above implementation methods, in this embodiment of the disclosure, the first search statement can be vectorized to obtain the semantics of the first statement, and the semantics of the first statement can be used as the first single feature of the first search statement under the semantic feature dimension. Simultaneously, the risk category label of the first search statement is vectorized to obtain the semantics of the first category, and the semantics of the first category can be used as the first single feature of the first search statement under the semantic feature dimension. In this way, not only can the deep semantic information of the first search statement itself be captured, but the deep semantic information of the risk category label of the first search statement can also be explicitly introduced, making the first single feature of the first search statement under the semantic feature dimension richer and more targeted. This provides a more accurate semantic dependency for domain name selection in the domain name admission model, thereby improving the learning efficiency and effectiveness of the domain name admission model.

[0047] (2) Multiple feature representation dimensions include meta-information feature dimensions In this embodiment of the disclosure, when multiple feature representation dimensions include meta-information feature dimensions, "constructing a first single feature of the first retrieval statement under each of the multiple feature representation dimensions" may include: Obtain multiple first-element information of the first search statement; Based on multiple first-level metadata, the first single-item feature of the first search statement under the metadata feature dimension is obtained.

[0048] The first meta-information may include at least one of the following: statement length, language type, relevant location, and real-time requirement information. Here, statement length can be the number of characters; language type can be Chinese, English, Korean, Japanese, etc.; relevant location can be the geographical location of the first user initiating the first search statement; real-time requirement information can characterize whether the first search statement has a timeliness requirement, which can be determined by detecting whether the first search statement includes timeliness keywords such as "latest," "today," or "real-time." For example, if the first search statement includes timeliness keywords, it is determined that the first search statement has a timeliness requirement; if the first search statement does not include timeliness keywords, it is determined that the first search statement does not have a timeliness requirement.

[0049] In one example, after obtaining multiple first meta-information pieces of the first search statement, a one-dimensional feature vector can be obtained based on each of the multiple first meta-information pieces. The first single feature of the first search statement under the meta-information feature dimension is then composed of multi-dimensional feature vectors that correspond one-to-one with the multiple first meta-information pieces. Here, "the first single feature of the first search statement under the meta-information feature dimension is composed of multi-dimensional feature vectors that correspond one-to-one with the multiple first meta-information pieces" can be characterized as: Where m is used to represent the first single feature of the first retrieval statement under the meta-information feature dimension; Used for characterization 3D eigenvectors; The first single feature used to characterize the first search statement under the meta-information feature dimension belongs to A dimensional eigenvector. Here, It can be a positive integer, and its specific value can be the number of first element information. This disclosure does not limit this.

[0050] Through the above implementation methods, in this embodiment of the disclosure, multiple first meta-information of the first search statement can be obtained, and based on the multiple first meta-information, a first single feature of the first search statement under the meta-information feature dimension can be obtained. The multiple first meta-information may include at least one of statement length, language type, relevant position, and real-time requirement information. In this way, structured contextual information (i.e., statement length, language type, relevant position, and real-time requirement information of the first search statement) other than the first overall semantics can be added to the domain name admission model, thereby improving the scenario adaptability of the domain name admission model.

[0051] (3) Multiple feature representation dimensions include risk feature dimensions In this embodiment of the disclosure, when multiple feature representation dimensions include risk feature dimensions, "constructing a first single feature of the first search statement under each feature representation dimension" may include: Based on the risk category label of the first search statement, the high-risk confidence of the first search statement, and the search complexity of the first search statement, the first content risk level of the first search statement is obtained. The first content risk level is used as the first single feature of the first search statement under the risk feature dimension.

[0052] In one example, the retrieval complexity of the first search query can be obtained in the following way: Obtain multiple first-statement features of the first search statement; For each of the multiple first statement features, the first statement feature is normalized to obtain the first feature processing result; The retrieval complexity of the first retrieval statement is obtained by weighted fusion of the processing results of multiple first feature features that correspond one-to-one with the first statement features.

[0053] The first-statement features can include at least one of the following: statement length, number of entities, sentence structure complexity, intent mixing degree, and number of time-sensitive keywords. Here, statement length can be the number of characters; entities can be specific objects, concepts, or proper nouns; sentence structure complexity can characterize whether the first search statement includes multiple levels of nesting or conditional clauses. For example, the sentence structure complexity can be set to "1" when the first search statement includes multiple levels of nesting or conditional clauses, and set to "0" when the first search statement does not include multiple levels of nesting or conditional clauses; intent mixing degree can characterize the number of search intents expressed by the first search statement; time-sensitive keywords can be "latest," "today," "real-time," etc.

[0054] In one example, when obtaining the first content risk level of a first search query based on its risk category label, high-risk confidence level, and retrieval complexity, a content risk awareness model can be used. This model can be a neural network model, specifically a multilayer perceptron (MLP).

[0055] The above process can be characterized as follows: in, Risk category labels used to characterize the first search statement; Used to characterize the high-risk confidence level of the first search statement; Used to characterize the retrieval complexity of the first retrieval statement; This is used to characterize the processing of the risk category label, high-risk confidence level, and retrieval complexity of the first search statement using the content risk perception model. Used to characterize the first content risk level of the first search query.

[0056] Furthermore, it should be noted that in this embodiment of the disclosure, the content risk perception model may be trained. In some optional implementations, the training process of the content risk perception model may include: Obtain a sample of the second search statement; Using a content risk perception model, based on the risk category label of the second search statement sample, the high-risk confidence of the second search statement sample, and the retrieval complexity of the second search statement sample, the content risk prediction result of the second search statement sample is obtained. The risk loss of the content risk prediction result relative to the true content risk value of the second search query sample is obtained. Based on the risk level loss, the content risk perception model is trained to obtain the trained content risk perception model.

[0057] The true value of content risk level can be obtained through manual annotation, and this disclosure does not limit this method.

[0058] In this embodiment of the disclosure, when multiple feature representation dimensions include risk feature dimensions, "constructing a first single feature of the first search statement under each feature representation dimension" may also include: Based on the frequency of occurrence, coverage, and location of sensitive words in the first search query, the risk level of triggering the first sensitive word in the first search query is obtained. The risk level triggered by the first sensitive word is used as the first single feature of the first search statement under the risk feature dimension.

[0059] The frequency of sensitive words can be defined as the number of sensitive words matched by the first search query in the sensitive word dictionary; the sensitive word coverage can represent the proportion of the number of sensitive word categories in the first search query to the total number of sensitive word categories in the sensitive word dictionary; and the location of sensitive words can represent the positional information of sensitive words in the first search query, such as at the beginning, middle, or end of the sentence, or their positional offset relative to core entities in the first search query. The sensitive word dictionary can be generated manually and stores sensitive words of different categories, such as "diagnosis," "complications," "lawyer's opinion," "dosage," and "profit forecast."

[0060] In one example, the risk level of the first sensitive word in the first search query can be obtained in the following way: The frequency of sensitive words in the first search query is normalized to obtain the first data processing result; The sensitivity word coverage of the first search statement is normalized to obtain the first coverage result; The positions of sensitive words in the first search statement are normalized to obtain the first position processing result; The results of the first data processing, the first coverage processing, and the first position processing are weighted and fused to obtain the first sensitive word trigger risk level of the first search statement.

[0061] Based on the above, it can be understood that in this embodiment of the disclosure, the first overall risk level of the first search statement may include the first content risk level and the first sensitive word triggering risk level of the first search statement.

[0062] Through the above implementation methods, in this embodiment of the disclosure, a first content risk level of the first search statement can be obtained based on the risk category label of the first search statement, the high-risk confidence level of the first search statement, and the search complexity of the first search statement. This first content risk level is then used as the first single feature of the first search statement under the risk feature dimension. Simultaneously, a first sensitive word trigger risk level of the first search statement is obtained based on the frequency of occurrence of sensitive words, the coverage of sensitive words, and the location of sensitive words in the first search statement. This first sensitive word trigger risk level is also used as the first single feature of the first search statement under the risk feature dimension. In other words, in this embodiment of the disclosure, the first search statement can be jointly risk-quantified from two sub-risk feature dimensions: content risk level and sensitive word trigger risk level. This considers both the macro-level thematic risk of the first search statement and captures the micro-level sensitive signals of the first search statement. This not only enables the domain name access model to implement stricter domain name access control for high-risk search statements but also helps the domain name access model to more accurately assess search risks. This allows for the implementation of differentiated security control strategies in authoritative domain name selection, search scope control, and evidence constraints, thereby further improving the scenario adaptability of the domain name access model.

[0063] (4) Multiple feature representation dimensions include user behavior feature dimensions In this embodiment of the disclosure, when multiple feature representation dimensions include user behavior feature dimensions, "constructing a first single feature of the first search statement under each feature representation dimension" may include: Determine the first user who initiates the first search query; Obtain the first user's primary behavioral characteristics; Based on the first behavioral feature, the first single feature of the first search statement under the user behavior feature dimension is obtained.

[0064] The first behavioral feature may include multiple first sub-features. Here, the first sub-feature may include at least one of the following: Authoritative source adoption rate: The proportion of search results from authoritative domains that were explicitly adopted by the first user within the first historical period (e.g., bookmarked, clicked "helpful", "trustworthy", or gave a positive review); Approval rate: The percentage of search results obtained from the first user's query within the first historical period that are deemed compliant after review (manual or machine review); Negative feedback rate: The proportion of search results obtained from the first user's query within the first historical period that received negative feedback from the first user (e.g., clicking "not helpful", "unreliable", or giving a negative rating); High-risk search ratio: The proportion of search statements that belong to high-risk topics in the first user's query statement within the first historical period; Average interaction time: The average time a user spends on a page when browsing the search results obtained from a query initiated by the first user within the first historical time period. Correction request ratio: Within the first historical period, the proportion of search results obtained from the query statement initiated by the first user that were requested to be regenerated by the first target user.

[0065] In the above description, authoritative domain names may include domain names of educational and research institutions (e.g., .edu, websites of well-known universities, etc.), official websites of international organizations (e.g., who.int, un.org, cdc.gov, etc.), portals of international authoritative institutions, etc.; the first historical time period may be a time period prior to the current time point with a duration of a first preset duration. The first preset duration can be set according to the application requirements of the target domain name access model, and this disclosure does not impose any limitations on it.

[0066] As mentioned above, in this embodiment of the disclosure, after determining the first user who initiates the first search statement and obtaining the first user's first behavioral features including multiple first sub-features, the first single feature of the first search statement under the user behavioral feature dimension will be obtained based on the multiple first sub-features. For example, multiple first sub-features can be concatenated to obtain the first single feature of the first search statement under the user behavioral feature dimension.

[0067] Through the above implementation methods, in this embodiment of the disclosure, the first user initiating the first search statement can be determined, and the first behavioral characteristics of the first user can be obtained. Then, based on the first behavioral characteristics, the first single feature of the first search statement under the user behavioral characteristic dimension can be obtained. In other words, in this embodiment of the disclosure, by characterizing the user's personalized behavioral patterns, the domain name admission model's perception of the user's personalized behavioral patterns can be improved. This improves the user acceptance of the search results while ensuring the reliability and security of the search results, thereby enhancing the user's experience with the target platform.

[0068] As mentioned above, in this embodiment of the present disclosure, after constructing the first single feature of the first search statement under each of the multiple feature representation dimensions, the first feature representation of the first search statement is obtained based on the multiple first single features that correspond one-to-one with the multiple feature representation dimensions. For example, the multiple first single features can be concatenated to obtain the first feature representation of the first search statement.

[0069] The above process can be characterized as follows: in, Used to characterize the first single feature of the first search statement under the semantic feature dimension, specifically the first statement semantics of the first search statement; The first single feature used to characterize the first search statement under the semantic feature dimension is specifically the first category semantic of the first search statement. Used to characterize the first single feature of the first search statement under the meta-information feature dimension; The first single feature used to characterize the first search statement under the risk feature dimension is specifically the first content risk degree of the first search statement. The first single feature used to characterize the first search statement under the risk feature dimension is the risk degree triggered by the first sensitive word in the first search statement. The first single feature of the first search statement is used to characterize the user behavior feature dimension; x is used to characterize the first feature representation of the first search statement.

[0070] In summary, in this embodiment of the present disclosure, when executing step S102, a first single feature of the first search statement under each of the multiple feature representation dimensions can be constructed, and a first feature representation of the first search statement can be obtained based on the multiple first single features that correspond one-to-one with the multiple feature representation dimensions. Specifically, the first single features of the first search statement under the semantic feature dimension, the first single features of the first search statement under the meta-information feature dimension, the first single features of the first search statement under the risk feature dimension, and the first single features of the first search statement under the user behavior feature dimension can be integrated into a unified first feature representation, constructing a comprehensive, refined, and risk-aware input state for the domain name admission model. This input state can support the domain name admission model to make differentiated and safe and controllable domain name selection decisions in different search scenarios (for example, tightening or widening the domain name admission range, i.e., the search range, according to the overall risk level of the search statement), fundamentally overcoming the problem of insufficient reliability and security caused by traditional solutions relying solely on semantic relevance.

[0071] Furthermore, in this embodiment of the present disclosure, during the execution of step S103, in the process of determining the first set of domain names related to the first search statement based on the first feature representation using the domain name admission model, when determining multiple first preliminary domain names related to the first search statement from the full set of domain names based on the first feature representation using the domain name admission model to form the first set of domain names, the first admission weight of each of the multiple first preliminary domain names can also be obtained.

[0072] Based on this, in this embodiment of the disclosure, the domain name admission model, in addition to the main model (i.e., the encoder), may also include a first output head and a second output head connected in parallel to the output of the main model. The first output head may include a combined network of fully connected layers and activation functions to output multiple first preliminary domain names; the second output head may also include a combined network of fully connected layers and activation functions to output the first admission weight of each of the multiple first preliminary domain names. For each of the multiple first preliminary domain names, its first admission weight can characterize the dependence on the document content under the first preliminary domain name when obtaining the first search result for the first search statement. Specifically, for each of the multiple first preliminary domain names, the larger its first admission weight, the higher the reference priority for the document content under the first preliminary domain name when obtaining the first search result for the first search statement.

[0073] For ease of understanding, in this embodiment of the disclosure, the execution process of step S103 can be contextualized as a Markov decision process, including: (1) State space The input state can be represented as: in, The first feature representation used to characterize the first search statement; Used to characterize the input state.

[0074] (2) Action space In this embodiment of the disclosure, in order to adapt to the two-stage characteristics of "first selecting a domain name and then allocating admission weights", the action space can be decomposed into two types of complementary actions: the action of selecting the initial domain name and the action of allocating admission weights.

[0075] The selection of the initial domain name can be represented as follows: in, Used to represent the complete set of domain names; Used to represent the first set of domain names; Used to represent the number of the first initially selected domain names in the first set of domain names; The threshold used to characterize the number of domain names can be set according to the application requirements of the target domain name admission model, and this disclosure does not limit it.

[0076] The allocation of admission weights can be characterized as follows: in, Used to represent a specific first preliminary domain name (hereinafter referred to as the target first preliminary domain name) in the first domain name set (including multiple first preliminary domain names); Used to represent the complete set of domain names; The first admission weight used to characterize the first target initial selection domain name in the first domain name set is located in the numerical range [0, 1]. Used to represent multiple first admission weights that correspond one-to-one with multiple first initial domain names.

[0077] Based on the above, in this embodiment of the disclosure, step S104, namely, "updating the parameters of the domain name admission model based on the first domain name set to obtain the trained domain name admission model," may include: Based on multiple first-selection domains and the first admission weight of each first-selection domain, the overall domain admission reward related to the first search statement is obtained; Based on the overall domain name admission reward, the parameters of the domain name admission model are updated to obtain the trained domain name admission model.

[0078] The overall domain admission reward can be obtained based on the ranking reward of multiple first-selection domains. Here, the ranking reward of multiple first-selection domains can represent the reliability of the weight ranking result after ranking the multiple first-selection domains according to their first admission weight.

[0079] Based on this, in the embodiments of this disclosure, "obtaining the overall domain name admission reward related to the first search statement based on multiple first preliminary domain names and the first admission weight of each of the multiple first preliminary domain names" may include: Based on the first admission weight of each of the multiple first-selection domains, a ranking reward is obtained for the multiple first-selection domains; Based on the ranking reward, the overall domain name access reward related to the first search query is obtained.

[0080] For ranking rewards for multiple first-selection domains, the methods for obtaining them can be: For each of the multiple initial domain names, an evaluation is conducted from multiple domain evaluation dimensions to obtain multiple individual domain evaluation results corresponding one-to-one with the multiple domain evaluation dimensions. Based on the multiple individual domain evaluation results, an overall domain evaluation result for the initial domain name is obtained. The multiple domain evaluation dimensions may include at least one of the following: content relevance dimension, content quality dimension, and authority level dimension.

[0081] The evaluation results of a single domain name for the first search statement under the content relevance dimension can characterize the semantic relevance between the first query statement and the document content under the first initially selected domain name; The evaluation results of a single domain for the first search query under the content quality dimension can represent the comprehensive content quality of the document content under the first initially selected domain under multiple quality evaluation dimensions such as structural clarity, content richness, and expression accuracy. The evaluation result of a single domain name for the first search query under the authority level dimension can represent the authority level of the first initially selected domain name.

[0082] In the above description, the authority level of the first initially selected domain name can be obtained through manual annotation (e.g., official website of international organizations = portal of international authoritative institutions > educational and research institutions > trusted media), and this disclosed embodiment does not limit this.

[0083] In this embodiment of the disclosure, "obtaining an overall domain name evaluation result for the first initial domain name based on multiple individual domain name evaluation results" may include: The evaluation results of individual domain names for the first search query under the content relevance dimension are normalized to obtain the first evaluation result. The evaluation results of individual domain names for the first search query under the content quality dimension are normalized to obtain the second evaluation result; The evaluation results of individual domain names for the first search query under the authority level dimension are normalized to obtain the third evaluation result. The first, second, and third evaluation results are weighted and merged to obtain the overall domain name evaluation result for the first initial domain name.

[0084] In this embodiment of the disclosure, after obtaining multiple overall domain name evaluation results corresponding one-to-one with multiple first initial domain names, the positive correlation between the multiple overall domain name evaluation results and the weight ranking results of the multiple first initial domain names can be obtained as a ranking reward for the multiple first initial domain names. Specifically: Multiple initial domain name pairs are constructed based on multiple first initial domain names; For each of the multiple initial domain name pairs, the relevance contribution value of the initial domain name pair is obtained based on the consistency between the weight ranking result and the evaluation ranking result. Based on the multiple correlation contribution values ​​corresponding one-to-one with multiple initial domain names, the positive correlation between the multiple overall domain name evaluation results and the weight ranking results of multiple first-selection domain names is obtained, which serves as the ranking reward for multiple first-selection domain names.

[0085] When constructing multiple initial domain name pairs based on multiple first initial domain names, an exhaustive list of binary ordered combinations of the multiple first initial domain names can be used to construct multiple initial domain name pairs. For example, the multiple first initial domain names include first initial domain name A1, first initial domain name A2, and first initial domain name A3. Therefore, the multiple initial domain name pairs constructed based on the multiple first initial domain names can include: The initial domain name pairs are A1 & A2; The initial domain name pairs are A1 and A3; The initial domain name pairs are A2 and A1; The initial domain name pairs are A2 and A3; The initial domain name pairs are A3 and A1; The initial domain name pairs are A3 and A2.

[0086] After constructing multiple initial domain name pairs based on multiple initial domain names, for each initial domain name pair, the relevance contribution value of the initial domain name pair is obtained based on the consistency between the weight ranking result and the evaluation ranking result. For example, when the weight ranking result and the evaluation ranking result of the initial domain name pair are consistent, the relevance contribution value of the initial domain name pair can be set to a positive value. Specifically, it can be the difference between the overall domain name evaluation results of the two initial domain names in the initial domain name pair. Subsequently, the sum of multiple relevance contribution values ​​can be obtained as the first parameter to be used.

[0087] The above process can be characterized as follows: in, Used to represent the number of initial domain names; Used to represent the first initial domain name in multiple (specifically M) first initial domain names. The first admission weight of the first initial domain name; Used to characterize the first initial domain name among multiple first initial domain names The first admission weight of the first initial domain name; Used to characterize indicator functions: if If true, count to 1; if true, count to 1. If not true, then the count is 0; Used to characterize the first initial domain name among multiple first initial domain names Overall domain name evaluation results for the first initial domain name; Used to characterize the first initial domain name among multiple first initial domain names Overall domain name evaluation results for the first initial domain name; Used to characterize the first parameter to be used.

[0088] After obtaining multiple correlation contribution values ​​corresponding one-to-one with multiple initial domain pairs, the maximum value of multiple correlation contribution values ​​under the ideal correlation state (that is, the weight ranking result and evaluation ranking result of each initial domain pair in multiple initial domain pairs are consistent) can be obtained as the second parameter to be used.

[0089] The above process can be characterized as follows: in, Used to represent the number of initial domain names; Used to represent the first initial domain name in multiple (specifically M) first initial domain names. Overall domain name evaluation results for the first initial domain name; Used to characterize the first initial domain name among multiple first initial domain names Overall domain name evaluation results for the first initial domain name; Used to characterize the second parameter to be used.

[0090] After obtaining the first and second pending parameters, the quotient of the first and second pending parameters can be calculated as the positive correlation between the overall domain name evaluation results and the weight ranking results of the multiple initial domain name selections. This process can be characterized as follows: in, Used to characterize the first parameter to be used; Used to characterize the second parameter to be used; This is used to characterize the positive correlation between the overall domain name evaluation results and the weighted ranking results of the multiple first-selection domain names, and it will serve as a ranking reward for the multiple first-selection domain names.

[0091] In this embodiment of the disclosure, after obtaining the ranking reward of the multiple first preliminary domain names based on the first admission weight of each first preliminary domain name among the multiple first preliminary domain names, the ranking reward can be used as the overall domain name admission reward related to the first search statement. Alternatively, the overall domain name admission reward related to the first search statement can be obtained in the following ways: Based on multiple first-selected domain names and the first admission weight of each first-selected domain name, the first search result for the first search statement is obtained; For each of the multiple reward dimensions, based on the first search result, obtain the single domain name access reward related to the first search statement under the reward dimension; Based on the ranking reward and multiple individual domain admission rewards corresponding to multiple reward dimensions, the overall domain admission reward related to the first search statement is obtained.

[0092] The first search result can be obtained by using a Large Language Model (LLM). For each of the multiple initial domain names, based on the first admission weight of the initial domain name, the first search content is selected from the document content under the initial domain name. Based on the multiple first search contents that correspond one-to-one with the multiple initial domain names, the first search result for the first search statement is obtained. Here, the LLM can be a pre-trained neural network model (e.g., an autoregressive generative model with a Transformer architecture) that possesses general language knowledge, world knowledge, and domain-specific expertise (e.g., expertise in the field of computer technology).

[0093] The method for obtaining the first search result can also be found in the description of the second optional implementation method in the retrieval method embodiment, which is "based on multiple second preliminary domain names and the second admission weight of each of the multiple second preliminary domain names, to obtain the second search result for the second search statement". It will not be repeated here.

[0094] Furthermore, in this embodiment of the disclosure, the multiple reward dimensions may include at least one of the following: adoption reward dimension, review reward dimension, relevance reward dimension, and response speed reward dimension.

[0095] The single-domain admission reward related to the first search query under the adoption reward dimension can represent the degree of adoption of the first search result by the first user. Specifically, when the first search result is explicitly adopted by the first user (e.g., by adding it to favorites, clicking "helpful", "trustworthy", or giving a positive review), the single-domain admission reward related to the first search query under the adoption reward dimension can be set to "1"; when the first search result is given negative feedback by the first user (e.g., by clicking "not helpful", "untrustworthy", or giving a negative review), the single-domain admission reward related to the first search query under the adoption reward dimension can be set to "0". Under the review reward dimension, the single domain admission reward related to the first search statement can be set to "1" if the first search result is deemed compliant after review (manual or machine review); otherwise, the single domain admission reward related to the first search statement can be set to "0". The single domain admission reward related to the first search statement under the relevance reward dimension can represent the semantic relevance between the first search statement and the first search result. Specifically, the greater the semantic relevance between the first search statement and the first search result, the greater the single domain admission reward related to the first search statement under the relevance reward dimension. The single domain admission reward related to the first search statement under the response speed reward dimension can represent the speed at which the first search result is obtained. Specifically, the faster the first search result is obtained, the greater the single domain admission reward related to the first search statement under the response speed reward dimension.

[0096] As described above, in this embodiment of the disclosure, after obtaining multiple individual domain name admission rewards corresponding one-to-one with multiple reward dimensions, an overall domain name admission reward related to the first search statement is obtained based on the ranking reward and the multiple individual domain name admission rewards corresponding one-to-one with multiple reward dimensions. For example, the ranking reward and the multiple individual domain name admission rewards can be weighted and merged to obtain the overall domain name admission reward related to the first search statement. For example, the multiple reward dimensions include adoption reward dimension, review reward dimension, relevance reward dimension, and response speed reward dimension. Therefore, "weighting and merging the ranking reward and the multiple individual domain name admission rewards to obtain the overall domain name admission reward related to the first search statement" can be characterized as: in, Used to characterize the first reward fusion weight; Used to characterize the single domain admission reward related to the first search statement under the adoption reward dimension; Used to characterize the fusion weight of the second reward; Used to characterize the single domain access reward related to the first search statement under the review reward dimension; Used to characterize the third reward fusion weight; Used to represent ranking reward, specifically the aforementioned ranking reward. ; Used to characterize the fourth reward fusion weight; Used to characterize the single domain name admission reward related to the first search statement under the relevance reward dimension; Used to represent the fifth reward fusion right; Used to characterize the single domain admission reward related to the first search statement under the response speed reward dimension; This is used to characterize the overall domain admission reward related to the first search query. Here, , , , and The specific value can be set according to the application requirements of the target domain name admission model, and this embodiment does not limit it.

[0097] As described above, in this embodiment of the disclosure, after obtaining the overall domain admission reward related to the first search statement based on multiple first initially selected domain names and the first admission weight of each first initially selected domain name, the parameters of the domain admission model are updated based on the overall domain admission reward to obtain a trained domain admission model. In some optional implementations, "updating the parameters of the domain admission model based on the overall domain admission reward to obtain a trained domain admission model" may include: Based on multiple initial domain names, the overall security cost related to the first search statement is obtained; Based on the overall domain name admission reward and overall security cost, the parameters of the domain name admission model are updated to obtain the trained domain name admission model.

[0098] The overall security cost can be a multi-dimensional security cost vector.

[0099] In one example, "the overall security cost associated with the first search query based on multiple initial domain names" may include: For each of the multiple security assessment dimensions, based on multiple initial domain names, the first single security cost related to the first search statement under the security assessment dimension is obtained; The multiple individual security costs, each corresponding to one of the multiple security assessment dimensions, are taken as the overall security cost related to the first search statement.

[0100] Among them, multiple security assessment dimensions may include at least one of the following: authority assessment dimension, security assessment dimension, and audit assessment dimension.

[0101] The first single security cost related to the first search statement under the authority assessment dimension can characterize the proportion of non-authoritative domains among multiple first-selected domains, and can be used to measure whether the domain admission model deviates from the principle of "preferentially selecting authoritative domains".

[0102] In a specific example, the first single security cost associated with the first search query under the authority assessment dimension can be characterized as: in, Used to characterize the first cost weighting coefficient; Used to represent the proportion of non-authoritative domains among multiple initial domain selections; This is used to characterize the first individual security cost associated with the first search query under the authority assessment dimension. Here, The specific value can be set according to the application requirements of the target domain name admission model, and this embodiment does not limit it.

[0103] The first single security cost related to the first search statement under the security assessment dimension: When the first search result triggers the target platform's security response policy (e.g., generating a risk warning), the baseline value of the first single security cost related to the first search statement under the security assessment dimension can be set to "1"; when the first search result does not trigger the target platform's security response policy, the baseline value of the first single security cost related to the first search statement under the security assessment dimension can be set to "0".

[0104] In a specific example, the first single security cost associated with the first search query under the security assessment dimension can be characterized as: in, Used to characterize the second cost weighting coefficient; A benchmark value used to characterize the first individual security cost associated with the first search statement under the security assessment dimension; This is used to characterize the first individual security cost associated with the first search query under the security assessment dimension. Here, The specific value can be set according to the application requirements of the target domain name admission model, and this embodiment does not limit it.

[0105] The first single security cost related to the first search statement under the audit and evaluation dimension: After the first search result is audited (manual or machine audit), if the first search result is judged to be in violation, the baseline value of the first single security cost related to the first search statement under the audit and evaluation dimension is set to "1"; if the first search result is judged to be compliant, the baseline value of the first single security cost related to the first search statement under the audit and evaluation dimension is set to "0".

[0106] In a specific example, the first single security cost related to the first search statement under the audit assessment dimension can be characterized as: in, Used to characterize the third cost weighting coefficient; A benchmark value used to characterize the first individual security cost associated with the first search statement under the audit evaluation dimension; This is used to characterize the first individual security cost related to the first search statement under the audit evaluation dimension. Here, The specific value can be set according to the application requirements of the target domain name admission model, and this embodiment does not limit it.

[0107] As previously described, in this embodiment of the disclosure, after obtaining the overall security cost related to the first search statement based on multiple first initially selected domain names, the domain admission model is updated with parameters based on the overall domain admission reward and the overall security cost to obtain a trained domain admission model. In one example, "updating the parameters of the domain admission model based on the overall domain admission reward and the overall security cost to obtain a trained domain admission model" can be: constructing an overall reward target based on the overall domain admission reward and the overall security cost, and updating the parameters of the domain admission model by optimizing the overall reward target to obtain a trained domain admission model. The overall reward target may include a long-term reward target and a long-term cost target.

[0108] Based on this, in a specific example, "updating the parameters of the domain admission model based on the overall domain admission reward and overall security cost to obtain a trained domain admission model" can include: A long-term reward target is constructed based on multiple overall domain name access rewards that correspond one-to-one with multiple first search statements. Based on multiple overall security costs that correspond one-to-one with multiple first search statements, a long-term cost target is constructed. By optimizing the long-term reward and long-term cost objectives, the parameters of the domain admission model are updated to obtain a trained domain admission model.

[0109] The first search statement can be different.

[0110] In a specific example, the long-term reward objective can be the sum of multiple overall domain admission rewards (hereinafter referred to as the reward sum); correspondingly, optimizing the long-term reward objective can be maximizing the reward sum.

[0111] When the overall security cost includes the first individual security cost related to the first search statement under the authority assessment dimension, the first individual security cost related to the first search statement under the security assessment dimension, and the first individual security cost related to the first search statement under the audit assessment dimension, the long-term cost objective may include the sum of the first individual security costs related to the first search statement under multiple authority assessment dimensions (hereinafter referred to as the first cost sum), the sum of the first individual security costs related to the first search statement under multiple security assessment dimensions (hereinafter referred to as the second cost sum), and the sum of the first individual security costs related to the first search statement under multiple audit assessment dimensions (hereinafter referred to as the third cost sum). Correspondingly, optimizing the long-term cost objective may include minimizing the first cost sum, minimizing the second cost sum, and minimizing the third cost sum.

[0112] In another specific example, the long-term reward objective can be the expected value of multiple rewards obtained based on the domain admission model at different times; correspondingly, optimizing the long-term reward objective can be maximizing the expected value of multiple rewards.

[0113] Specifically, optimizing long-term reward objectives can be characterized as follows: in, Used to characterize the domain name admission model; The policy space used to characterize the domain name admission model; Used to characterize long-term reward objectives; Used to represent a specific reward sum among multiple reward sums (hereinafter referred to as the target reward sum); This is used to characterize the t-th overall domain admission reward (hereinafter referred to as the target overall domain admission reward) used when obtaining the target reward. The reward discount factor used to characterize the overall domain name admission reward can be determined based on the interval between the acquisition time of the overall domain name admission reward and the current time, and can be negatively correlated with the interval. Used to characterize the expected value of obtaining multiple rewards.

[0114] When the overall security cost includes the first individual security cost related to the first search statement under the authority assessment dimension, the first individual security cost related to the first search statement under the security assessment dimension, and the first individual security cost related to the first search statement under the audit assessment dimension, the long-term cost target can include the expected values ​​of multiple first cost sums, multiple second cost sums, and multiple third cost sums obtained based on the domain name admission model at different times. Correspondingly, optimizing the long-term cost target can include constraining the expected values ​​of multiple first cost sums below a first cost threshold, constraining the expected values ​​of multiple second cost sums below a second cost threshold, and constraining the expected values ​​of multiple third cost sums below a third cost threshold. The first cost threshold, second cost threshold, and third cost threshold can be set according to the application requirements of the target domain name admission model, and this disclosure embodiment does not impose any limitations on this.

[0115] Specifically, optimizing long-term cost objectives can be characterized as follows: in, Used to characterize the domain name admission model; Used to characterize long-term cost targets; Used to characterize a certain i-th cost sum among multiple i-th cost sums (hereinafter referred to as the target i-th cost sum); Used to characterize the t-th individual security cost used when obtaining the target cost ith; The reward discount factor used to characterize the sum of the i-th target cost can be determined based on the interval between the acquisition time of the sum of the i-th target cost and the current time, and can be negatively correlated with the interval. Used to characterize the expected value of obtaining multiple i-th cost sums; Used to characterize the cost threshold of the i-th cost.

[0116] Furthermore, it should be noted that in this embodiment of the disclosure, when setting parameters for the domain name admission model ( When updating and obtaining the trained domain admission model, the Constrained Policy Optimization (CPO) framework can be used. Specifically: Using the advantage function, we estimate the direction for policy improvement in the domain name admission model; The policy update step size of the domain admission model is limited by the relative entropy (also known as Kullback-Leibler divergence) trust region constraint.

[0117] In this embodiment of the present disclosure, based on the plurality of first preliminary domain names and the first admission weight of each of the plurality of first preliminary domain names, an overall domain name admission reward related to the first search statement can be obtained. Based on the overall domain name admission reward, the parameters of the domain name admission model are updated to obtain the trained domain name admission model. In other words, in this embodiment of the present disclosure, the domain name admission model can output differentiated sets of domain names and fine-grained admission weights for different search statements, introducing refined control over the retrieval contribution of preliminary domain names: preliminary domain names with high admission weights have higher reference priority, while preliminary domain names with low admission weights are effectively suppressed. This significantly improves the security control capability of the domain name admission model, thereby further enhancing the security of the search results obtained using the domain name admission model.

[0118] In this embodiment of the disclosure, when obtaining the overall domain admission reward related to the first search statement based on multiple first preliminary domain names and the first admission weight of each of the multiple first preliminary domain names, the ranking reward of the multiple first preliminary domain names can be obtained based on the first admission weight of each of the multiple first preliminary domain names, and the overall domain admission reward related to the first search statement can be obtained based on the ranking reward. That is to say, in this embodiment of the disclosure, by converting the admission weight into a ranking reward, a direct and clear mapping relationship can be established between the overall domain admission reward and the admission weight. In this way, the overall domain admission reward not only quantifies the impact of admission weight allocation on the retrieval contribution, but more importantly, it transforms security preferences (e.g., providing higher admission weights for authoritative domains) into a learnable target for the domain admission model, enabling the domain admission model to more sensitively perceive the impact of admission weight adjustments on the retrieval results during the training process, thereby strengthening the domain admission model's ability to identify authoritative domains and improving the security control capability of the domain admission model from the reward level.

[0119] Furthermore, when obtaining the overall domain admission reward related to the first search statement based on the ranking reward, the first search result for the first search statement can be obtained based on multiple first preliminary domains and the first admission weight of each of the multiple first preliminary domains. Then, for each of the multiple reward dimensions, based on the first search result, the individual domain admission reward related to the first search statement under that reward dimension can be obtained. Finally, based on the ranking reward and the multiple individual domain admission rewards corresponding one-to-one with the multiple reward dimensions, the overall domain admission reward related to the first search statement can be obtained. In other words, in this embodiment, by introducing multiple reward dimensions (e.g., adoption reward dimension, review reward dimension, relevance reward dimension, and response speed reward dimension), the decision-making merits of the domain admission model can be comprehensively evaluated. This prevents the domain admission model from sacrificing one dimension to gain a false capability improvement in another dimension during training, effectively preventing strategy bias caused by single-dimensional optimization.

[0120] In this embodiment of the disclosure, when updating the parameters of the domain admission model based on the overall domain admission reward to obtain a trained domain admission model, the overall security cost related to the first search statement can be obtained based on multiple first initially selected domains (for example, for each of multiple security assessment dimensions, based on multiple first initially selected domains, the first individual security cost related to the first search statement under the security assessment dimension is obtained; the multiple first individual security costs corresponding one-to-one with multiple security assessment dimensions are used as the overall security cost related to the first search statement). Based on the overall domain admission reward and the overall security cost, the parameters of the domain admission model are updated to obtain the trained domain admission model. In other words, in this embodiment of the disclosure, a quantifiable security constraint can be simultaneously applied to the parameter updates of the domain admission model by introducing an overall security cost. Specifically, the introduced overall security cost will directly reflect the comprehensive performance of domain name selection in terms of authority, security, and compliance. This will ensure that the domain name admission model is subject to clear security boundaries in the process of pursuing the maximization of overall domain name admission rewards, and will prevent the decline in authority or compliance risks caused by policy drift. It is particularly suitable for retrieval scenarios under high-risk topics such as medical diagnosis, legal interpretation, and financial advice.

[0121] Furthermore, in this embodiment, when updating the parameters of the domain admission model based on the overall domain admission reward and the overall security cost to obtain a trained domain admission model, a long-term reward target can be constructed based on multiple overall domain admission rewards corresponding one-to-one with multiple first search statements, and a long-term cost target can be constructed based on multiple overall security costs corresponding one-to-one with multiple first search statements. Then, by optimizing the long-term reward target and the long-term cost target, the parameters of the domain admission model are updated to obtain the trained domain admission model. In other words, in this embodiment, long-term reward targets and long-term cost targets can be constructed separately and jointly optimized, transforming the training of the domain admission model into a constrained long-term optimization problem to improve the training effect of the domain admission model.

[0122] Furthermore, the model training method provided in this disclosure embodiment may further include: Obtain the performance evaluation results for the trained domain admission model; If, based on the effectiveness evaluation results, it is determined that the trained domain name admission model meets the application requirements, then the trained domain name admission model will be used as the target domain name admission model.

[0123] The application requirements can be set according to the application needs of the target domain name admission model, and this disclosure does not impose any restrictions on them.

[0124] In this embodiment of the disclosure, "obtaining the performance evaluation results for the trained domain name admission model" may include: Using a trained domain admission model, a new set of domain names related to the first search query is determined based on the first feature representation; Based on the first set of domain names and the new set of domain names, the single-sample evaluation contribution value corresponding to the first search statement is obtained; Based on the contribution value of a single sample, the effect evaluation results for the trained domain name admission model are obtained.

[0125] The phrase "using the trained domain name admission model and based on the first feature representation to determine the new set of domain names related to the first search statement" can be found in step S103, which is the description of "using the domain name admission model and based on the first feature representation to determine the first set of domain names related to the first search statement", and will not be repeated here.

[0126] As described above, in this embodiment of the disclosure, after determining the set of new domain names related to the first search statement using the trained domain name admission model based on the first feature representation, a single-sample evaluation contribution value corresponding to the first search statement is obtained based on the first domain name set and the new domain name set. For example, if the first domain name set and the new domain name set are the same, a positive evaluation reward value can be obtained as the single-sample evaluation contribution factor corresponding to the first search statement; or, if the first domain name set and the new domain name set are different, a negative evaluation reward value can be obtained as the single-sample evaluation contribution factor corresponding to the first search statement. Subsequently, the single-sample evaluation contribution value can be obtained based on the overall domain name admission reward and the single-sample evaluation contribution factor. Wherein, the first domain name set and the new domain name set being the same can mean that multiple first-selected domain names in the first domain name set correspond one-to-one with multiple new-selected domain names in the new domain name set; otherwise, the first domain name set and the new domain name set can be considered different. The positive evaluation reward value can be "1"; the negative evaluation reward value can be "0".

[0127] Furthermore, it should be noted that in this embodiment of the disclosure, there can be multiple first search statements, and these multiple first search statements can be different. Therefore, in this embodiment of the disclosure, multiple single-sample evaluation contribution values ​​corresponding one-to-one with the multiple first search statements can be obtained, and based on the multiple single-sample evaluation contribution values, an effect evaluation result for the trained domain name admission model can be obtained. The above process can be characterized as follows: Where N represents the number of the first search statement; Used to characterize the single-sample evaluation contribution value corresponding to the i-th first search statement among multiple (specifically N) first search statements; Used to represent the overall domain name admission reward corresponding to the i-th first search statement among multiple first search statements; The effect correction value used to characterize the i-th first search statement among multiple first search statements can be set according to the application requirements of the target domain name admission model. This disclosure does not limit this. Used to characterize the performance evaluation results of the trained domain name admission model.

[0128] Through the above methods, in this embodiment of the disclosure, the effect evaluation results of the trained domain name admission model can be obtained. Based on the effect evaluation results, if it is determined that the trained domain name admission model meets the application requirements, the trained domain name admission model is used as the target domain name admission model. That is, in this embodiment of the disclosure, after executing steps S101, S102, S103, and S104 to obtain the trained domain name admission model, its actual effect is evaluated, and based on the effect evaluation results, a decision is made on whether to use the current trained domain name admission model as the final deployed target domain name admission model. This closed-loop mechanism of "training-evaluation-selection" effectively avoids overfitting or policy bias problems that may occur during the training process, ensuring the reliability and security of the final deployed target domain name admission model in practical applications.

[0129] In this embodiment of the disclosure, when obtaining the performance evaluation result of the trained domain name admission model, the trained domain name admission model can be used to determine the set of new domain names related to the first search statement based on the first feature representation. Based on the first domain name set and the new domain name set, a single-sample evaluation contribution value corresponding to the first search statement can be obtained (for example, if the first domain name set and the new domain name set are the same, a positive evaluation reward value is obtained as the single-sample evaluation contribution factor corresponding to the first search statement; or, if the first domain name set and the new domain name set are different, a negative evaluation reward value is obtained as the single-sample evaluation contribution factor corresponding to the first search statement; based on the overall domain name admission reward and the single-sample evaluation contribution factor, the single-sample evaluation contribution value is obtained). Then, based on the single-sample evaluation contribution value, the performance evaluation result of the trained domain name admission model is obtained. In other words, in this embodiment of the disclosure, when evaluating the trained domain name admission model, the consistency of domain name decisions before and after model training can be used as a consideration indicator, and positive / negative reward values ​​can be used to quantify and reward / penalize it. When domain name decisions are consistent before and after model training, a positive reward factor is used to amplify the overall domain name admission reward, serving as a contribution factor for single-sample evaluation. Conversely, when domain name decisions are inconsistent before and after model training, a negative reward factor is used to reduce or even eliminate the overall domain name admission reward, serving as a contribution factor for single-sample evaluation. In this way, based on the performance evaluation results, it is possible to effectively distinguish between "reasonable changes brought about by strategy optimization" and "random fluctuations caused by model instability or overfitting," thereby ensuring that only high-quality models that improve the overall domain name admission reward while maintaining model stability can pass the evaluation, thus ensuring the reliability of the target domain name admission model.

[0130] This disclosure provides a retrieval method that can be applied to electronic devices. The electronic device can be a server, workbench, mainframe computer, conventional computer, or other similar computing device. The following will be combined with... Figure 2 The flowchart shown illustrates a retrieval method provided by an embodiment of this disclosure. It should be noted that, although in Figure 2 The flowchart shown illustrates the logical order; however, in some cases, the steps shown or described in the flowchart may be performed in a different order.

[0131] Step S201: Obtain the second search statement.

[0132] The first search statement can be a word, phrase, statement, or paragraph used to express the search intent.

[0133] Step S202: Construct the second feature representation of the second retrieval statement from multiple feature representation dimensions.

[0134] Among them, multiple feature representation dimensions may include at least one of semantic feature dimension, meta-information feature dimension, risk feature dimension, and user behavior feature dimension.

[0135] Based on this, in this embodiment of the disclosure, when multiple feature representation dimensions include semantic feature dimension, meta-information feature dimension, risk feature dimension, and user behavior feature dimension, the second individual feature of the second search statement under the semantic feature dimension, the second individual feature of the second search statement under the meta-information feature dimension, the second individual feature of the second search statement under the risk feature dimension, and the second individual feature of the second search statement under the user behavior feature dimension can be obtained, and based on these second individual features, the second feature representation of the second search statement is obtained. Specifically, the second individual feature of the second search statement under the semantic feature dimension can characterize the second overall semantics of the second search statement; the second individual feature of the second search statement under the meta-information feature dimension can characterize multiple second meta-information of the second search statement; the second individual feature of the second search statement under the risk feature dimension can characterize the second overall risk level of the second search statement; and the second individual feature of the second search statement under the user behavior feature dimension can characterize the second behavioral feature of the second user who initiated the second search statement.

[0136] Step S203: Using the target domain name admission model, based on the second feature representation, determine the set of second domain names related to the second search statement.

[0137] The target domain admission model can be obtained by training the domain admission model using a model training method. Based on the second feature representation, it can determine multiple second preliminary domains related to the second search statement from the full set of domains to form a second domain set.

[0138] Step S204: Based on the second set of domain names, obtain the search results for the second search statement.

[0139] The retrieval method provided in this disclosure, when the target domain admission model is applied to the target platform for content retrieval, no longer relies solely on semantic relevance selection as in traditional solutions. Instead, after obtaining the second search statement, it constructs a second feature representation of the second search statement from multiple feature representation dimensions (e.g., semantic feature dimension, meta-information feature dimension, risk feature dimension, and user behavior feature dimension). Using the target domain admission model, based on the second feature representation, it determines a set of second domains related to the second search statement. Then, based on the set of second domains, it obtains the second search result for the second search statement. This multi-dimensional comprehensive feature representation method, while ensuring high semantic relevance between the second search result and the second search statement, also enables the target domain admission model to more comprehensively and deeply understand the reliability requirements and security risks implied by the second search statement. This proactively filters out potentially untrusted and high-risk domains during the domain selection stage, controlling the source quality of the second search result from the source and effectively improving the reliability and security of the second search result.

[0140] In this embodiment of the disclosure, after performing step S201 to obtain the second search statement, the second search statement can be classified to determine the risk category label of the second search statement. The risk category label of the second search statement can be one of a high-risk label, a medium-risk label, and a low-risk label.

[0141] In one example, when classifying the second search statement to determine its risk category label, a high-risk confidence level can be obtained to characterize the probability that the second search statement belongs to a high-risk topic. Subsequently, if the second search statement matches a set of high-risk topics and its high-risk confidence level is greater than or equal to a high-risk threshold, the risk category label of the second search statement can be determined as a high-risk label; if it matches a set of high-risk topics and its high-risk confidence level is less than a high-risk threshold, the risk category label of the second search statement can be determined as a medium-risk label; or, if it matches a set of medium-risk topics, the risk category label of the second search statement can be determined as a medium-risk label; in all other cases, the risk category label of the second search statement can be determined as a low-risk label. The high-risk topic set and the medium-risk topic set can be obtained manually, and this embodiment of the disclosure does not impose any limitations on this; the high-risk threshold can be set according to the application requirements of the search method, and this embodiment of the disclosure also does not impose any limitations on this.

[0142] Furthermore, it should be noted that in this embodiment of the disclosure, the high-risk confidence level of the second search statement can be obtained by a confidence model.

[0143] Based on the above, in this embodiment of the disclosure, step S202, namely, "constructing a second feature representation of the second retrieval statement from multiple feature representation dimensions", may include: For each feature representation dimension among multiple feature representation dimensions, construct the second single-item feature of the second retrieval statement under the feature representation dimension; Based on multiple second single features that correspond one-to-one with multiple feature representation dimensions, the second feature representation of the second retrieval statement is obtained.

[0144] As previously stated, in this embodiment of the disclosure, the multiple feature representation dimensions may include at least one of semantic feature dimensions, meta-information feature dimensions, risk feature dimensions, and user behavior feature dimensions.

[0145] Based on this, the implementation method of "constructing a second single feature of the second retrieval statement under each of the multiple feature representation dimensions" in this disclosure embodiment may include at least one of the following: (1) Multiple feature representation dimensions include semantic feature dimensions In this embodiment of the disclosure, when multiple feature representation dimensions include semantic feature dimensions, "constructing a second single feature of the second retrieval statement under each of the multiple feature representation dimensions" may include: The second search statement is vectorized to obtain its semantics. The semantics of the second statement are used as the second single feature of the second retrieval statement under the semantic feature dimension.

[0146] The semantics of the second statement can represent the position of the second retrieval statement in the semantic space.

[0147] In one example, when vectorizing the second search statement to obtain its semantics, a pre-trained BGE can be used to vectorize the second search statement to obtain its semantics, so that the semantics of the second statement can be used as the second single feature of the second search statement under the semantic feature dimension.

[0148] In this embodiment of the disclosure, when multiple feature representation dimensions include semantic feature dimensions, "constructing a second single feature of the second retrieval statement under each feature representation dimension" may also include: The risk category labels of the second search statement are vectorized to obtain the second category semantics; The second category of semantics is used as the second single feature of the second search statement under the semantic feature dimension.

[0149] The semantics of the second statement can characterize the position of the risk category label of the second retrieval statement in the semantic space.

[0150] In one example, when vectorizing the risk category label of the second search statement to obtain the second category semantics, an embedding model can be used to vectorize the risk category label of the second search statement to obtain the second category semantics, so that the second category semantics can be used as the second single feature of the second search statement under the semantic feature dimension.

[0151] Based on the above, it can be understood that in this embodiment of the disclosure, the second overall semantics of the aforementioned second search statement may include the second statement semantics and the second category semantics of the second search statement.

[0152] Through the above implementation methods, in this embodiment of the disclosure, the second search statement can be vectorized to obtain the semantics of the second statement, and the semantics of the second statement can be used as the second single feature of the second search statement under the semantic feature dimension. Simultaneously, the risk category label of the second search statement is vectorized to obtain the second category semantics, and the second category semantics can be used as the second single feature of the second search statement under the semantic feature dimension. In this way, not only can the deep semantic information of the second search statement itself be captured, but the deep semantic information of the risk category label of the second search statement can also be explicitly introduced, making the second single feature of the second search statement under the semantic feature dimension richer and more targeted. This provides a more accurate semantic dependency for the domain name selection of the target domain name admission model, thereby improving the decision accuracy of the target domain name admission model.

[0153] (2) Multiple feature representation dimensions include meta-information feature dimensions In this embodiment of the disclosure, when multiple feature representation dimensions include meta-information feature dimensions, "constructing a second single feature of the second retrieval statement under each feature representation dimension" may include: Obtain multiple second-order information from the second search query; Based on multiple second-level meta-information, the second single-item feature of the second retrieval statement under the meta-information feature dimension is obtained.

[0154] The second-level information may include at least one of the following: statement length, language type, relevant location, and real-time requirement information. Here, statement length can be the number of characters; language type can be Chinese, English, Korean, Japanese, etc.; relevant location can be the geographical location of the second user who initiated the second search statement; real-time requirement information can characterize whether the second search statement has a timeliness requirement, which can be determined by detecting whether the second search statement includes timeliness keywords such as "latest," "today," or "real-time." For example, if the second search statement includes timeliness keywords, it is determined that the second search statement has a timeliness requirement; if the second search statement does not include timeliness keywords, it is determined that the second search statement does not have a timeliness requirement.

[0155] In one example, after obtaining multiple second meta-information of the second search statement, a one-dimensional feature vector can be obtained based on each of the multiple second meta-information. The second single feature of the second search statement under the meta-information feature dimension is composed of the multi-dimensional feature vectors that correspond one-to-one with the multiple second meta-information.

[0156] Through the above implementation methods, in this embodiment of the disclosure, multiple second meta-information of the second search statement can be obtained, and based on the multiple second meta-information, a second single feature of the second search statement under the meta-information feature dimension can be obtained. The multiple second meta-information may include at least one of statement length, language type, relevant position, and real-time requirement information. In this way, structured contextual information (i.e., statement length, language type, relevant position, and real-time requirement information of the second search statement) can be added to the domain name admission model in addition to the second overall semantics, thereby improving the scenario adaptability of the domain name admission model.

[0157] (3) Multiple feature representation dimensions include risk feature dimensions In this embodiment of the disclosure, when multiple feature representation dimensions include risk feature dimensions, "constructing a second single feature of the second search statement under each feature representation dimension" may include: Based on the risk category label of the second search statement, the high-risk confidence of the second search statement, and the search complexity of the second search statement, the second content risk level of the second search statement is obtained. The second content risk level is used as the second single feature of the second search statement under the risk feature dimension.

[0158] In one example, the retrieval complexity of the second search query can be obtained in the following way: Obtain multiple features of the second search statement; For each of the multiple second statement features, the second statement feature is normalized to obtain the second feature processing result; The retrieval complexity of the second retrieval statement is obtained by weighted fusion of the processing results of multiple second feature features that correspond one-to-one with the features of multiple second statements.

[0159] The second-statement features can include at least one of the following: statement length, number of entities, sentence structure complexity, intent mixing degree, and number of time-sensitive keywords. Here, statement length can be the number of characters; entities can be specific objects, concepts, or proper nouns; sentence structure complexity can characterize whether the second-search statement includes multiple levels of nesting or conditional clauses. For example, the sentence structure complexity can be set to "1" when the second-search statement includes multiple levels of nesting or conditional clauses, and set to "0" when the second-search statement does not include multiple levels of nesting or conditional clauses; intent mixing degree can characterize the number of search intents expressed by the second-search statement; time-sensitive keywords can be "latest," "today," "real-time," etc.

[0160] In one example, when obtaining the second content risk level of a second search statement based on its risk category label, high-risk confidence level, and retrieval complexity, a content risk perception model can be used to obtain the second content risk level of the second search statement based on its risk category label, high-risk confidence level, and retrieval complexity.

[0161] In this embodiment of the disclosure, when multiple feature representation dimensions include risk feature dimensions, "constructing a second single feature of the second search statement under each feature representation dimension" may also include: Based on the frequency of occurrence of sensitive words, the coverage of sensitive words, and the location of sensitive words in the second search statement, the risk level of triggering the second sensitive word in the second search statement is obtained. The risk level triggered by the second sensitive word is used as the second single feature of the second search statement under the risk feature dimension.

[0162] Among them, the number of times a sensitive word appears can be the number of sensitive words that the second search statement hits in the sensitive word dictionary; the sensitive word coverage can represent the proportion of the number of sensitive word categories in the second search statement to the number of sensitive word categories in the sensitive word dictionary; the location of a sensitive word can represent the location information of the sensitive word in the second search statement, such as the beginning, middle, or end of the sentence, or the position offset relative to the core entity in the second search statement.

[0163] In one example, the risk level of the second sensitive word in the second search query can be obtained in the following way: The frequency of sensitive words in the second search statement is normalized to obtain the second data processing result. The sensitive word coverage of the second search statement is normalized to obtain the second coverage result; The positions of sensitive words in the second search statement are normalized to obtain the second position processing result; The results of the second data processing, the second coverage processing, and the second position processing are weighted and fused to obtain the second sensitive word trigger risk level of the second search statement.

[0164] Based on the above, it can be understood that in this embodiment of the disclosure, the second overall risk level of the aforementioned second search statement may include the second content risk level and the second sensitive word triggering risk level of the second search statement.

[0165] Through the above implementation methods, in this embodiment of the disclosure, a second content risk level of the second search statement can be obtained based on the risk category label, high-risk confidence level, and retrieval complexity of the second search statement, and this second content risk level is used as the second single feature of the second search statement under the risk feature dimension. Simultaneously, a second sensitive word trigger risk level is obtained based on the frequency of occurrence, coverage, and location of sensitive words in the second search statement, and this second sensitive word trigger risk level is used as the second single feature of the second search statement under the risk feature dimension. In other words, in this embodiment of the disclosure, the second search statement can be jointly risk-quantified from two sub-risk feature dimensions: content risk level and sensitive word trigger risk level. This considers both the macro-level thematic risk of the second search statement and captures its micro-level sensitive signals. This not only enables the target domain admission model to implement stricter domain admission control for high-risk search statements but also helps it more accurately assess retrieval risks. This allows for the implementation of differentiated security control strategies in authoritative domain selection, retrieval scope control, and evidence constraints, thereby further improving the scenario adaptability of the target domain admission model.

[0166] (4) Multiple feature representation dimensions include user behavior feature dimensions In this embodiment of the disclosure, when multiple feature representation dimensions include user behavior feature dimensions, "constructing a second single feature of the second search statement under each feature representation dimension" may include: Determine the second user who initiated the second search statement; Obtain the second user's second behavioral characteristics; Based on the second behavioral feature, the second single feature of the second search statement under the user behavior feature dimension is obtained.

[0167] The second behavioral feature may include multiple second sub-features. Here, the second sub-features may include at least one of the following: Authoritative source adoption rate: The proportion of search results from authoritative domains that were explicitly adopted by a second user within the second historical period (e.g., bookmarked, clicked "helpful", "trustworthy", or gave a positive review); Approval rate: Within the second historical period, the proportion of search results obtained from the second user's query that are deemed compliant after review (manual or machine review); Negative feedback rate: The proportion of search results obtained from a query initiated by a second user within the second historical period that received negative feedback from the second user (e.g., clicking "not helpful", "unreliable", or giving a negative rating); High-risk retrieval ratio: The proportion of retrieval statements belonging to high-risk topics among the queries initiated by the second user within the second historical period; Average interaction time: The average time a second user spends browsing the search results obtained from a query initiated by the second user within the second historical time period; Correction request ratio: In the second historical period, the proportion of search results obtained from the query statement initiated by the second user that were requested to be regenerated by the second target user.

[0168] In the above description, the second historical time period can be a time period located before the current time point and with a duration of a second preset duration. The second preset duration can be set according to the application requirements of the retrieval method, and this embodiment does not impose any limitations on it.

[0169] As mentioned above, in this embodiment of the disclosure, after determining the second user who initiates the second search statement and obtaining the second user's second behavioral features including multiple second sub-features, the second single feature of the second search statement under the user behavioral feature dimension will be obtained based on the multiple second sub-features. For example, multiple second sub-features can be concatenated to obtain the second single feature of the second search statement under the user behavioral feature dimension.

[0170] Through the above implementation methods, in this embodiment of the disclosure, a second user initiating a second search statement can be identified, and the second user's second behavioral characteristics can be obtained. Based on these second behavioral characteristics, a second single-item feature of the second search statement under the user behavioral characteristic dimension can then be obtained. In other words, in this embodiment of the disclosure, by characterizing the user's personalized behavioral patterns, the target domain name admission model's perception of the user's personalized behavioral patterns can be improved. This, in turn, enhances the user acceptance of the search results while ensuring their reliability and security, thereby improving the user experience of the target platform.

[0171] As mentioned above, in this embodiment of the disclosure, after constructing the second single feature of the second search statement under each of the multiple feature representation dimensions, the second feature representation of the second search statement is obtained based on the multiple second single features that correspond one-to-one with the multiple feature representation dimensions. For example, the multiple second single features can be concatenated to obtain the second feature representation of the second search statement.

[0172] In summary, in this embodiment of the present disclosure, when executing step S202, a second single feature of the second search statement under each of the multiple feature representation dimensions can be constructed, and a second feature representation of the second search statement can be obtained based on the multiple second single features that correspond one-to-one with the multiple feature representation dimensions. Specifically, the second single features of the second search statement under the semantic feature dimension, the second single features of the second search statement under the meta-information feature dimension, the second single features of the second search statement under the risk feature dimension, and the second single features of the second search statement under the user behavior feature dimension can be integrated into a unified second feature representation to construct a comprehensive, refined, and risk-aware input state for the target domain name admission model. This input state can support the target domain name admission model to make differentiated and safe and controllable domain name selection decisions in different search scenarios (for example, tightening or widening the domain name admission range, i.e., the search range, according to the overall risk level of the search statement), fundamentally overcoming the problem of insufficient reliability and security caused by traditional solutions relying solely on semantic relevance.

[0173] Furthermore, in this embodiment of the disclosure, step S203, namely, "using the target domain name admission model and based on the second feature representation, to determine the set of second domain names related to the second search statement," may include: Determine the complete set of domain names; Perform security pre-filtering on the full set of domain names to obtain a set of candidate domain names; Using the target domain admission model, based on the second feature representation, multiple second preliminary domains related to the second search statement are selected from the candidate domain set to form a second domain set.

[0174] As previously stated, in this embodiment of the disclosure, the full set of domain names may include all Internet domain names that the target platform has access to.

[0175] Furthermore, in this embodiment of the disclosure, the purpose of pre-filtering can be to filter out known risky sites from the entire domain set, such as malicious sites, phishing links, and websites hosting harmful content. These known risky sites can be manually labeled and stored in a site blacklist. Based on this, in this embodiment of the disclosure, the security filtering function can be characterized as follows: in, Used to represent a specific Internet domain name (hereinafter referred to as the target Internet domain name) in the full set of domain names. This is used to indicate that when a target internet domain name does not match the site blacklist, it is determined that the target internet domain name belongs to a safe site, and it is retained and stored in the candidate domain name set; This is used to characterize the determination that a target internet domain belongs to a known risky site when it hits a site blacklist, and to filter it out.

[0176] Through the above methods, in this embodiment of the disclosure, anticipated risky sites can be filtered out in advance before the domain name selection stage, effectively blocking them from entering the subsequent retrieval pipeline. On the one hand, this ensures the reliability and security of the second retrieval results from the source, avoiding potential risks caused by improper output of the target domain name admission model; on the other hand, by narrowing the processing scope of the target domain name admission model, its computational overhead on the candidate domain name set is reduced, thereby improving retrieval efficiency.

[0177] Furthermore, in this embodiment of the present disclosure, during the execution of step S203, which involves determining the set of second domain names related to the second search statement based on the second feature representation using the target domain name admission model, when determining multiple second preliminary domain names related to the second search statement from the full domain name set using the target domain name admission model based on the second feature representation to form the second domain name set, a second admission weight can also be obtained for each of the multiple second preliminary domain names. Specifically, multiple second preliminary domain names can be output using the first output head connected to the main model output end in the target domain name admission model, and the second admission weight of each of the multiple second preliminary domain names can be output using the second output head connected to the main model output end in the target domain name admission model. For each of the multiple second preliminary domain names, its second admission weight can characterize the dependence on the document content under the second preliminary domain name when obtaining the second search result for the second search statement. Specifically, for each of the multiple second preliminary domain names, the greater its second admission weight, the higher the reference priority for the document content under the second preliminary domain name when obtaining the second search results for the second search statement.

[0178] Based on the above, in this embodiment of the disclosure, step S204, namely, "obtaining the second search result for the second search statement based on the second domain name set", may include: Based on multiple second preliminary domain names and the second admission weight of each of the multiple second preliminary domain names, the second search results for the second search statement are obtained.

[0179] The second search result can be obtained through the first optional implementation method: using LLM, for each of the multiple second initial domain names, based on the second admission weight of the second initial domain name, the second search content is selected from the document content under the second initial domain name, and based on the multiple second search contents that correspond one-to-one with the multiple second initial domain names, the second search result for the second search statement is obtained.

[0180] The second search result can also be obtained through a second optional implementation method: For each of the multiple preliminary second domain names, the value prediction result of the preliminary second domain name is obtained based on the similarity between the second search statement and the document content under the preliminary second domain name, as well as the second admission weight of the preliminary second domain name. From multiple preliminary second-choice domains, select the second-choice domains whose value prediction results meet the first value requirement as intermediate domains to form an intermediate domain set; Based on the intermediate domain name set, the second search result is obtained for the second search statement.

[0181] The first value requirement can be that the value prediction results are ranked among the top K1. Here, K1 is a first quantity threshold, and its specific value can be set according to the application requirements of the retrieval method. This embodiment of the present disclosure does not impose any restrictions on it.

[0182] Based on this, in the above implementation, "for each of the multiple second preliminary domain names, based on the similarity between the second search statement and the document content under the second preliminary domain name, and the second admission weight of the second preliminary domain name, the value prediction result of the second preliminary domain name is obtained; from the multiple second preliminary domain names, the second preliminary domain name whose value prediction result meets the first value requirement is selected as the intermediate domain name to form the intermediate domain name set" can be characterized as: in, Used to characterize the second search statement; Used to represent the document content under a specific second preliminary domain (hereinafter referred to as the target second preliminary domain) among multiple second preliminary domains; Used to characterize the similarity between the second search query and the content of documents under the target second initial domain name; Used to characterize the second admission weight of the target second initial domain name; K1 is used to represent multiple second initial domain names, that is, the set of second domain names; K1 is used to represent the first quantity threshold. Used to represent a set of intermediate domain names.

[0183] In the above embodiments, for each of the multiple second preliminary domain names, based on the similarity between the second search statement and the document content under the second preliminary domain name, and the second admission weight of the second preliminary domain name, a value prediction result for the second preliminary domain name is obtained. Then, from the multiple second preliminary domain names, the second preliminary domain name whose value prediction result meets the first value requirement is selected as an intermediate domain name to form an intermediate domain name set. Finally, based on the intermediate domain name set, a second search result for the second search statement is obtained. In one example, "obtaining a second search result for the second search statement based on the intermediate domain name set" may include: For each intermediate domain in the intermediate domain set, obtain multiple individual application values ​​of the intermediate domain, and based on the multiple individual application values, obtain the comprehensive application value of the intermediate domain. From the set of intermediate domain names, select the intermediate domain names whose comprehensive application value meets the second value requirement as target domain names to form a set of target domain names; Based on the target domain name set, the second search results are obtained for the second search statement.

[0184] In this context, multiple individual application values ​​can correspond one-to-one with multiple value assessment dimensions. These multiple value assessment dimensions may include at least one of the following: semantic relevance dimension, sentence vector similarity dimension, admission weight dimension, authority level dimension, and timeliness dimension.

[0185] The single application value of intermediate domains under the semantic relevance dimension: It can characterize the semantic relevance between the second search statement and the document content under the intermediate domain; The single application value of intermediate domains under the sentence vector similarity dimension: It can represent the sentence vector similarity between the second search statement and the document content under the intermediate domain; The single application value of intermediate domains under the admission weight dimension can be represented by the second admission weight of intermediate domains; The single application value of intermediate domains under the authority level dimension: It can represent the authority level of intermediate domains; The single application value of intermediate domains under the timeliness dimension: It can characterize the publication time of document content under the intermediate domain. For example, the smaller the interval between the publication time and the current time, the greater the single application value of the intermediate domain under the timeliness dimension.

[0186] In the above description, the authority level of the intermediate domain name can be obtained through manual annotation (e.g., official website of international organizations = portal of international authoritative institutions > educational and research institutions > trusted media), and this disclosed embodiment does not limit this.

[0187] In the above example, after obtaining multiple individual application values ​​for each intermediate domain in the intermediate domain set, the comprehensive application value of the intermediate domain is obtained based on these individual application values. For example, the comprehensive application value of the intermediate domain can be obtained by weighted fusion of the multiple individual application values. Exemplarily, the multiple value assessment dimensions may include semantic relevance, sentence vector similarity, admission weight, authority level, and timeliness. Therefore, "weighted fusion of multiple individual application values ​​to obtain the comprehensive application value of the intermediate domain" can be characterized as: in, Used to characterize the first value fusion weight; Used to characterize the single application value of intermediate domain names under the semantic relevance dimension; Used to characterize the second value fusion weight; Used to characterize the single application value of intermediate domain names under the sentence vector similarity dimension; Used to characterize the third value fusion weight; It is used to characterize the single application value of intermediate domains under the admission weight dimension, that is, the second admission weight of intermediate domains; Used to characterize the fourth value fusion weight; Used to characterize the individual application value of intermediate domains under the authority level dimension; Used to characterize the fifth value fusion weight; Used to characterize the single application value of intermediate domain names in the dimension of timeliness; Used to characterize the comprehensive application value of intermediate domain names. Here, , , , and The specific value can be set according to the application requirements of the retrieval method, and this embodiment does not limit it.

[0188] Furthermore, in the above examples, the second value requirement can be that the comprehensive application value is ranked among the top K2. Here, K2 is a second quantity threshold, and its specific value can be set according to the application requirements of the retrieval method. This embodiment of the present disclosure does not impose any restrictions on this. Based on this, in the above examples, after obtaining multiple individual application values ​​for each intermediate domain in the intermediate domain set, and obtaining the comprehensive application value of the intermediate domain based on the multiple individual application values, when selecting intermediate domains whose comprehensive application value meets the second value requirement from the intermediate domain set as target domains to form a target domain set, the intermediate domains ranked among the top K2 in comprehensive application value can be selected as target domains to form a target domain set, so as to obtain the second retrieval result for the second retrieval statement based on the target domain set.

[0189] In a specific example, when obtaining the second search result for the second search statement based on the target domain set, LLM can be used to select the third search content from the document content under the target domain for each of the multiple target domains included in the target domain set, based on the second admission weight of the target domain. Based on the multiple third search contents that correspond one-to-one with the multiple target domains, the second search result for the second search statement can be obtained.

[0190] In another specific example, "obtaining the second search result based on the target domain name set for the second search statement" can include: Determine the number of authoritative domains among the multiple target domains included in the target domain set; Obtain the consistency assessment results of document content under authoritative domains in the target domain set.

[0191] In a more specific example, when obtaining the consistency assessment result of the target domain set, for each authoritative domain in the target domain set, the semantic representation result of the document content under the authoritative domain can be obtained, and the semantic concentration of multiple semantic representation results corresponding one-to-one with multiple authoritative domains in the target domain set can be obtained as the consistency assessment result of the document content under the authoritative domains in the target domain set. This process can be characterized as: in, Used to represent a subset of authoritative domains within a target domain set; A specific authoritative domain (hereinafter referred to as the target authoritative domain) is used to represent a subset of authoritative domains in the middle. Used to characterize the key content extracted from the document content under the target authoritative domain, so as to obtain the key content under the target authoritative domain; It can be cluster compactness, inverse embedding variance function, or other functions used to measure semantic concentration; This is used to characterize the consistency assessment results of document content under the authoritative domain subset, that is, the consistency assessment results of document content under the authoritative domain in the target domain set.

[0192] In the specific examples above, after determining the number of authoritative domains among the multiple target domains included in the target domain set and obtaining the consistency assessment results of the document content under the authoritative domains in the target domain set, a second search result for the second search statement can be obtained based on the target domain set, provided that the number of authoritative domains among the multiple target domains meets the requirement for the number of authoritative domains and the consistency assessment results of the document content under the authoritative domains in the target domain set meet the content consistency requirement.

[0193] The requirement for the number of authoritative domains can be that the number of authoritative domains among multiple target domains is greater than or equal to a third quantity threshold; the requirement for content consistency can be that the consistency evaluation result of the document content under the authoritative domains in the target domain set is greater than or equal to an evaluation result threshold. Here, the third quantity threshold and the evaluation result threshold can be set according to the application requirements of the retrieval method, and this disclosure embodiment does not limit them. For example, the requirement for the number of authoritative domains is that the number of authoritative domains among multiple target domains is greater than or equal to the third quantity threshold; the requirement for content consistency is that the consistency evaluation result of the document content under the authoritative domains in the target domain set is greater than or equal to the evaluation result threshold. Then, after determining the number of authoritative domains among the multiple target domains included in the target domain set and obtaining the consistency evaluation result of the document content under the authoritative domains in the target domain set, it is possible to: If the number of authoritative domains among multiple target domains is greater than or equal to the third quantity threshold, and the consistency evaluation result of the document content under the authoritative domains in the target domain set is greater than or equal to the evaluation result threshold, then based on the target domain set, the second search result for the second search statement is obtained. If the number of authoritative domains among multiple target domains is greater than zero and less than the third threshold, a supplementary prompt is generated to prompt the second user to supplement the second search statement. If the number of authoritative domains among multiple target domains is zero, generate a rejection response message.

[0194] The above process can be characterized as follows: in, Used to represent the number of authoritative domains among multiple target domains; Used for the third quantity threshold; Used to characterize the consistency assessment results of document content under authoritative domains in the target domain set; Thresholds used to characterize evaluation results; Used to represent the second search result for the second search statement; RequestClarification is used to represent supplementary suggestions; Rejcect is used to represent a rejection response suggestion; Used to characterize the search output results of the target platform in response to the second search suggestion.

[0195] Furthermore, in the specific example above, "obtaining the second search result for the second search statement based on the target domain name set" can include: Based on multiple target domains in the target domain set, the initial search results are obtained; Several key conclusions were identified from the initial search results; Based on multiple target domains, source tags are added to each key conclusion in the initial search results to obtain a second search result for the second search statement.

[0196] The initial search results can be obtained by using LLM (Limited Ledger Management) to select fourth search content from the document content of each of the multiple target domains based on the second admission weight of the target domain. Based on the multiple second search contents corresponding one-to-one with the multiple target domains, initial search results for the second search statement are obtained. These initial search results can include multiple key conclusions, which can also be determined by LLM. Here, a key conclusion can be the smallest semantic unit carrying core information in the initial search results. For example, for the second search statement "What are the dangers of hypertension?", "causes heart disease", "leads to kidney damage", and "increases the risk of stroke" in the initial search results can be considered key conclusions.

[0197] After obtaining initial search results based on multiple target domains in the target domain set, and determining multiple key conclusions from the initial search results, when adding source tags to each key conclusion in the initial search results based on multiple target domains, the similarity between each key conclusion and the document content under each target domain in the multiple target domains can be obtained. Target domains whose similarity meets the similarity requirement are selected from the multiple target domains as source tags for the key conclusions. The similarity requirement can be that the similarity is greater than or equal to a similarity threshold, and the similarity threshold can be set according to the application requirements of the search method; this disclosure does not impose any limitations on this.

[0198] Based on this, the above process can be characterized as follows: in, Used to characterize one of the key conclusions among multiple key conclusions (hereinafter referred to as the target key conclusion); Used to represent the document content under a specific target domain (hereinafter referred to as the specified target domain) among multiple target domains; Used to characterize the similarity between key conclusions of the target and the content of documents under a specified target domain; Used to characterize similarity thresholds.

[0199] Through the above methods, in this embodiment of the disclosure, a second search result for a second search statement can be obtained based on multiple second preliminary domain names and the second admission weight of each of the multiple second preliminary domain names. Specifically, for each of the multiple second preliminary domain names, a value prediction result for the second preliminary domain name can be obtained based on the similarity between the second search statement and the document content under the second preliminary domain name, as well as the second admission weight of the second preliminary domain name. Then, from the multiple second preliminary domain names, second preliminary domain names whose value prediction results meet the first value requirement are selected as intermediate domain names to form an intermediate domain name set. Based on this intermediate domain name set, a second search result for the second search statement is obtained. In other words, in this embodiment of the disclosure, a value prediction mechanism is introduced in the search stage, combining the second admission weight with semantic similarity to achieve a quantitative assessment of the comprehensive value of each second preliminary domain name. Only second preliminary domain names whose value prediction results meet the first value requirement can enter the intermediate domain name set, thereby filtering out low-value or irrelevant domain names early in the domain name selection stage, effectively reducing the computational overhead of subsequent processing stages, and improving the accuracy of the second search result for the second search statement.

[0200] In this embodiment, when obtaining the second search result for the second search statement based on the intermediate domain name set, multiple individual application values ​​corresponding one-to-one with multiple value assessment dimensions can be obtained for each intermediate domain name in the intermediate domain name set. Based on these multiple individual application values, the comprehensive application value of the intermediate domain name is obtained. Then, intermediate domain names whose comprehensive application value meets the second value requirement are selected from the intermediate domain name set as target domain names to form a target domain name set. Based on this target domain name set, the second search result for the second search statement is obtained. In other words, in this embodiment, by conducting multi-dimensional application value assessments on each intermediate domain name in the intermediate domain name set (e.g., semantic relevance dimension, sentence vector similarity dimension, admission weight dimension, authority level dimension, and timeliness dimension), the comprehensive application value of the intermediate domain name is obtained, and target domain names are selected accordingly, achieving a secondary refinement of domain name value. This ensures that the target domain name entering the final search stage is not only relevant to the second search statement but also performs excellently in terms of content quality, authority, timeliness, and other aspects, further improving the reliability and security of the second search result for the second search statement.

[0201] Furthermore, when obtaining the second search result for the second search statement based on the target domain set, the number of authoritative domains among the multiple target domains included in the target domain set can be determined, and the consistency assessment result of the document content under the authoritative domains in the target domain set can be obtained. Then, if the number of authoritative domains among the multiple target domains meets the authoritative domain quantity requirement, and the consistency assessment result of the document content under the authoritative domains in the target domain set meets the content consistency requirement, the second search result for the second search statement can be obtained based on the target domain set. In other words, in this embodiment, by further double-checking the number of authoritative domains and content consistency in the target domain set, strict evidence admission conditions can be constructed in high-risk scenarios: only when the number of authoritative domains is sufficient and the content is consistent is a second search result for the second search statement allowed; otherwise, the target platform will require the second user to supplement the second search statement or refuse to respond. This design effectively avoids model illusions and erroneous outputs caused by insufficient authoritative evidence or contradictory viewpoints, and is particularly suitable for search scenarios under high-risk topics such as medical diagnosis, legal interpretation, and financial advice.

[0202] Furthermore, when obtaining the second search results for the second search statement based on the target domain set, an initial search result can be obtained based on multiple target domains in the target domain set. Multiple key conclusions can then be determined from these initial search results. Furthermore, based on the multiple target domains, source tags can be added to each of these key conclusions in the initial search results to obtain the second search results for the second search statement. In other words, in this embodiment, the initial search results under the target domain set can be broken down into multiple key conclusions, and source tags can be added independently to each key conclusion, achieving an upgrade from "overall citation" to "fine-grained tracing." This allows the second user to intuitively see the specific source of evidence corresponding to each key conclusion, improving the user experience. Simultaneously, it provides transparent data support for subsequent auditing and dispute tracing, fundamentally improving the transparency and verifiability of the second search results.

[0203] Furthermore, in this embodiment of the disclosure, the aforementioned "obtaining the value prediction result of the second preliminary domain name based on the similarity between the second search statement and the document content under the second preliminary domain name, and the second admission weight corresponding to the second preliminary domain name" may also include: Based on the similarity between the second search query and the document content under the second initial domain name, and the second admission weight corresponding to the second initial domain name, the initial value prediction result of the second initial domain name is obtained. Obtain the trust score for the second initial domain name; Based on the initial value prediction results and trust scores, the value prediction results of the second preliminary domain name are obtained.

[0204] The trust score can be obtained based on historical statistical data related to the second initial domain name. The historical statistical data can be updated in real time and can include historical adoption rate, historical approval rate, historical negative feedback rate, etc. for the search results obtained based on the second initial domain name. This disclosure does not limit this.

[0205] Based on the similarity between the second search query and the document content under the second initial domain name, as well as the second admission weight corresponding to the second initial domain name, the initial value prediction result of the second initial domain name is obtained. After obtaining the trust score for the second initial domain name, the initial value prediction result and the trust score can be weighted and fused to obtain the value prediction result of the second initial domain name.

[0206] In this embodiment of the present disclosure, the initial value prediction result of the second preliminary domain name can be obtained based on the similarity between the second search statement and the document content under the second preliminary domain name, as well as the second admission weight corresponding to the second preliminary domain name. A trust score for the second preliminary domain name is then obtained, and the final value prediction result of the second preliminary domain name is obtained based on the initial value prediction result and the trust score. In other words, in this embodiment of the present disclosure, the stability of the second preliminary domain name's performance in historical interactions can be incorporated into the value assessment system. That is, the trust score reflects the degree of confidence of the target domain name admission model in the current value prediction result: for second preliminary domain names with high historical adoption rates, high historical approval rates, and low historical negative feedback rates, their value prediction results are amplified; for second preliminary domain names with low historical adoption rates, low historical approval rates, and high historical negative feedback rates, their value prediction results are reduced. This effectively reduces erroneous domain name selection decisions caused by data sparsity or random noise, thereby improving the accuracy and generalization ability of target domain name selection.

[0207] Furthermore, the retrieval method provided in this disclosure embodiment may further include: If the target domain name admission model is found to have security vulnerabilities, a conservative domain name admission strategy is obtained.

[0208] Among these, the conservative domain name admission strategy can be used to execute subsequent retrieval methods, specifically: Strictly adhere to the domain whitelist for domain filtering; Domain name filtering is performed according to static sorting rules with fixed thresholds; Disable the stable search mode with dynamic weight adjustment.

[0209] In some alternative implementations, "determining that the target domain name admission model has security vulnerabilities" may include: For each of the multiple security assessment dimensions, based on multiple second initial domain names, the second single security cost related to the second search statement under the security assessment dimension is obtained; If at least one of the multiple second-item security costs that correspond one-to-one with multiple security assessment dimensions meets the requirements for identifying potential risks, then the target domain name access model is determined to have security risks.

[0210] As previously stated, in this embodiment of the disclosure, the multiple security assessment dimensions may include at least one of the following: authority assessment dimension, security assessment dimension, and audit assessment dimension.

[0211] In this embodiment of the disclosure, the method for obtaining the second single security cost can be found in the model training method embodiment, which describes the "first single security cost related to the first search statement under the security assessment dimension based on multiple first initial domain names", and will not be repeated here.

[0212] Furthermore, in this embodiment of the disclosure, the existence of security vulnerabilities in the target domain name admission model can be determined based on the number of most recent second target retrieval processes on the target platform (each retrieval process is a process of "using the target platform to obtain a second retrieval result for a second retrieval statement." That is, there may be multiple second retrieval statements, and these multiple second retrieval statements can be different). The specific value of the number of second targets can be set according to the application requirements of the retrieval method, and this embodiment of the disclosure does not impose any restrictions on it.

[0213] Specifically, in this embodiment of the disclosure, for each retrieval process (hereinafter referred to as the target retrieval process) included in the second target number of retrieval processes, for each security assessment dimension (hereinafter referred to as the target security assessment dimension) among the multiple security assessment dimensions, a second single-item security cost related to the second retrieval statement used in the target retrieval process under the target security assessment dimension will be obtained. Finally, a second target number of second single-item security costs corresponding one-to-one with the second target number of retrieval processes can be obtained. Based on this, in this embodiment of the disclosure, the average of the second target number of second single-item security costs can be obtained as a hazard identification parameter corresponding to the target security assessment dimension. This process can be characterized as follows: Where T is used to characterize the number of second targets; Used to characterize the t-th second individual security cost in the second target quantity of second individual security costs; The average value of the second individual safety costs used to characterize the number of second objectives will serve as the hazard identification parameter corresponding to the target safety assessment dimension.

[0214] Furthermore, in this embodiment of the disclosure, the target safety assessment dimension is determined to meet the hazard identification requirements if the hazard identification parameter corresponding to the target safety assessment dimension is greater than its corresponding hazard threshold; or, the target safety assessment dimension is determined to meet the hazard identification requirements if the hazard identification parameter corresponding to the target safety assessment dimension is greater than the sum of its corresponding hazard threshold and a specific tolerance value. The hazard threshold and the specific tolerance value can be set according to the application requirements of the retrieval method, and this embodiment of the disclosure does not impose any restrictions on them.

[0215] The above statement, "When the hazard identification parameter corresponding to the target safety assessment dimension is greater than the sum of its corresponding hazard threshold and specific tolerance value, the target safety assessment dimension is determined to meet the hazard identification requirements," can be represented as follows: > in, Thresholds used to characterize potential hazards corresponding to the target security assessment dimensions; Used to characterize a specific tolerance value; The average value of the second individual safety costs used to characterize the number of second objectives will serve as the hazard identification parameter corresponding to the target safety assessment dimension.

[0216] Through the above methods, this embodiment of the disclosure can establish a real-time security monitoring mechanism during the actual operation of the target domain name admission model. For example, by acquiring the second individual security cost under multiple security assessment dimensions, and when at least one security assessment dimension meets the requirements for vulnerability identification, it can promptly determine that the target domain name admission model has security vulnerabilities, thereby quickly sensing the risk increase caused by policy drift, environmental changes, or abnormal input. Once triggered, it will automatically switch to a conservative domain name admission policy, thereby completing isolation and circuit breaking before the security threat spreads. In this way, a safety net security defense line can be built for the actual deployment of the target domain name admission model, effectively curbing risk events caused by uncontrolled model output, and ensuring the continuous and reliable operation of the target platform.

[0217] Please see Figure 3 This diagram illustrates an application scenario of a model training and retrieval method provided in this embodiment, which can be applied to electronic devices. The electronic device can be a server, workbench, mainframe computer, conventional computer, or other similar computing device.

[0218] In this embodiment of the disclosure, when the model training method is applied to an electronic device, the electronic device is used to: Retrieve the first search statement; A first feature representation of the first retrieval statement is constructed from multiple feature representation dimensions; wherein, the multiple feature representation dimensions include at least one of semantic feature dimension, meta-information feature dimension, risk feature dimension, and user behavior feature dimension; Using the domain name admission model, based on the first feature representation, determine the first set of domain names related to the first search statement; Based on the first set of domain names, the parameters of the domain name admission model are updated to obtain the trained domain name admission model.

[0219] In this embodiment of the disclosure, when the retrieval method is applied to an electronic device, the electronic device is used to: Obtain the second search statement; A second feature representation of the second retrieval statement is constructed from multiple feature representation dimensions; wherein, the multiple feature representation dimensions include at least one of semantic feature dimension, meta-information feature dimension, risk feature dimension, and user behavior feature dimension; Using the target domain admission model, based on the second feature representation, determine the set of second domains related to the second search statement; Based on the second set of domain names, the search results for the second search statement are obtained.

[0220] It should be noted that, in the embodiments disclosed herein, Figure 3 The application scenario diagrams shown are for illustrative purposes only and are not restrictive. Those skilled in the art can use them as a basis for their own interpretation. Figure 3 The examples may be modified in various obvious ways and / or substitutions, and the resulting technical solutions still fall within the scope of the disclosure of the embodiments of this disclosure.

[0221] To better implement the model training method, this disclosure also provides a model training apparatus that can be integrated into an electronic device. The electronic device can be a server, workbench, mainframe computer, conventional computer, or other similar computing device. The following will be combined with... Figure 4 The schematic block diagram shown illustrates a model training device 400 provided in the disclosed embodiment.

[0222] Model training device 400, including: The first statement acquisition unit 401 is used to acquire the first search statement; The first feature construction unit 402 is used to construct a first feature representation of the first retrieval statement from multiple feature representation dimensions; wherein, the multiple feature representation dimensions include at least one of semantic feature dimension, meta-information feature dimension, risk feature dimension and user behavior feature dimension; The first domain name determination unit 403 is used to determine the first domain name set related to the first search statement based on the first feature representation using the domain name admission model; The model training unit 404 is used to update the parameters of the domain admission model based on the first set of domain names, so as to obtain the trained domain admission model.

[0223] In some optional implementations, the first domain name set may include a plurality of first preliminary domain names, and each of the plurality of first preliminary domain names has a first admission weight; the model training unit 404 may be used for: Based on multiple first-selection domains and the first admission weight of each first-selection domain, the overall domain admission reward related to the first search statement is obtained; Based on the overall domain name admission reward, the parameters of the domain name admission model are updated to obtain the trained domain name admission model.

[0224] In some alternative implementations, the model training unit 404 can be used for: Based on the first admission weight of each of the multiple first-selection domains, a ranking reward is obtained for the multiple first-selection domains; Based on the ranking reward, the overall domain name access reward related to the first search query is obtained.

[0225] In some alternative implementations, the model training unit 404 can be used for: Based on multiple first-selected domain names and the first admission weight of each first-selected domain name, the first search result for the first search statement is obtained; For each of the multiple reward dimensions, based on the first search result, obtain the single domain name access reward related to the first search statement under the reward dimension; Based on the ranking reward and multiple individual domain admission rewards corresponding to multiple reward dimensions, the overall domain admission reward related to the first search statement is obtained.

[0226] In some alternative implementations, the model training unit 404 can be used for: Based on multiple initial domain names, the overall security cost related to the first search statement is obtained; Based on the overall domain name admission reward and overall security cost, the parameters of the domain name admission model are updated to obtain the trained domain name admission model.

[0227] In some alternative implementations, the model training unit 404 can be used for: For each of the multiple security assessment dimensions, based on multiple initial domain names, the first single security cost related to the first search statement under the security assessment dimension is obtained; The overall security cost includes multiple individual security costs that correspond one-to-one with multiple security assessment dimensions.

[0228] In some alternative implementations, the model training unit 404 can be used for: A long-term reward target is constructed based on multiple overall domain name access rewards that correspond one-to-one with multiple first search statements. Based on multiple overall security costs that correspond one-to-one with multiple first search statements, a long-term cost target is constructed. By optimizing the long-term reward and long-term cost objectives, the parameters of the domain admission model are updated to obtain a trained domain admission model.

[0229] In some alternative implementations, the model training apparatus 400 further includes: The performance evaluation unit can be used to obtain performance evaluation results for the trained domain admission model; The model determination unit can be used to use the trained domain name admission model as the target domain name admission model when the trained domain name admission model meets the application requirements based on the effect evaluation results.

[0230] In some alternative implementations, the effect evaluation unit can be used for: Using a trained domain admission model, a new set of domain names related to the first search query is determined based on the first feature representation; Based on the first set of domain names and the new set of domain names, the single-sample evaluation contribution value corresponding to the first search statement is obtained; Based on the contribution value of a single sample, the effect evaluation results for the trained domain name admission model are obtained.

[0231] In some alternative implementations, the effect evaluation unit can be used for: If the first set of domain names is the same as the new set of domain names, obtain the positive evaluation reward value as the single-sample evaluation contribution factor corresponding to the first search statement; Alternatively, if the first set of domain names is different from the new set of domain names, obtain the negative evaluation reward value as the single-sample evaluation contribution factor corresponding to the first search statement; The single-sample evaluation contribution value is obtained based on the overall domain name access reward and the single-sample evaluation contribution factor.

[0232] In some alternative implementations, the first feature building unit 402 may be used for: For each feature representation dimension among multiple feature representation dimensions, construct the first single feature of the first search statement under the feature representation dimension; Based on multiple first single features that correspond one-to-one with multiple feature representation dimensions, the first feature representation of the first retrieval statement is obtained.

[0233] In some optional implementations, the multiple feature representation dimensions may include semantic feature dimensions; the first feature construction unit 402 may be used for at least one of the following: The first search statement is vectorized to obtain its semantics; the semantics of the first statement are then used as the first single feature of the first search statement under the semantic feature dimension. The risk category label of the first search statement is vectorized to obtain the first category semantics; the first category semantics is used as the first single feature of the first search statement under the semantic feature dimension.

[0234] In some optional implementations, the multiple feature representation dimensions may include meta-information feature dimensions; the first feature construction unit 402 may be used for: Obtain multiple first-level information of the first search statement; wherein, the multiple first-level information includes at least one of the following: statement length, language type, relevant location, and real-time requirement information; Based on multiple first-level metadata, the first single-item feature of the first search statement under the metadata feature dimension is obtained.

[0235] In some optional implementations, the multiple feature representation dimensions may include risk feature dimensions; the first feature construction unit 402 may be used for at least one of the following: Based on the risk category label of the first search statement, the high-risk confidence of the first search statement, and the search complexity of the first search statement, the first content risk degree of the first search statement is obtained; the first content risk degree is used as the first single feature of the first search statement under the risk feature dimension. Based on the frequency of occurrence, coverage, and location of sensitive words in the first search statement, the first sensitive word trigger risk level of the first search statement is obtained; the first sensitive word trigger risk level is used as the first single feature of the first search statement under the risk feature dimension.

[0236] In some optional implementations, the multiple feature representation dimensions may include user behavior feature dimensions; the first feature construction unit 402 may be used for: Determine the first user who initiates the first search query; Obtain the first user's primary behavioral characteristics; Based on the first behavioral feature, the first single feature of the first search statement under the user behavior feature dimension is obtained.

[0237] The specific functions and examples of each unit in the model training device 400 in this embodiment can be found in the relevant descriptions of the corresponding steps in the foregoing model training method embodiments, and will not be repeated here.

[0238] To better implement the retrieval method, embodiments of this disclosure also provide a retrieval device, which can be integrated into an electronic device. The electronic device can be a server, workbench, mainframe computer, conventional computer, or other similar computing device. The following will be combined with... Figure 5 The schematic block diagram shown illustrates a retrieval device 500 provided in a disclosed embodiment.

[0239] The retrieval device 500 includes: The second statement acquisition unit 501 is used to acquire the second search statement; The second feature construction unit 502 is used to construct a second feature representation of the second retrieval statement from multiple feature representation dimensions; wherein, the multiple feature representation dimensions include at least one of semantic feature dimension, meta-information feature dimension, risk feature dimension and user behavior feature dimension; The second domain name determination unit 503 is used to determine the set of second domain names related to the second search statement based on the second feature representation using the target domain name admission model. The retrieval result acquisition unit 504 is used to obtain retrieval results for the second retrieval statement based on the second domain name set.

[0240] In some optional implementations, the second domain name set may include multiple second preliminary domain names, and each of the multiple second preliminary domain names has a second admission weight; the retrieval result acquisition unit 504 may be used for: Based on multiple second preliminary domain names and the second admission weight of each of the multiple second preliminary domain names, the second search results for the second search statement are obtained.

[0241] In some optional implementations, the retrieval result acquisition unit 504 can be used for: For each of the multiple preliminary second domain names, the value prediction result of the preliminary second domain name is obtained based on the similarity between the second search statement and the document content under the preliminary second domain name, as well as the second admission weight of the preliminary second domain name. From multiple preliminary second-choice domains, select the second-choice domains whose value prediction results meet the first value requirement as intermediate domains to form an intermediate domain set; Based on the intermediate domain name set, the second search result is obtained for the second search statement.

[0242] In some optional implementations, the retrieval result acquisition unit 504 can be used for: For each intermediate domain in the intermediate domain set, obtain multiple individual application values ​​of the intermediate domain; based on the multiple individual application values, obtain the comprehensive application value of the intermediate domain; where the multiple individual application values ​​correspond one-to-one with multiple value assessment dimensions; From the set of intermediate domain names, select the intermediate domain names whose comprehensive application value meets the second value requirement as target domain names to form a set of target domain names; Based on the target domain name set, the second search results are obtained for the second search statement.

[0243] In some optional implementations, the retrieval result acquisition unit 504 can be used for: Determine the number of authoritative domains among the multiple target domains included in the target domain set; Obtain the consistency assessment results of document content under authoritative domains in the target domain set; If the number of authoritative domains among multiple target domains meets the requirement for the number of authoritative domains, and the consistency assessment results of the document content under the authoritative domains in the target domain set meet the content consistency requirement, then a second search result for the second search statement is obtained based on the target domain set.

[0244] In some optional implementations, the retrieval result acquisition unit 504 can be used for: For each authoritative domain in the target domain set, obtain the semantic representation results of the document content under the authoritative domain; Obtain the semantic concentration of multiple semantic representation results that correspond one-to-one with multiple authoritative domains in the target domain set, and use it as the consistency assessment result of the document content under the authoritative domains in the target domain set.

[0245] In some optional implementations, the retrieval result acquisition unit 504 can be used for: Based on multiple target domains in the target domain set, the initial search results are obtained; Several key conclusions were identified from the initial search results; Based on multiple target domains, source tags are added to each key conclusion in the initial search results to obtain a second search result for the second search statement.

[0246] In some optional implementations, the retrieval result acquisition unit 504 can be used for: Based on the similarity between the second search query and the document content under the second initial domain name, and the second admission weight corresponding to the second initial domain name, the initial value prediction result of the second initial domain name is obtained. Obtain the trust score for the second initial domain name; Based on the initial value prediction results and trust scores, the value prediction results of the second preliminary domain name are obtained.

[0247] In some optional embodiments, the retrieval device 500 further includes: The strategy fallback unit can be used to obtain a conservative domain admission strategy when it is determined that there are security risks in the target domain admission model; the conservative domain admission strategy is used to execute subsequent retrieval methods.

[0248] In some alternative implementations, the policy rollback unit can be used to: For each of the multiple security assessment dimensions, based on multiple second initial domain names, the second single security cost related to the second search statement under the security assessment dimension is obtained; If at least one of the multiple second-item security costs that correspond one-to-one with multiple security assessment dimensions meets the requirements for identifying potential risks, then the target domain name access model is determined to have security risks.

[0249] In some alternative implementations, the second feature building unit 502 can be used for: For each feature representation dimension among multiple feature representation dimensions, construct the second single-item feature of the second retrieval statement under the feature representation dimension; Based on multiple second single features that correspond one-to-one with multiple feature representation dimensions, the second feature representation of the second retrieval statement is obtained.

[0250] In some optional implementations, the multiple feature representation dimensions may include semantic feature dimensions; the second feature construction unit 502 may be used for at least one of the following: The second search statement is vectorized to obtain its semantics; the semantics of the second search statement are then used as the second single feature of the second search statement under the semantic feature dimension. The risk category label of the second search statement is vectorized to obtain the second category semantics; the second category semantics is used as the second single feature of the second search statement under the semantic feature dimension.

[0251] In some optional implementations, the multiple feature representation dimensions may include meta-information feature dimensions; the second feature construction unit 502 may be used for: Obtain multiple second-level information of the second search statement; wherein, the multiple second-level information includes at least one of the following: statement length, language type, relevant location, and real-time requirement information; Based on multiple second-level meta-information, the second single-item feature of the second retrieval statement under the meta-information feature dimension is obtained.

[0252] In some optional implementations, the multiple feature representation dimensions may include risk feature dimensions; the second feature construction unit 502 may be used for at least one of the following: Based on the risk category label of the second search statement, the high-risk confidence of the second search statement, and the search complexity of the second search statement, the second content risk degree of the second search statement is obtained; the second content risk degree is used as the second single feature of the second search statement under the risk feature dimension. Based on the frequency of occurrence, coverage, and location of sensitive words in the second search statement, the second sensitive word trigger risk level of the second search statement is obtained; the second sensitive word trigger risk level is used as the second single feature of the second search statement under the risk feature dimension.

[0253] In some optional implementations, the multiple feature representation dimensions may include user behavior feature dimensions; the second feature construction unit 502 may be used for: Determine the second user who initiated the second search statement; Obtain the second user's second behavioral characteristics; Based on the second behavioral feature, the second single feature of the second search statement under the user behavior feature dimension is obtained.

[0254] In some alternative implementations, the second domain name determination unit 503 may be used for: Determine the complete set of domain names; Perform security pre-filtering on the full set of domain names to obtain a set of candidate domain names; Using the target domain admission model, based on the second feature representation, multiple second preliminary domains related to the second search statement are selected from the candidate domain set to form a second domain set.

[0255] In this embodiment of the disclosure, the specific functions and examples of each unit in the retrieval device 500 can be found in the relevant descriptions of the corresponding steps in the retrieval method embodiment, and will not be repeated here.

[0256] In this disclosed embodiment, the acquisition, storage, and application of user personal information are all known to the user and have been agreed to by the user, and all comply with the provisions of relevant laws and regulations, and do not violate public order and good morals.

[0257] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0258] Figure 6A schematic structural block diagram of an example electronic device 600 that can be used to implement embodiments of the present disclosure is shown. Electronic device 600 is intended to represent various forms of digital computers, such as in-vehicle computing devices, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic device 600 may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0259] like Figure 6 As shown, the electronic device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 602 or a computer program loaded from a storage unit 608 into a random access memory (RAM) 603. The RAM 603 may also store various programs and data required for the operation of the electronic device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0260] Multiple components in electronic device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, etc.; output unit 607, such as various types of renderers, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows electronic device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0261] The computing unit 601 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as model training methods and / or retrieval methods. For example, in some embodiments, the model training methods and / or retrieval methods may be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by the computing unit 601, one or more steps of the model training methods and / or retrieval methods described above may be performed. Alternatively, in other embodiments, the computing unit 601 may be configured as a model training method and / or a retrieval method by any other suitable means (e.g., by means of firmware).

[0262] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-chip (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transferring data and instructions to the storage system, the at least one input device, and the at least one output device.

[0263] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data optimization device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0264] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, RAM, ROM, erasable programmable read-only memory (EPROM) or flash memory, optical fibers, compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0265] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a rendering device (e.g., a cathode ray tube (CRT) renderer or a liquid crystal display (LCD)) for rendering information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices are also used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0266] The systems and technologies described herein can be implemented in computing systems that include back-end components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include front-end components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0267] A computer system can include client and server components. Clients and servers are generally located far apart and typically interact via a communication network. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, a server in a distributed system, or a server incorporating blockchain technology.

[0268] This disclosure also provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute a model training method and / or a retrieval method.

[0269] This disclosure also provides a computer program product, including a computer program that, when executed by a processor, implements a model training method and / or a retrieval method.

[0270] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure is achieved, and this is not limited herein. Furthermore, in this disclosure, relational terms such as "first," "second," "third," and "fourth" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Additionally, "multiple" in this disclosure can be understood as at least two.

[0271] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A model training method, comprising: Retrieve the first search statement; A first feature representation of the first retrieval statement is constructed from multiple feature representation dimensions; wherein, the multiple feature representation dimensions include at least one of semantic feature dimension, meta-information feature dimension, risk feature dimension, and user behavior feature dimension; Using the domain name admission model, based on the first feature representation, determine the first set of domain names related to the first retrieval statement; Based on the first set of domain names, the parameters of the domain name admission model are updated to obtain a trained domain name admission model.

2. The method according to claim 1, wherein, The first domain name set includes multiple first preliminary domain names, and each of the multiple first preliminary domain names has a first admission weight; The step of updating the parameters of the domain admission model based on the first set of domain names to obtain a trained domain admission model includes: Based on the plurality of first preliminary domain names and the first admission weight of each of the plurality of first preliminary domain names, the overall domain name admission reward related to the first search statement is obtained; Based on the overall domain name admission reward, the parameters of the domain name admission model are updated to obtain the trained domain name admission model.

3. The method according to claim 2, wherein, The overall domain admission reward related to the first search statement is obtained based on the plurality of first initially selected domain names and the first admission weight of each of the plurality of first initially selected domain names, including: Based on the first admission weight of each of the plurality of first initial domain names, the ranking reward of the plurality of first initial domain names is obtained; Based on the ranking reward, the overall domain name access reward related to the first search statement is obtained.

4. The method according to claim 3, wherein, The overall domain name access reward related to the first search statement is obtained based on the ranking reward, including: Based on the plurality of first preliminary domain names and the first admission weight of each of the plurality of first preliminary domain names, a first search result is obtained for the first search statement; For each of the multiple reward dimensions, based on the first search result, a single domain name access reward related to the first search statement under the reward dimension is obtained; Based on the ranking reward and the multiple individual domain access rewards corresponding one-to-one with the multiple reward dimensions, the overall domain access reward related to the first search statement is obtained.

5. The method according to claim 2, wherein, The step of updating the parameters of the domain admission model based on the overall domain admission reward to obtain the trained domain admission model includes: Based on the multiple first initially selected domain names, the overall security cost related to the first search statement is obtained; Based on the overall domain name admission reward and the overall security cost, the parameters of the domain name admission model are updated to obtain the trained domain name admission model.

6. The method according to claim 5, wherein, The overall security cost related to the first search statement, based on the plurality of first initially selected domain names, includes: For each of the multiple security assessment dimensions, based on the multiple first initially selected domain names, the first single security cost related to the first search statement under the security assessment dimension is obtained; The overall security cost includes multiple first individual security costs that correspond one-to-one with the multiple security assessment dimensions.

7. The method according to claim 5, wherein, The step of updating the parameters of the domain name admission model based on the overall domain name admission reward and the overall security cost to obtain the trained domain name admission model includes: A long-term reward target is constructed based on multiple overall domain name access rewards that correspond one-to-one with multiple first search statements. Based on multiple overall security costs that correspond one-to-one with the multiple first search statements, a long-term cost target is constructed. By optimizing the long-term reward objective and the long-term cost objective, the parameters of the domain name admission model are updated to obtain the trained domain name admission model.

8. The method according to claim 2, further comprising: Obtain the performance evaluation results for the trained domain name admission model; If, based on the performance evaluation results, the trained domain name admission model is determined to meet the application requirements, then the trained domain name admission model shall be used as the target domain name admission model.

9. The method according to claim 8, wherein, The process of obtaining the performance evaluation results for the trained domain name admission model includes: Using the trained domain name admission model, a new set of domain names related to the first retrieval statement is determined based on the first feature representation; Based on the first set of domain names and the new set of domain names, a single-sample evaluation contribution value corresponding to the first search statement is obtained; Based on the single-sample evaluation contribution value, the effect evaluation result for the trained domain name admission model is obtained.

10. The method according to claim 9, wherein, The step of obtaining the single-sample evaluation contribution value corresponding to the first search statement based on the first domain name set and the new domain name set includes: If the first domain name set is the same as the new domain name set, obtain a positive evaluation reward value as a single-sample evaluation contribution factor corresponding to the first search statement; Alternatively, if the first set of domain names is different from the new set of domain names, a negative evaluation reward value is obtained as a single-sample evaluation contribution factor corresponding to the first search statement; The single-sample evaluation contribution value is obtained based on the overall domain name access reward and the single-sample evaluation contribution factor.

11. The method according to any one of claims 1 to 10, wherein, The construction of the first feature representation of the first retrieval statement from multiple feature representation dimensions includes: For each of the multiple feature representation dimensions, construct a first single feature of the first retrieval statement under that feature representation dimension; Based on multiple first single-item features that correspond one-to-one with the multiple feature representation dimensions, the first feature representation of the first retrieval statement is obtained.

12. The method according to claim 11, wherein, The plurality of feature representation dimensions include semantic feature dimensions; constructing a first single feature of the first retrieval statement under each of the plurality of feature representation dimensions includes at least one of the following: The first search statement is vectorized to obtain the semantics of the first statement; the semantics of the first statement are used as the first single feature of the first search statement under the semantic feature dimension. The risk category label of the first search statement is vectorized to obtain the first category semantics; the first category semantics is used as the first single feature of the first search statement under the semantic feature dimension.

13. The method according to claim 11, wherein, The plurality of feature representation dimensions include meta-information feature dimensions; the step of constructing a single feature of the first retrieval statement under each of the plurality of feature representation dimensions includes: Obtain multiple first-level information of the first search statement; wherein, the multiple first-level information includes at least one of statement length, language type, relevant position and real-time requirement information; Based on the multiple first meta-information, the first single-item feature of the first retrieval statement under the meta-information feature dimension is obtained.

14. The method according to claim 11, wherein, The plurality of feature representation dimensions include a risk feature dimension; the construction of a first single feature of the first retrieval statement under each of the plurality of feature representation dimensions includes at least one of the following: Based on the risk category label of the first search statement, the high-risk confidence level of the first search statement, and the search complexity of the first search statement, the first content risk level of the first search statement is obtained. The first content risk level is used as the first single feature of the first search statement under the risk feature dimension; Based on the number of times sensitive words appear, the coverage of sensitive words, and the location of sensitive words in the first search statement, the first sensitive word triggering risk level of the first search statement is obtained; The risk level triggered by the first sensitive word is used as the first single feature of the first search statement under the risk feature dimension.

15. The method according to claim 11, wherein, The plurality of feature representation dimensions include user behavior feature dimensions; the step of constructing a first single feature of the first search statement under each of the plurality of feature representation dimensions includes: Determine the first user who initiated the first search statement; Obtain the first behavioral characteristics of the first user; Based on the first behavioral feature, the first single feature of the first search statement under the user behavior feature dimension is obtained.

16. A retrieval method, comprising: Obtain the second search statement; A second feature representation of the second retrieval statement is constructed from multiple feature representation dimensions; wherein, the multiple feature representation dimensions include at least one of semantic feature dimension, meta-information feature dimension, risk feature dimension, and user behavior feature dimension; Using the target domain admission model, based on the second feature representation, determine the set of second domains related to the second search statement; Based on the second set of domain names, a second search result is obtained for the second search statement.

17. The method according to claim 16, wherein, The second set of domain names includes a plurality of second preliminary domain names, and each of the plurality of second preliminary domain names has a second admission weight; The step of obtaining the second search result based on the second domain name set for the second search statement includes: Based on the plurality of second preliminary domain names and the second admission weight of each of the plurality of second preliminary domain names, a second search result is obtained for the second search statement.

18. The method according to claim 17, wherein, The second search result for the second search statement is obtained based on the plurality of second preliminary domain names and the second admission weight of each of the plurality of second preliminary domain names, including: For each of the plurality of second preliminary domain names, the value prediction result of the second preliminary domain name is obtained based on the similarity between the second search statement and the document content under the second preliminary domain name, and the second admission weight of the second preliminary domain name; From the plurality of second preliminary domain names, select the second preliminary domain name whose value prediction result meets the first value requirement as the intermediate domain name to form an intermediate domain name set; Based on the intermediate domain name set, a second search result is obtained for the second search statement.

19. The method according to claim 18, wherein, The process of obtaining a second search result for the second search statement based on the intermediate domain name set includes: For each intermediate domain in the intermediate domain set, obtain multiple individual application values ​​of the intermediate domain; based on the multiple individual application values, obtain the comprehensive application value of the intermediate domain; wherein, the multiple individual application values ​​correspond one-to-one with multiple value assessment dimensions; From the set of intermediate domain names, select intermediate domain names whose comprehensive application value meets the second value requirement as target domain names to form a set of target domain names; Based on the target domain name set, a second search result is obtained for the second search statement.

20. The method according to claim 19, wherein, The step of obtaining the second search result for the second search statement based on the target domain name set includes: Determine the number of authoritative domains among the multiple target domains included in the target domain set; Obtain the consistency assessment results of document content under authoritative domains in the target domain set; If the number of authoritative domains among the multiple target domains meets the requirement for the number of authoritative domains, and the consistency assessment result of the document content under the authoritative domains in the target domain set meets the content consistency requirement, then a second search result for the second search statement is obtained based on the target domain set.

21. The method according to claim 20, wherein, The process of obtaining the consistency assessment results of document content under authoritative domains in the target domain set includes: For each authoritative domain in the target domain set, obtain the semantic representation results of the document content under the authoritative domain; Obtain the semantic concentration of multiple semantic representation results that correspond one-to-one with multiple authoritative domains in the target domain set, and use it as the consistency evaluation result of the document content under the authoritative domains in the target domain set.

22. The method according to claim 19, wherein, The step of obtaining the second search result for the second search statement based on the target domain name set includes: Based on multiple target domains in the target domain set, initial search results are obtained; Several key conclusions were determined from the initial search results; Based on the multiple target domain names, a source tag is added to each of the multiple key conclusions in the initial search results to obtain a second search result for the second search statement.

23. The method according to claim 18, wherein, The step of obtaining the value prediction result of the second preliminary domain name based on the similarity between the second search statement and the document content under the second preliminary domain name, and the second admission weight corresponding to the second preliminary domain name, includes: Based on the similarity between the second search statement and the document content under the second initial domain name, and the second admission weight corresponding to the second initial domain name, the initial value prediction result of the second initial domain name is obtained. Obtain the trust score for the second initially selected domain name; Based on the initial value prediction results and trust scores, the value prediction results of the second initially selected domain name are obtained.

24. The method of claim 18, further comprising: If it is determined that the target domain name admission model has security risks, a conservative domain name admission strategy is obtained; wherein, the conservative domain name admission strategy is used to execute the subsequent retrieval method.

25. The method according to claim 24, wherein, The determination that the target domain name admission model has security vulnerabilities includes: For each of the multiple security assessment dimensions, based on the multiple second initial domain names, the second single security cost related to the second search statement under the security assessment dimension is obtained; If at least one of the multiple second individual security costs corresponding to the multiple security assessment dimensions meets the requirements for identifying potential risks, then the target domain name access model is determined to have security risks.

26. The method according to any one of claims 16 to 25, wherein, The construction of the second feature representation of the second retrieval statement from multiple feature representation dimensions includes: For each of the multiple feature representation dimensions, construct a second single feature of the second retrieval statement under that feature representation dimension; Based on multiple second single-item features that correspond one-to-one with the multiple feature representation dimensions, the second feature representation of the second retrieval statement is obtained.

27. The method according to claim 26, wherein, The plurality of feature representation dimensions include semantic feature dimensions; the construction of a second single feature of the second retrieval statement under each of the plurality of feature representation dimensions includes at least one of the following: The second search statement is vectorized to obtain the semantics of the second statement; the semantics of the second statement are used as the second single feature of the second search statement under the semantic feature dimension. The risk category label of the second search statement is vectorized to obtain the second category semantics; the second category semantics is used as the second single feature of the second search statement under the semantic feature dimension.

28. The method according to claim 26, wherein, The plurality of feature representation dimensions include meta-information feature dimensions; the step of constructing a second single-item feature of the second retrieval statement under each of the plurality of feature representation dimensions includes: Obtain multiple second-level information of the second search statement; wherein, the multiple second-level information includes at least one of statement length, language type, relevant location, and real-time requirement information; Based on the multiple second meta-information, the second single-item feature of the second retrieval statement under the meta-information feature dimension is obtained.

29. The method according to claim 26, wherein, The plurality of feature representation dimensions include a risk feature dimension; the construction of a second single feature of the second search statement under each of the plurality of feature representation dimensions includes at least one of the following: Based on the risk category label of the second search statement, the high-risk confidence level of the second search statement, and the search complexity of the second search statement, the second content risk level of the second search statement is obtained. The second content risk level is used as the second single feature of the second search statement under the risk feature dimension; Based on the number of times sensitive words appear, the coverage of sensitive words, and the location of sensitive words in the second search statement, the second sensitive word trigger risk level of the second search statement is obtained; The risk level triggered by the second sensitive word is used as the second single feature of the second search statement under the risk feature dimension.

30. The method of claim 26, wherein, The plurality of feature representation dimensions include user behavior feature dimensions; the step of constructing a second single feature of the second search statement under each of the plurality of feature representation dimensions includes: Determine the second user who initiated the second search statement; Obtain the second behavioral characteristics of the second user; Based on the second behavioral feature, the second single-item feature of the second search statement under the user behavior feature dimension is obtained.

31. The method according to any one of claims 16 to 25, wherein, The step of using the target domain name admission model to determine the set of second domain names related to the second search statement based on the second feature representation includes: Determine the complete set of domain names; The full set of domain names is subjected to security pre-filtering to obtain a set of candidate domain names; Using the target domain name admission model, based on the second feature representation, multiple second preliminary domain names related to the second search statement are selected from the candidate domain name set to form the second domain name set.

32. A model training device, comprising: The first statement acquisition unit is used to acquire the first search statement; The first feature construction unit is used to construct a first feature representation of the first retrieval statement from multiple feature representation dimensions; wherein, the multiple feature representation dimensions include at least one of semantic feature dimension, meta-information feature dimension, risk feature dimension and user behavior feature dimension; The first domain name determination unit is used to determine the first set of domain names related to the first retrieval statement based on the first feature representation using the domain name admission model. The model training unit is used to update the parameters of the domain admission model based on the first set of domain names to obtain a trained domain admission model.

33. A retrieval device, comprising: The second statement acquisition unit is used to acquire the second search statement; The second feature construction unit is used to construct a second feature representation of the second retrieval statement from multiple feature representation dimensions; wherein, the multiple feature representation dimensions include at least one of semantic feature dimension, meta-information feature dimension, risk feature dimension and user behavior feature dimension; The second domain name determination unit is used to determine the set of second domain names related to the second retrieval statement based on the second feature representation using the target domain name admission model. The search result acquisition unit is used to obtain a second search result for the second search statement based on the second set of domain names.

34. An electronic device, comprising: At least one processor; A memory that is communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method according to any one of claims 1 to 31.

35. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1 to 31.

36. A computer program product comprising a computer program; wherein, When the computer program is executed by a processor, it can implement the method of any one of claims 1 to 31.

Images