Intelligent search method, apparatus, device, and storage medium

Abstract

The application discloses an intelligent search method and device, equipment and a storage medium, and relates to the technical field of intelligent search. The method comprises the following steps: in response to a search request, obtaining a to-be-searched statement corresponding to the search request, and judging whether the to-be-searched statement exists in preset historical search records; if the to-be-searched statement exists in the preset historical search records, taking a search result corresponding to the to-be-searched statement in the preset historical search records as a target search result; if the to-be-searched statement does not exist in the preset historical search records, determining the semantic complexity of the to-be-searched statement, calling a preset target model corresponding to the semantic complexity to search the to-be-searched statement, and obtaining the target search result. The application can improve the intelligent search efficiency on the basis of guaranteeing the accuracy of the intelligent search result.

Description

Technical Field

[0001] This application relates to the field of intelligent search technology, and in particular to an intelligent search method, apparatus, device and storage medium. Background Technology

[0002] As enterprises deepen their digital transformation, search systems have evolved from simple retrieval tools into core entry points for accessing business services and resources. Among them, intelligent search, with its ability to understand user intent and provide accurate results, is widely used in business scenarios and has become a key infrastructure for improving search efficiency.

[0003] Currently, the most common intelligent search method is large-model search, which involves uniformly dispatching user queries to a large language model for retrieval or generative responses to achieve understanding and response to the user's natural language input. However, this traditional approach requires calling the large model for each high-frequency, explicit query, which can lead to situations where complex models are used to handle simple search tasks, thus affecting the speed at which intelligent search returns results. Furthermore, it lacks effective semantic understanding and hierarchical processing mechanisms when dealing with low-frequency, colloquial, or highly specialized long-tail queries, resulting in poor relevance of search results and low click-through rates. Therefore, the aforementioned intelligent search methods suffer from overall inefficiency.

[0004] Therefore, how to improve the efficiency of intelligent search while ensuring the accuracy of intelligent search results is an urgent problem that needs to be solved. Summary of the Invention

[0005] The main objective of this application is to provide an intelligent search method, apparatus, device, and storage medium, which aims to improve the efficiency of intelligent search while ensuring the accuracy of intelligent search results.

[0006] To achieve the above objectives, this application provides an intelligent search method, which includes: In response to a search request, the system obtains the search query corresponding to the search request and determines whether the search query exists in a preset historical search record. If the search query exists in the preset historical search records, the search result corresponding to the search query in the preset historical search records will be used as the target search result. If the search query does not exist in the preset historical search records, the semantic complexity of the search query is determined, and the preset target model corresponding to the semantic complexity is invoked to search for the search query to obtain the target search result.

[0007] In one embodiment, the step of determining the semantic complexity of the search query includes: The search query is normalized to obtain the rewritten query. The mean cosine similarity of the rewritten statement is determined, and the semantic complexity of the statement to be searched is determined based on the mean cosine similarity, wherein the mean cosine similarity is negatively correlated with the semantic complexity.

[0008] In one embodiment, the step of normalizing the search statement to obtain the rewritten statement includes: Perform a format normalization operation on the statement to be searched to obtain a normalized statement; The standardized statement is semantically split to obtain multiple semantic units; For each semantic unit, a semantic normalization operation is performed on the semantic unit to obtain the target semantic unit; The statement formed by the target semantic units corresponding to each of the aforementioned semantic units is taken as the rewritten statement.

[0009] In one embodiment, the step of performing semantic normalization on the semantic unit to obtain the target semantic unit includes: The semantic units are matched with standard semantic units in a preset corpus to obtain multiple semantic similarities; Determine the maximum semantic similarity among the given semantic similarities; If the maximum semantic similarity is greater than a preset semantic similarity threshold, the standard semantic unit corresponding to the maximum semantic similarity in the preset corpus is taken as the target semantic unit. If the maximum semantic similarity is less than or equal to a preset semantic similarity threshold, the semantic unit is taken as the target semantic unit.

[0010] In one embodiment, before the step of calling the preset target model corresponding to the semantic complexity to search the search statement and obtain the target search result, the method further includes: When the semantic complexity is greater than or equal to a preset semantic complexity threshold, the first preset model is used as the target model, wherein the first preset model refers to a model based on a deep learning architecture used to handle complex semantic understanding and generation tasks. If the semantic complexity is less than the preset semantic complexity threshold, the target model is determined based on the historical frequency of the rewritten statement.

[0011] In one embodiment, the step of determining the target model based on the historical frequency of the rewritten statement includes: Obtain historical query statements within a preset time period, and count the frequency of each historical query statement within the preset time period. The historical query statements are sorted according to their historical occurrence frequency to obtain a statement sequence. The frequency ranking of the rewritten statement in the statement sequence is determined based on the historical occurrence frequency of the rewritten statement. If the frequency ranking falls within a preset high-frequency ranking range, then the second preset model is used as the target model, wherein the second preset model refers to a retrieval model based on word frequency statistics and keyword matching. If the frequency ranking falls within a preset mid-frequency ranking range, then the third preset model is used as the target model, wherein the third preset model refers to a model that performs similarity matching based on semantic vectorization representation; If the frequency ranking falls within a preset low-frequency ranking range, then the first preset model is used as the target model.

[0012] In one embodiment, after the step of calling the preset target model corresponding to the semantic complexity to search the search statement and obtain the target search result, the method includes: Obtain the multi-dimensional features corresponding to the search request, wherein the multi-dimensional features characterize the intent profile of the search request; The target search results are reordered based on the multi-dimensional features to obtain optimized target search results.

[0013] Furthermore, to achieve the above objectives, this application also provides an intelligent search device, the intelligent search device comprising: The judgment module is used to respond to a search request, obtain the search statement corresponding to the search request, and determine whether the search statement exists in a preset historical search record; The matching module is used to select the search result corresponding to the search statement in the preset historical search records as the target search result if the search statement exists in the preset historical search records. The scheduling module is used to determine the semantic complexity of the search statement if the search statement does not exist in the preset historical search records, and then call the preset target model corresponding to the semantic complexity to search for the search statement to obtain the target search result.

[0014] In addition, to achieve the above objectives, this application also proposes an intelligent search device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the intelligent search method as described above.

[0015] In addition, to achieve the above objectives, this application also provides a storage medium, which is a computer-readable storage medium, on which a program implementing the intelligent search method is stored, and the program implementing the intelligent search method is executed by a processor to implement the steps of the intelligent search method as described above.

[0016] In addition, to achieve the above objectives, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the intelligent search method described above.

[0017] This application provides an intelligent search method. The method obtains the statement to be searched and determines whether the statement to be searched exists in the preset historical search records. If the statement to be searched exists in the preset historical search records, the target search result is returned based on the preset historical search records. If the statement to be searched does not exist in the preset historical search records, the semantic complexity is determined and the target model corresponding to the semantic complexity is called to perform the search and return the target search result.

[0018] Thus, compared to traditional solutions that directly dispatch user queries to a large language model for retrieval or generative answers, this application directly returns the target search result based on the historical search records when the search term exists. This avoids the problem of low search efficiency caused by calling complex models to process simple search tasks for high-frequency and easily understood search terms, thereby improving the efficiency of intelligent search. For search terms not existing in the historical search records, the semantic complexity is determined, and the target model corresponding to the semantic complexity is called for searching. In this way, when faced with low-frequency, colloquial, or highly specialized long-tail queries, a model matching the semantic complexity of the search term can be selected. This ensures the accuracy of intelligent search results while avoiding waste of model resources, further improving the efficiency of intelligent search and optimizing the user search experience. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the first embodiment of the intelligent search method of this application; Figure 2This is a schematic diagram illustrating the target model determination process in an embodiment of the intelligent search method of this application; Figure 3 This is a schematic diagram of a fallback mechanism involved in an embodiment of the intelligent search method of this application; Figure 4 This is a schematic diagram of the module structure of the intelligent search device of this application; Figure 5 This is a schematic diagram of the device structure of the hardware operating environment involved in the intelligent search method in this application embodiment.

[0022] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0024] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0025] Currently, the most common intelligent search method is large-model search, which involves uniformly dispatching user queries to a large language model for retrieval or generative responses to achieve understanding and response to the user's natural language input. However, this traditional approach requires calling the large model for each high-frequency, explicit query, which can lead to situations where complex models are used to handle simple search tasks, thus affecting the speed at which intelligent search returns results. Furthermore, it lacks effective semantic understanding and hierarchical processing mechanisms when dealing with low-frequency, colloquial, or highly specialized long-tail queries, resulting in poor relevance of search results and low click-through rates. Therefore, the aforementioned intelligent search methods suffer from overall inefficiency.

[0026] Therefore, how to improve the efficiency of intelligent search while ensuring the accuracy of intelligent search results is an urgent problem that needs to be solved.

[0027] This application improves intelligent search efficiency by directly returning the target search result based on the historical search records when the search query exists in the historical search records. This avoids the problem of low search efficiency caused by calling complex models to process simple search tasks for high-frequency and easily understood search queries, thus improving the efficiency of intelligent search. For search queries that do not exist in the historical search records, the semantic complexity is determined, and the target model corresponding to the semantic complexity is called for searching. In this way, when faced with low-frequency, colloquial, or highly specialized long-tail queries, a model that matches the semantic complexity of the search query can be selected. This ensures the accuracy of intelligent search results, avoids the waste of model resources, further improves the efficiency of intelligent search, and optimizes the user search experience.

[0028] It should be noted that the execution entity of the intelligent search method in various embodiments of this application can be the backend service system of the search system, or a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an intelligent search device capable of the above functions. This embodiment does not specifically limit it in this way. The following uses the backend service system of the search system as the execution entity as an example to describe this embodiment and the following embodiments.

[0029] Based on this, this application proposes an intelligent search method according to the first embodiment, please refer to... Figure 1 The intelligent search method includes steps S10 to S30: Step S10: In response to the search request, obtain the search statement corresponding to the search request, and determine whether the search statement exists in the preset historical search records.

[0030] A search request refers to the instruction information generated when a user initiates a search in the search system. This instruction information is submitted by the front-end page to the back-end service system, instructing the back-end service system to return search results that match the user's input. The search query refers to the original query content input by the user, carried in the search request; that is, the text information entered by the user in the front-end search box. Pre-set historical search records refer to historical search data pre-stored in the back-end service system. This data includes historical query statements and the corresponding search results for each historical query statement, used to directly reuse historical search results when the user initiates the same or similar query again.

[0031] The backend service system responds to the search request triggered by the user on the search page, extracts the search query entered by the user from the request, and then matches the search query with the preset historical search records to determine whether the search query exists in the historical search records; if it exists, proceed to step S20; if it does not exist, proceed to step S30.

[0032] For example, after a user enters "limit settings" in the search box of the online banking platform and clicks the search button, the front-end page generates a search request and submits it to the back-end service system. The back-end service system parses the search term "limit settings" from the request and performs a matching query in the preset historical search records to determine whether the search term exists in the historical search records.

[0033] Step S20: If the search query exists in the preset historical search records, then the search result corresponding to the search query in the preset historical search records is taken as the target search result.

[0034] The target search results refer to the set of search results that need to be returned to the user for the current search request. In scenarios where historical search records are matched, this set of results can directly reuse historical search results without having to call the retrieval model again to search.

[0035] Once the backend service system determines that the search query exists in the historical search records, it directly reads the corresponding historical search results for that query and uses them as the target search result for the current search, returning them to the user. This avoids repeatedly calling the search model and achieves millisecond-level response times. The pre-configured historical search records can be maintained through a combination of manual pre-configuration and dynamic rule generation. For example, business personnel can pre-configure the mapping relationship between high-frequency search terms and related results, and dynamically filter high-frequency search terms and update the mapping list between high-frequency search terms and click results daily based on historical search data to maintain the timeliness and accuracy of the historical search records.

[0036] For example, once the backend service system determines that there is a historical search result corresponding to "quota setting" in the historical search records, it directly returns the historical search result as the target search result to the user's front-end page. The entire search process does not require calling a large language model, and the response time is extremely short.

[0037] Step S30: If the search query does not exist in the preset historical search records, determine the semantic complexity of the search query, call the preset target model corresponding to the semantic complexity to search for the search query, and obtain the target search result.

[0038] Semantic complexity is a quantitative indicator that represents the ease or difficulty of understanding a search query at the semantic level. Higher semantic complexity indicates greater difficulty in understanding and generating the query's semantics, requiring a more powerful model for semantic processing. The target model refers to a pre-defined model determined based on the semantic complexity or historical frequency of the search query and used to perform the search task. This includes a first pre-defined model, a second pre-defined model, and a third pre-defined model.

[0039] After determining that the search query does not exist in the historical search records, the backend service system first determines the semantic complexity of the search query. Then, based on the semantic complexity, it matches the corresponding target model, calls the target model to search for the search query, obtains the target search result, and returns it to the user.

[0040] For example, after determining that the query "cross-border trade finance full-process solution" does not exist in the historical search records, the backend service system determines that its semantic complexity is high, calls the preset model used to handle complex semantic understanding and generation tasks to search, obtains the target search result and returns it.

[0041] Thus, compared to traditional solutions that directly dispatch user queries to a large language model for retrieval or generative answers, this application's embodiment directly returns the target search result based on the historical search record when the search query exists in the historical search record. This avoids the problem of low search efficiency caused by calling complex models to process simple search tasks for high-frequency and easily understood search queries, thereby improving intelligent search efficiency. For search queries that do not exist in the historical search record, the semantic complexity is determined, and the target model corresponding to the semantic complexity is called for searching. In this way, when faced with low-frequency, colloquial, or highly specialized long-tail queries, a model that matches the semantic complexity of the search query can be selected. This ensures the accuracy of intelligent search results while avoiding waste of model resources, further improving intelligent search efficiency and optimizing the user search experience.

[0042] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description and will not be repeated hereafter. Based on this, the step of determining the semantic complexity of the search statement in step S30 may include: Step S301: Perform a normalization operation on the statement to be searched to obtain the rewritten statement; Step S302: Determine the average cosine similarity of the rewritten statement, and determine the semantic complexity of the statement to be searched based on the average cosine similarity, wherein the average cosine similarity is negatively correlated with the semantic complexity.

[0043] Normalization refers to the standardization of search queries in terms of format and semantics, converting queries with different expressions into a standardized form to facilitate subsequent semantic analysis. The rewritten statement is the statement obtained after performing normalization on the search query.

[0044] Mean cosine similarity refers to the average cosine similarity between each semantic unit in the rewritten sentence and a predefined set of common simple word vectors. It is used to measure the commonness of the sentence in the semantic space. Mean cosine similarity is negatively correlated with semantic complexity; that is, the smaller the mean cosine similarity, the greater the semantic distance between the sentence and common simple words, and the higher the semantic complexity.

[0045] The backend service system performs normalization operations on the search statement, including format normalization and semantic normalization, to obtain the rewritten statement. It then uses a word vector model to obtain the vector representation of the rewritten statement, calculates the average cosine similarity between the rewritten statement and a preset set of common simple word vectors, and determines the semantic complexity of the search statement based on the average cosine similarity.

[0046] For example, when a user inputs the search term "derivatives", the backend service system performs a normalization operation on the term and obtains the rewritten term "derivatives". For the financial term "derivatives", its mean cosine similarity with the vectors of common and simple words such as "stocks" and "funds" is low, so its semantic complexity is determined to be high.

[0047] In this embodiment, step S301 may include: Step S3011: Perform a format normalization operation on the statement to be searched to obtain the normalized statement; Step S3012: Semantically decompose the standardized statement to obtain multiple semantic units; Step S3013: For each semantic unit, perform semantic normalization operation to obtain the target semantic unit; Step S3014: The statement formed by the target semantic units corresponding to each semantic unit is taken as the rewritten statement.

[0048] Format normalization refers to the process of standardizing the format information in a search query, including but not limited to correcting spelling errors, unifying full-width and half-width characters, and filtering sensitive words and meaningless characters. A semantic unit is the smallest semantically independent component obtained by breaking down a normalized query according to semantic boundaries. A target semantic unit is the standard semantic unit obtained after performing semantic normalization on the semantic units.

[0049] The backend service system performs a format standardization operation on the search query. It uses a preset regular expression and terminology database to match and check the search query, correcting spelling errors, standardizing input format, and filtering sensitive words to obtain a standardized query. The standardized query is then semantically split into multiple semantic units. Semantic standardization is performed on each semantic unit to convert each semantic unit into the corresponding target semantic unit.

[0050] For example, when a user enters "Check out corporate loans" (including extra spaces), the backend service system performs a format standardization operation to obtain the standardized statement "corporate loans". The standardized statement "corporate loans" is then broken down into multiple semantic units, namely "corporate" and "loan". After semantic standardization, the semantic unit "loan" is transformed into the target semantic unit "credit loan". The target semantic units corresponding to each semantic unit are combined in their original order, and the resulting statement is used as the rewritten statement. That is, the target semantic units "corporate" and "credit loan" are combined in their original order to form the rewritten statement "corporate credit loan".

[0051] In this embodiment, step S3013 may include: Step S401: Match the semantic unit with each standard semantic unit in the preset corpus to obtain multiple semantic similarities; Step S402: Determine the maximum semantic similarity among all the semantic similarities; Step S403: If the maximum semantic similarity is greater than the preset semantic similarity threshold, the standard semantic unit corresponding to the maximum semantic similarity in the preset corpus is taken as the target semantic unit. Step S404: If the maximum semantic similarity is less than or equal to a preset semantic similarity threshold, the semantic unit is taken as the target semantic unit.

[0052] A pre-built corpus refers to a pre-constructed set of standard semantic units containing multiple standard semantic units. Semantic similarity refers to the degree of similarity between a semantic unit and a standard semantic unit at the semantic level, used to measure the semantic consistency between the two. The maximum semantic similarity is the corresponding semantic similarity with the highest numerical value among all semantic similarities. The pre-set semantic similarity threshold refers to a pre-defined threshold used to determine whether a semantic unit is sufficiently similar to a standard semantic unit.

[0053] The backend service system matches each semantic unit with a standard semantic unit in a preset corpus, calculates the semantic similarity between the semantic unit and each standard semantic unit, and obtains multiple semantic similarities. It then determines the maximum value among these multiple semantic similarities and uses this maximum value as the maximum semantic similarity. The maximum semantic similarity is compared with a preset semantic similarity threshold. If the maximum semantic similarity is greater than the preset semantic similarity threshold, it is determined that the semantic unit has been matched with a sufficiently similar standard semantic unit in the preset corpus, and this standard semantic unit is used as the target semantic unit. If the maximum semantic similarity is less than or equal to the preset semantic similarity threshold, it is determined that the semantic unit cannot be matched with a sufficiently similar standard semantic unit in the preset corpus, and this semantic unit is directly used as the target semantic unit.

[0054] For example, if the semantic similarity between the semantic unit "loan" and the standard semantic unit "credit loan" in the preset corpus is 0.9, which is greater than the preset semantic similarity threshold of 0.85, then "credit loan" will be used as the target semantic unit.

[0055] In this embodiment, before the step of "calling the preset target model corresponding to the semantic complexity to search for the search statement" in step S30, the intelligent search method of this application further includes: Step S501: If the semantic complexity is greater than or equal to a preset semantic complexity threshold, the first preset model is used as the target model, wherein the first preset model refers to a model based on a deep learning architecture used to handle complex semantic understanding and generation tasks. Step S502: If the semantic complexity is less than the preset semantic complexity threshold, determine the target model based on the historical frequency of the rewritten statement.

[0056] The preset semantic complexity threshold refers to a pre-set threshold used to determine whether the search query needs to invoke the model with the highest semantic processing capability. The first preset model refers to a model based on a deep learning architecture used to handle complex semantic understanding and generation tasks, such as LLM (Large Language Model).

[0057] The backend service system compares the determined semantic complexity with a preset semantic complexity threshold. If the semantic complexity is greater than or equal to the preset semantic complexity threshold, it indicates that the semantics of the search statement is relatively complex, and the backend service system uses the first preset model as the target model. If the semantic complexity is less than the preset semantic complexity threshold, the target model is determined based on the frequency of occurrence of the rewritten statement in historical queries.

[0058] For example, if the semantic complexity of the search term "cross-border trade finance full-process solution" is 0.9, which is greater than the preset semantic complexity threshold of 0.8, then the first preset model (large language model) is used as the target model; if the semantic complexity of the search term "limit setting" is 0.3, which is less than the preset semantic complexity threshold of 0.8, then the target model is determined based on the historical frequency of occurrence of "limit setting".

[0059] In this embodiment, step S502 may include: Step S601: Obtain historical query statements within a preset time period, and count the historical occurrence frequency of each historical query statement within the preset time period.

[0060] Step S602: Sort each of the historical query statements according to their historical occurrence frequency to obtain a statement sequence, and determine the frequency ranking of the rewritten statement in the statement sequence based on the historical occurrence frequency of the rewritten statement. Step S603: If the frequency ranking belongs to a preset high-frequency ranking range, then the second preset model is used as the target model, wherein the second preset model refers to a retrieval model based on word frequency statistics and keyword matching. Step S604: If the frequency ranking belongs to the preset mid-frequency ranking range, then the third preset model is used as the target model, wherein the third preset model refers to a model that performs similarity matching based on semantic vectorization representation. Step S605: If the frequency ranking belongs to a preset low frequency ranking range, then the first preset model is used as the target model.

[0061] The preset time period refers to a pre-defined time range for statistically analyzing query frequency, such as the past month. Historical query statements refer to all queries initiated by users within this preset time period. Historical occurrence frequency refers to the number of times each historical query statement appeared within the preset time period, counted as a single statement. Statement sequence refers to the sequence of historical query statements arranged from highest to lowest historical occurrence frequency. Frequency ranking refers to the position of the rewritten statement within the statement sequence.

[0062] The second preset model refers to a retrieval model based on word frequency statistics and keyword matching, such as a traditional retrieval model based on the BM25 (BestMatch 25) algorithm. The preset high-frequency ranking range refers to a pre-defined frequency ranking range used to determine if a query is a high-frequency query, such as the top 20%. The third preset model refers to a model based on semantic vectorization for similarity matching, such as the BGE-M3 (Bidirectional General Embedding Model-M3) vector retrieval model based on a semantic vector knowledge base for semantic matching. The preset mid-frequency ranking range refers to a pre-defined frequency ranking range used to determine if a query is a mid-frequency query, such as the middle 50%. The preset low-frequency ranking range refers to a pre-defined frequency ranking range used to determine if a query is a low-frequency query, such as the bottom 30%.

[0063] The backend service system obtains historical query statements within a preset time period, counts the occurrence frequency of each historical query statement within the preset time period, and obtains the historical occurrence frequency of each historical query statement. The historical query statements are then sorted according to their historical occurrence frequencies to obtain a statement sequence. The frequency ranking of the rewritten statement within the statement sequence is determined based on its historical occurrence frequency. It is then determined whether the frequency ranking of the rewritten statement falls within a preset high-frequency ranking range. If so, the second preset model is used as the target model. If the frequency ranking of the rewritten statement falls within a preset medium-frequency ranking range, the third preset model is used as the target model. If the frequency ranking of the rewritten statement falls within a preset low-frequency ranking range, the first preset model is used as the target model, leveraging its deep semantic understanding and generation capabilities to handle low-frequency complex queries.

[0064] For example, the backend service system analyzes historical query statements from the past month. "Limit setting" appears 1200 times, "corporate credit loan" appears 200 times, and "cross-border trade finance full-process solution" appears 5 times. "Limit setting" ranks high in frequency (top 20%), "corporate credit loan" ranks in the middle (middle 50%), and "cross-border trade finance full-process solution" ranks low (bottom 30%). If "Limit setting" falls within the high-frequency range (top 20%), the second preset model (traditional retrieval model based on the BM25 algorithm) is used as the target model. If "corporate credit loan" falls within the medium-frequency range (middle 50%), the third preset model (vector retrieval model) is used as the target model. If "cross-border trade finance full-process solution" falls within the low-frequency range (bottom 30%), the first preset model (large language model) is used as the target model.

[0065] For example, such as Figure 2 The diagram illustrates the target model determination process. Based on the rewritten statement, the semantic complexity of the rewritten statement is determined, and it is judged whether this semantic complexity is less than a preset semantic complexity threshold. If it is not less than the preset semantic complexity threshold, the first preset model is used as the target model. If it is less than the preset semantic complexity threshold, the historical occurrence frequencies are counted and sorted. Based on the sorting results, the frequency ranking of historical query statements matching the rewritten statement is determined, and then it is determined which frequency ranking range the frequency ranking belongs to. If it belongs to the high-frequency ranking range, the second preset model is used as the target model; if it belongs to the medium-frequency ranking range, the third preset model is used as the target model; if it belongs to the low-frequency ranking range, the first preset model is used as the target model.

[0066] In this embodiment, after the step of "calling the preset target model corresponding to the semantic complexity to search the search statement and obtain the target search result" in step S30, the intelligent search method of this application further includes: Step S701: Obtain the multi-dimensional features corresponding to the search request, wherein the multi-dimensional features represent the intent profile of the search request; Step S702: Based on the multi-dimensional features, the target search results are reordered to obtain optimized target search results.

[0067] Multi-dimensional features refer to the multi-dimensional feature information that characterizes the intent profile of a search request, including but not limited to user-dimensional features, product features, user behavior features, sequence features, and query features. An intent profile is a set of features used to describe the user's intent behind a search request. Re-ranking processing refers to the process of personalized re-ranking of target search results based on multi-dimensional features. Optimized target search results refer to the search results obtained after re-ranking.

[0068] The backend service system obtains multi-dimensional features corresponding to the search request and constructs an intent profile of the search request. Based on the multi-dimensional features, the system re-ranks the target search results through a preset feature fusion and ranking model to obtain optimized target search results and returns them to the user.

[0069] For example, the backend service system obtains user identity features, historical click behavior features, and query statement features corresponding to the current search request to form multi-dimensional features; based on the multi-dimensional feature fusion model, the target search results are personalized and re-ranked, with the ranking taking about 20 milliseconds, and relevant and diverse result sets are returned to the user.

[0070] This embodiment of the application obtains a rewritten statement by normalizing the statement to be searched. The semantic complexity of the statement to be searched is determined based on the mean cosine similarity of the rewritten statement. When the semantic complexity is greater than or equal to a preset semantic complexity threshold, a first preset model is invoked for searching. When the semantic complexity is less than the preset threshold, a hierarchical scheduling is further performed based on the frequency of the rewritten statement in historical query records: a second preset model is invoked for high frequency, a third preset model for medium frequency, and the first preset model for low frequency. After obtaining the target search results, the target search results are reordered based on multi-dimensional features. Thus, this embodiment of the application achieves precise matching between query complexity and model processing capabilities through a two-layer decision-making mechanism combining semantic complexity determination and frequency-based hierarchical scheduling. This avoids low-complexity queries consuming high-cost model resources, and optimizes search result ranking through multi-dimensional feature reordering for personalized search scenarios, further improving the efficiency of intelligent search and user experience.

[0071] Based on the first or second embodiment of this application, in the third embodiment of this application, the content that is the same as or similar to the first or second embodiment described above can be referred to the above description and will not be repeated hereafter. Furthermore, the intelligent search method of this application also optimizes the generation and maintenance of historical search records, the training of the semantic understanding model, and the fallback processing for retrieval degradation, specifically including: Step A10: Generate and maintain the preset historical search records.

[0072] The generation and maintenance of historical search records can employ a combination of manual pre-configuration and dynamic rule generation for preservation. Manual pre-configuration refers to a mechanism where business personnel pre-configure the mapping relationship between search terms and related results. Business personnel can perform fine-grained configurations for core business scenarios, including configuring the mapping between search terms and related menus and results, and setting up quick jump entry points. The dynamic rule generation mechanism refers to a mechanism that statistically analyzes historical search data according to a preset period, automatically generating and updating hotspot rules. The hotspot rule list is a mapping list composed of high-frequency search terms and corresponding high-frequency click results, used to provide direct search results for high-frequency queries.

[0073] When generating historical search records, the backend service system supports two main mechanisms. First, it supports a manual pre-configuration mechanism, allowing business personnel to configure the mapping relationship between search terms and related results through the operations management module. During configuration, the system verifies the validity of the configuration item names and paths to prevent system failures caused by configuration errors. Second, the system runs a scheduled task daily to statistically analyze historical data: first, it cleans user behavior data, filtering special character sets and meaningless search terms; then, it statistically analyzes search records within a preset time period, selecting the top-ranking high-frequency search keywords to form a high-frequency search keyword list. Next, based on user search result click behavior data, it sorts the results from high to low, selecting the top-ranking high-frequency menu click items to construct a keyword-to-click menu mapping list. Finally, it integrates the high-frequency search keyword list and the keyword-to-click menu mapping list to initially generate a candidate list of hot topic rules. This candidate list undergoes a second manual review and evaluation to select the final hot topic rule list. Only rules that pass the second manual review become the final hot topic rule list and officially take effect.

[0074] For example, the backend service system compiles search records from the past month and filters out the top 800 high-frequency search keywords such as "limit setting", "transfer and remittance", and "financial products" to form a high-frequency search keyword list. It also selects the top 10 high-frequency menu click items based on user click behavior. The two are linked and integrated to generate a candidate list of hot topic rules. In this list, any high-frequency search keyword corresponds to 10 high-frequency menu click items. After a second manual review, 250 high-frequency search keywords and the 10 high-frequency menu click items corresponding to any high-frequency search keyword are selected. The selected results serve as the final hot topic rule list and officially take effect.

[0075] Step A20: Train a semantic matching model to perform semantic matching and understanding of the search query.

[0076] The initial corpus refers to the basic set of text data used to train the semantic matching model. Data cleaning refers to the process of removing garbled text, irrelevant symbols, and duplicate content from the initial corpus. Manually labeled data refers to data generated by manual semantic annotation of samples, used to improve the training effect of the model. The semantic matching model refers to a model based on a deep learning architecture used to understand and match the semantics of query statements, such as BERT (Bidirectional Encoder Representations from Transformers).

[0077] When training the semantic matching model, the backend service system first processes users' historical search records within a preset time period, extracting typos, similar words, and the last-level names of various menu paths in the business system as initial corpus. Next, the initial corpus undergoes data cleaning, removing garbled characters, irrelevant symbols, and duplicate content. To further improve model performance, manually labeled data is added, and the semantic matching model is pre-trained based on the initial corpus and manually labeled data, resulting in a basic model with fundamental semantic understanding capabilities. In practical applications, a preset semantic similarity threshold is set. When a rewrite request is received, the semantic matching model uses text similarity technology to quickly search the corpus for results that are semantically similar or related to the request text, thus initially and accurately identifying the user's intent and providing guidance for subsequent deeper understanding.

[0078] For example, the backend service system compiles typos, similar words, and the last-level names of more than 3,000 menu paths from the user search records of the online banking platform over the past month into an initial corpus. After cleaning, it combines manually labeled data to pre-train the BERT model to obtain the basic model. In practical applications, a semantic similarity threshold of 0.85 is set to quickly find standard semantic units with similar semantics when a rewrite request arrives.

[0079] Step A30: Configure a fallback retrieval mechanism for each target model.

[0080] Search degradation refers to a mechanism that automatically switches to a backup search channel to complete the search when the target model channel experiences an anomaly (such as model service unavailability or response timeout). Fallback ranking refers to a mechanism that uses a pre-defined simplified ranking strategy to sort search results when re-ranking processes malfunction or cannot be effectively executed, such as fallback ranking based on click-through rate.

[0081] The backend service system is configured with a fallback retrieval mechanism for each target model channel: a database-based fallback retrieval channel is configured for the retrieval model based on word frequency statistics and keyword matching (i.e., the second preset model); an ES (Elasticsearch)-based fallback retrieval channel is configured as a fallback for the model based on semantic vectorization for similarity matching (i.e., the third preset model); and an ES-based fallback retrieval channel is configured for the first preset model based on a deep learning architecture to cope with sudden changes in model service. At the same time, a fallback re-ranking strategy based on click-through rate is configured during the result ranking stage. When multi-dimensional feature fusion re-ranking cannot be effectively executed, the target search results are ranked based on click-through rate to ensure the stability and availability of the search results output.

[0082] For example, such as Figure 3 The diagram illustrates the fallback mechanism. When the target model is the first preset model, if the first preset model fails, an Elasticsearch (ES) search is performed. When the target model is the second preset model, if the second preset model fails, a database search is performed. When the target model is the third preset model, if the third preset model fails, an ES search is performed. After the target model is searched, multi-dimensional feature fusion training and re-ranking are performed based on the obtained target search results. If the re-ranking of multi-dimensional feature fusion training fails, a re-ranking process based on click volume is performed.

[0083] This application embodiment employs a management mechanism that combines manual pre-configuration with dynamic rule generation to preserve the freshness of historical search records. It also pre-trains the semantic matching model based on initial corpus and manually labeled data, and configures a fallback mechanism for retrieval degradation for each target model channel. This improves the timeliness and accuracy of historical search records, enhances the accuracy of semantic understanding of long-tail queries, and ensures the continued availability of search services when models malfunction, further optimizing the overall efficiency and stability of intelligent search.

[0084] This application also provides an intelligent search device, please refer to... Figure 4 The intelligent search device includes: The judgment module 10 is used to respond to a search request, obtain the search statement corresponding to the search request, and determine whether the search statement exists in a preset historical search record; The matching module 20 is used to use the search result corresponding to the search statement in the preset historical search record as the target search result if the search statement exists in the preset historical search record. The scheduling module 30 is used to determine the semantic complexity of the search statement if the search statement does not exist in the preset historical search records, and to call the preset target model corresponding to the semantic complexity to search for the search statement and obtain the target search result.

[0085] Optionally, the scheduling module 30 is further configured to: The search query is normalized to obtain the rewritten query. The mean cosine similarity of the rewritten statement is determined, and the semantic complexity of the statement to be searched is determined based on the mean cosine similarity, wherein the mean cosine similarity is negatively correlated with the semantic complexity.

[0086] Optionally, the scheduling module 30 is further configured to: Perform a format normalization operation on the search query to obtain a normalized query; The standardized statement is semantically split to obtain multiple semantic units; For each semantic unit, a semantic normalization operation is performed on the semantic unit to obtain the target semantic unit; The statement formed by the target semantic units corresponding to each of the aforementioned semantic units is taken as the rewritten statement.

[0087] Optionally, the scheduling module 30 is further configured to: The semantic units are matched with standard semantic units in a preset corpus to obtain multiple semantic similarities; Determine the maximum semantic similarity among the given semantic similarities; If the maximum semantic similarity is greater than a preset semantic similarity threshold, the standard semantic unit corresponding to the maximum semantic similarity in the preset corpus is taken as the target semantic unit. If the maximum semantic similarity is less than or equal to a preset semantic similarity threshold, the semantic unit is taken as the target semantic unit.

[0088] Optionally, the scheduling module 30 is further configured to: When the semantic complexity is greater than or equal to a preset semantic complexity threshold, the first preset model is used as the target model, wherein the first preset model refers to a model based on a deep learning architecture used to handle complex semantic understanding and generation tasks. If the semantic complexity is less than the preset semantic complexity threshold, the target model is determined based on the historical frequency of the rewritten statement.

[0089] Optionally, the scheduling module 30 is further configured to: Obtain historical query statements within a preset time period, and count the frequency of each historical query statement within the preset time period. The historical query statements are sorted according to their historical occurrence frequency to obtain a statement sequence. The frequency ranking of the rewritten statement in the statement sequence is determined based on the historical occurrence frequency of the rewritten statement. If the frequency ranking falls within a preset high-frequency ranking range, then the second preset model is used as the target model, wherein the second preset model refers to a retrieval model based on word frequency statistics and keyword matching. If the frequency ranking falls within a preset mid-frequency ranking range, then the third preset model is used as the target model, wherein the third preset model refers to a model that performs similarity matching based on semantic vectorization representation; If the frequency ranking falls within a preset low-frequency ranking range, then the first preset model is used as the target model.

[0090] Optionally, the intelligent search device further includes a reordering module, which is used for: Obtain the multi-dimensional features corresponding to the search request, wherein the multi-dimensional features characterize the intent profile of the search request; The target search results are reordered based on the multi-dimensional features to obtain optimized target search results.

[0091] The intelligent search device provided in this application, employing the intelligent search method described in the above embodiments, can solve the technical problem of low efficiency in intelligent search. Compared with the prior art, the beneficial effects of the intelligent search device provided in this application are the same as those of the intelligent search method provided in the above embodiments, and will not be repeated here.

[0092] This application provides an intelligent search device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the intelligent search method in the above embodiments.

[0093] The following is for reference. Figure 5 It shows a structural schematic diagram of an intelligent search device suitable for implementing embodiments of this application. Figure 5 The smart search device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0094] like Figure 5As shown, the intelligent search device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in the read-only memory 1002 or a program loaded from the storage device 1003 into the random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the intelligent search device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus 1005.

[0095] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments of this application.

[0096] The intelligent search device provided in this application, employing the intelligent search method described in the above embodiments, can solve the technical problem of low efficiency in intelligent search. Compared with the prior art, the beneficial effects of the intelligent search device provided in this application are the same as those of the intelligent search method provided in the above embodiments, and other technical features in this intelligent search device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0097] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0098] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0099] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the intelligent search method in the above embodiments.

[0100] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0101] The aforementioned computer-readable storage medium may be included in the intelligent search device; or it may exist independently and not be assembled into the intelligent search device.

[0102] The aforementioned computer-readable storage medium carries one or more programs. When the aforementioned one or more programs are executed by an intelligent search device, the intelligent search device: in response to a search request, obtains the search statement corresponding to the search request, and determines whether the search statement exists in a preset historical search record; if the search statement exists in the preset historical search record, the search result corresponding to the search statement in the preset historical search record is used as the target search result; if the search statement does not exist in the preset historical search record, the semantic complexity of the search statement is determined, and a preset target model corresponding to the semantic complexity is invoked to search for the search statement to obtain the target search result.

[0103] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server.

[0104] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings.

[0105] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0106] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described intelligent search method, thereby solving the technical problem of low efficiency in intelligent search. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the intelligent search method provided in the above embodiments, and will not be repeated here.

[0107] This application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the intelligent search method described above.

[0108] The computer program product provided in this application can improve the efficiency of intelligent search while ensuring the accuracy of intelligent search results. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the intelligent search method provided in the above embodiments, and will not be repeated here.

[0109] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent scope of this application.

Claims

1. An intelligent search method, characterized in that, The intelligent search method includes: In response to a search request, the system obtains the search query corresponding to the search request and determines whether the search query exists in a preset historical search record. If the search query exists in the preset historical search records, the search result corresponding to the search query in the preset historical search records will be used as the target search result. If the search query does not exist in the preset historical search records, the semantic complexity of the search query is determined, and the preset target model corresponding to the semantic complexity is invoked to search for the search query to obtain the target search result.

2. The intelligent search method as described in claim 1, characterized in that, The step of determining the semantic complexity of the search query includes: The search query is normalized to obtain the rewritten query. The mean cosine similarity of the rewritten statement is determined, and the semantic complexity of the statement to be searched is determined based on the mean cosine similarity, wherein the mean cosine similarity is negatively correlated with the semantic complexity.

3. The intelligent search method as described in claim 2, characterized in that, The step of normalizing the search statement to obtain the rewritten statement includes: Perform a format normalization operation on the search query to obtain a normalized query; The standardized statement is semantically split to obtain multiple semantic units; For each semantic unit, a semantic normalization operation is performed on the semantic unit to obtain the target semantic unit; The statement formed by the target semantic units corresponding to each of the aforementioned semantic units is taken as the rewritten statement.

4. The intelligent search method as described in claim 3, characterized in that, The step of performing semantic normalization on the semantic unit to obtain the target semantic unit includes: The semantic units are matched with standard semantic units in a preset corpus to obtain multiple semantic similarities; Determine the maximum semantic similarity among the given semantic similarities; If the maximum semantic similarity is greater than a preset semantic similarity threshold, the standard semantic unit corresponding to the maximum semantic similarity in the preset corpus is taken as the target semantic unit. If the maximum semantic similarity is less than or equal to a preset semantic similarity threshold, the semantic unit is taken as the target semantic unit.

5. The intelligent search method as described in claim 2, characterized in that, Before the step of calling the preset target model corresponding to the semantic complexity to search the search statement and obtain the target search result, the method further includes: When the semantic complexity is greater than or equal to a preset semantic complexity threshold, the first preset model is used as the target model, wherein the first preset model refers to a model based on a deep learning architecture used to handle complex semantic understanding and generation tasks. If the semantic complexity is less than the preset semantic complexity threshold, the target model is determined based on the historical frequency of the rewritten statement.

6. The intelligent search method as described in claim 5, characterized in that, The step of determining the target model based on the historical frequency of the rewritten statement includes: Obtain historical query statements within a preset time period, and count the frequency of each historical query statement within the preset time period. The historical query statements are sorted according to their historical occurrence frequency to obtain a statement sequence. The frequency ranking of the rewritten statement in the statement sequence is determined based on the historical occurrence frequency of the rewritten statement. If the frequency ranking falls within a preset high-frequency ranking range, then the second preset model is used as the target model, wherein the second preset model refers to a retrieval model based on word frequency statistics and keyword matching. If the frequency ranking falls within a preset mid-frequency ranking range, then the third preset model is used as the target model, wherein the third preset model refers to a model that performs similarity matching based on semantic vectorization representation; If the frequency ranking falls within a preset low-frequency ranking range, then the first preset model is used as the target model.

7. The intelligent search method as described in claim 1, characterized in that, After the step of calling the preset target model corresponding to the semantic complexity to search the search query and obtaining the target search result, the method includes: Obtain the multi-dimensional features corresponding to the search request, wherein the multi-dimensional features characterize the intent profile of the search request; The target search results are reordered based on the multi-dimensional features to obtain optimized target search results.

8. An intelligent search device, characterized in that, The intelligent search device includes: The judgment module is used to respond to a search request, obtain the search statement corresponding to the search request, and determine whether the search statement exists in a preset historical search record; The matching module is used to select the search result corresponding to the search statement in the preset historical search records as the target search result if the search statement exists in the preset historical search records. The scheduling module is used to determine the semantic complexity of the search statement if the search statement does not exist in the preset historical search records, and then call the preset target model corresponding to the semantic complexity to search for the search statement to obtain the target search result.

9. An intelligent search device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the intelligent search method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the intelligent search method as described in any one of claims 1 to 7.

Images