An adaptive retrieval strategy optimization method and system based on artificial intelligence

Abstract

This invention discloses an adaptive retrieval strategy optimization method and system based on artificial intelligence. The method includes acquiring user query statements, processing and filtering the corpus, matching with resource libraries to obtain query resource libraries, performing multi-perspective semantic expansion and hierarchical fusion to obtain three-level expanded queries, calculating the query-resource coupling degree of each query resource library and determining an initial retrieval strategy, executing the initial retrieval strategy to calculate the retrieval utility and cumulative regret value of each query resource library, determining an adaptive filtering threshold to divide positive feedback and negative feedback query resource libraries, performing characterization analysis to obtain feedback content and update the three-level expanded queries, iteratively updating the retrieval strategy for the negative feedback query resource library, and outputting an optimized retrieval strategy when no negative feedback query resource library exists. This method not only improves data retrieval efficiency and accuracy but also has good interpretability and can be directly applied to adaptive retrieval strategy optimization systems.

Description

Technical Field

[0001] This invention relates to the field of information retrieval technology, and in particular to an adaptive retrieval strategy optimization method and system based on artificial intelligence. Background Technology

[0002] With the rapid development of the internet and digital technologies, global data resources are growing explosively, and information retrieval technology has become a core means for users to obtain high-value knowledge from massive amounts of heterogeneous data. Especially in an environment where multiple heterogeneous resource repositories such as academic literature, standards and specifications, patent technologies, and industry reports coexist, the efficiency and quality of knowledge acquisition are directly determined by how to accurately match resource repositories according to user intent and dynamically optimize retrieval strategies.

[0003] However, existing retrieval technologies suffer from the following main drawbacks: First, query expansion methods are mostly simple expansions, lacking pre-match verification with the structural characteristics of the target resource database, resulting in a semantic gap between search terms and the resource database; second, retrieval strategy configuration usually adopts fixed rules or a single strategy, failing to make differentiated dynamic configurations based on the degree of coupling between query intent and resource database structure, leading to insufficient recall in resource databases with scattered topics and wasted precision in resource databases with well-structured structures; third, existing solutions only consider system-side indicators when evaluating retrieval utility, failing to explicitly incorporate deep user behavioral feedback into the utility model, lacking closed-loop correction based on users' real needs; fourth, fixed strategy selection thresholds fail to consider the cumulative bias of strategy selection, resulting in the permanent elimination of high-bias resource databases and severe strategy oscillations during long-term system operation. Therefore, this invention proposes an adaptive retrieval strategy optimization method and system based on artificial intelligence. By constructing a closed-loop technical framework of "three-level progressive query expansion - query-resource coupling degree pre-adaptation - dual-driven screening of retrieval utility and accumulated regret value - content similarity topology feedback correction - iterative negative feedback to positive", it not only solves the problem of the disconnect between traditional query expansion and resource structure, but also enables the retrieval system to continuously approach the optimal strategy in long-term operation by introducing an accumulated regret value correction mechanism and a content similarity-driven feedback content generation strategy. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive retrieval strategy optimization method and system based on artificial intelligence.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: The system obtains user query statements, performs corpus processing and filtering to obtain first-level extended queries, and calculates the representation similarity between the first-level extended queries and the resource databases to match the query resource databases. Perform multi-perspective semantic expansion on the first-level extended query to obtain the second-level extended query, and perform hierarchical fusion to obtain the third-level extended query; Calculate the query intent diffusion of the three-level extended query and the resource structure entropy of each query resource library. Calculate the query-resource coupling degree based on the query intent diffusion, resource structure entropy and representation similarity. Generate an initial retrieval strategy based on the query-resource coupling degree. The initial search strategy is executed to obtain search results and search feedback. The search utility and cumulative regret value of each query resource database are calculated, and an adaptive filtering threshold is determined based on the search utility and cumulative regret value. Based on the adaptive filtering threshold, the query resource library is divided into positive feedback query resource library and negative feedback query resource library. The feedback content is obtained by characterization analysis of each positive feedback query resource library, and the feedback content is used to update the three-level extended query. The retrieval strategy and retrieval utility of the negative feedback query resource library are updated based on the updated level 3 extended query. Based on the updated retrieval utility, the negative feedback query resource library is adjusted to a positive feedback query resource library. The retrieval strategy of the negative feedback query resource library is iteratively updated until there is no negative feedback query resource library. The optimized retrieval strategy is then output. The search results include search quality, response latency, and result diversity. The search feedback includes browsing time, click frequency, browsing depth, and content extraction volume.

[0006] Furthermore, the method for matching and querying the resource library includes: The user query statement is obtained and standardized, and then input into a pre-trained language encoder to obtain the original semantic vector; the user query statement is a colloquial search expression; the standardization process includes word segmentation, stop word filtering, and part-of-speech tagging; The user query is input into a pre-trained language model to generate candidate extended queries, and then input into a pre-trained language encoder to obtain candidate semantic vectors. The pre-trained language model embeds a table linking colloquial search terms with standardized domain search terms. The candidate extended queries are standardized domain search terms or search phrases. The original semantic vector is aligned with the candidate semantic vector using a semantic alignment constraint function, and the anchoring score is obtained by semantic anchoring metric on the candidate expanded query. The expression is as follows: ; ; in For semantic alignment constraint functions, For the user's query statement, For the first One candidate extended query, The original semantic vector, For the first candidate semantic vectors As a balance factor, For Jensen-Shannon divergence, for The probability distribution of the original semantic vector along the semantic dimension. for The probability distribution of candidate expanded queries in the semantic dimension. Expanding the query for candidates Anchored rating, For bandwidth parameters; Candidate extended queries with anchor scores greater than the anchor threshold are selected as first-level extended queries; Calculate the representation similarity between the first-level extended query vector and the representation feature vectors of each resource library, and select the resource libraries whose representation similarity is greater than the first representation similarity threshold. The content features are in multi-hot encoding form, including resource category, data format, resource purpose and applicable scenario.

[0007] Furthermore, the method for obtaining the three-level extended query includes: Fine-grained expansion variants are obtained by performing multi-perspective semantic expansion on each first-level expansion query using a semantic diversity enhancement function. Based on the expansion query optimization objective, these fine-grained expansion variants are selected as second-level expansion queries, expressed as follows: ; ; in For first-level extended queries The multi-perspective semantic extension results for semantic neighborhood The Query a neighbor's address For first-level extended queries Secondary extended queries, For first-level extended queries The Among the fine-grained extended variants, To estimate the relevance score for the retrieval, for and semantic alignment loss, For correlation-diversity balance weights, Alignment penalty coefficient; The first-level and second-level extended queries are hierarchically merged to obtain the third-level extended query; the third-level extended query includes a corresponding extended query weight.

[0008] Furthermore, the method for generating the initial retrieval strategy includes: The third-level extended query is input into the pre-trained language model encoder to obtain a semantic embedding vector. The semantic centroid closest to the semantic embedding vector is retrieved in the embedding space, and the query intent diffusion is calculated, expressed as: ; in For the set of semantic embedding vectors for query intent diffusion, for, for, To expand the number of queries to level three, For third-level extended queries The semantic centroid that is closest in the embedding space. For bandwidth parameters; Perform topic clustering on the document sets of each query resource library to obtain topic clusters and the corresponding document proportions, and calculate resource structure entropy; Calculate the query-resource repository global semantic similarity between the third-level extended query and each query repository. Concatenate the query intent diffusion, resource structure entropy, and query-resource repository global semantic similarity into a feature vector. Output the query-resource coupling degree through a learnable coupling degree network, expressed as: ; in For three-level extended queries and query resource database Query-resource coupling This is the coupling weight vector. To query the global semantic similarity of the resource repository, To query the resource library Resource structure entropy, It is the bias vector; Based on the query-resource coupling degree between the three-level extended query and each query resource library, the retrieval strategy for the corresponding query resource library is selected, and all retrieval strategies are combined to generate the initial retrieval strategy.

[0009] Furthermore, the method for determining the adaptive screening threshold includes: Execute the initial search strategy to obtain search results and search feedback, and calculate the search result score and search feedback coefficient respectively, as shown in the following expressions: ; in To query the retrieval feedback coefficient of the resource library, For browsing time, Standard browsing time, For click frequency, Standard click frequency, , For feedback weights, This is the hierarchy sensitivity coefficient. To browse levels, For memory sensitivity, Content extraction volume; The product of the search result score and the search feedback coefficient is taken as the search utility. The maximum search utility value in all query resource databases is taken as the utility benchmark. The search utility is normalized to obtain the normalized search utility. The difference between 1 and the normalized search utility is taken as the instantaneous regret value. All instantaneous regret values ​​are accumulated to obtain the cumulative regret value. The adaptive filtering threshold is calculated, and the expression is: ; in For adaptive filtering thresholds, This represents the average retrieval utility across all query databases. For regret adjustment coefficient, This is the cumulative regret value for all queried repositories. This represents the highest cumulative regret value in history. This is a smoothing factor.

[0010] Furthermore, the method for obtaining feedback content includes: Query resource repositories whose retrieval utility is greater than the adaptive filtering threshold are classified as positive feedback query resource repositories, and those whose retrieval utility is less than the threshold are classified as negative feedback query resource repositories. Calculate the representation similarity between each positive feedback query resource library. Two positive feedback query resource libraries with a representation similarity greater than the second representation similarity threshold are regarded as a centralized resource pair, and otherwise, they are regarded as a decentralized resource pair. The feedback content is formed by taking the intersection of the centralized resource pairs representing the content features and the non-intersection of the decentralized resource pairs representing the content features. The feedback content is then mapped to the third-level extended query, and the third-level extended query is updated.

[0011] 8. Furthermore, the method for iteratively updating the retrieval strategy of the negative feedback query resource library includes: The query intent diffusion is recalculated based on the updated level 3 extended query. The representation similarity between the updated level 3 extended query and the negative feedback query resource library is calculated. The query-resource coupling degree of each feedback query resource library is updated accordingly. The retrieval strategy of the negative feedback query resource library is readjusted based on the query-resource coupling degree. If the level 3 extended query is not updated, directly fine-tune the proportion of the retrieval strategy used in the negative feedback query resource library; The new retrieval strategy is implemented to obtain the retrieval results and retrieval feedback of the negative feedback query resource library. The retrieval utility of the negative feedback query resource library is updated. If the updated retrieval utility is greater than the original adaptive filtering threshold, the corresponding negative feedback query resource library is divided into a positive feedback query resource library and the third-level extended query is updated. Otherwise, the original division results are maintained. If, in the same round of iteration, no negative feedback query resource library is classified as a positive feedback query resource library, then the proportion of the retrieval strategy used in the negative feedback query resource library is continuously fine-tuned until at least one new positive feedback query resource library is added, and then the third-level extended query is updated. Repeat the previous iteration update steps until there is no negative feedback querying the resource library, then output the retrieval strategy used by each positive feedback query to form an optimized retrieval strategy.

[0012] Secondly, an adaptive retrieval strategy optimization system based on artificial intelligence includes: The query resource library matching module is used to obtain user query statements, process and filter the corpus to obtain first-level extended queries, and calculate the representation similarity between the first-level extended queries and each resource library to match the query resource library. Extended query acquisition module: used to perform multi-perspective semantic expansion on the first-level extended query to obtain the second-level extended query, and perform hierarchical fusion to obtain the third-level extended query; The retrieval strategy generation module is used to calculate the query intent diffusion of the three-level extended query and the resource structure entropy of each query resource library. It calculates the query-resource coupling degree based on the query intent diffusion, resource structure entropy and representation similarity, and generates the initial retrieval strategy based on the query-resource coupling degree. The query resource database partitioning module executes the initial retrieval strategy to obtain retrieval results and retrieval feedback, calculates the retrieval utility and cumulative regret value of each query resource database, determines the adaptive filtering threshold based on the retrieval utility and cumulative regret value, and partitions the query resource database into positive feedback query resource databases and negative feedback query resource databases based on the adaptive filtering threshold. The iterative update module is used to perform characterization analysis on each positive feedback query resource library to obtain feedback content, update the third-level extended query with the feedback content, update the retrieval strategy and retrieval utility of the negative feedback query resource library based on the updated third-level extended query, adjust the negative feedback query resource library to a positive feedback query resource library based on the updated retrieval utility, iteratively update the retrieval strategy of the negative feedback query resource library until there is no more negative feedback query resource library, and output the optimized retrieval strategy.

[0013] The beneficial effects of this invention are: This invention is an adaptive retrieval strategy optimization method and system based on artificial intelligence. Compared with the prior art, this invention has the following technical advantages: This invention, through the combined effect of a three-level progressive query expansion system, a semantic alignment constraint function, and a semantic anchoring scoring function, ensures that standardized search terms maintain the anchoring of the original query intent while being pre-adapted to the content distribution of the target resource library, significantly improving the matching degree between expanded queries and the content of heterogeneous resource libraries; This invention integrates query intent diffusion, resource structure entropy, and global semantic similarity to form a continuous query-resource coupling degree, replacing the coarse mode of parameter adjustment. It configures differentiated recall layer strategy combinations for different resource libraries, solving the problem of strategy mismatch for the same query on multi-source heterogeneous resource libraries. This invention breaks through the limitations of traditional methods that rely solely on system-side indicators, and innovatively introduces in-depth user behavior feedback, enabling strategy optimization to shift from a "single-objective of system performance" to a collaborative optimization of "dual-objectives of system and user". This invention introduces a cumulative regret value adjustment mechanism in the design of the screening threshold to ensure that resource libraries with large strategy deviations are not permanently eliminated, and provides an iterative positive channel for negative feedback resource libraries based on feedback content correction, which significantly enhances the stability and robustness of the system's online learning. This invention analyzes the content similarity topology of the positive feedback resource library, integrates the content intersection (common needs) of centralized resource pairs with the content non-intersection (differentiated needs) of low-dispersion resource pairs into structured feedback content, and performs content topology-driven query correction, thereby improving the completeness of the system's semantic coverage of multi-source heterogeneous resource libraries. This invention constructs a closed-loop mechanism of "negative feedback identification → query correction → strategy reconfiguration → utility re-evaluation → iterative positive conversion". By continuously fine-tuning the strategy ratio, it ensures that the negative feedback resource library is transformed into a positive feedback resource library within a finite number of iterations, forming a theoretical convergence guarantee for retrieval strategy optimization. It outputs a differentiated optimized retrieval strategy combination covering all heterogeneous resource libraries, realizing adaptive dynamic optimization of the entire retrieval stack. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of an adaptive retrieval strategy optimization method based on artificial intelligence according to the present invention. Detailed Implementation

[0015] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0016] The present invention provides an adaptive retrieval strategy optimization method and system based on artificial intelligence, comprising the following steps: like Figure 1 As shown, this embodiment includes the following steps: The system obtains user query statements, performs corpus processing and filtering to obtain first-level extended queries, and calculates the representation similarity between the first-level extended queries and the resource databases to match the query resource databases. Perform multi-perspective semantic expansion on the first-level extended query to obtain the second-level extended query, and perform hierarchical fusion to obtain the third-level extended query; Calculate the query intent diffusion of the three-level extended query and the resource structure entropy of each query resource library. Calculate the query-resource coupling degree based on the query intent diffusion, resource structure entropy and representation similarity. Generate an initial retrieval strategy based on the query-resource coupling degree. The initial search strategy is executed to obtain search results and search feedback. The search utility and cumulative regret value of each query resource database are calculated, and an adaptive filtering threshold is determined based on the search utility and cumulative regret value. Based on the adaptive filtering threshold, the query resource library is divided into positive feedback query resource library and negative feedback query resource library. The feedback content is obtained by characterization analysis of each positive feedback query resource library, and the feedback content is used to update the three-level extended query. The retrieval strategy and retrieval utility of the negative feedback query resource library are updated based on the updated level 3 extended query. Based on the updated retrieval utility, the negative feedback query resource library is adjusted to a positive feedback query resource library. The retrieval strategy of the negative feedback query resource library is iteratively updated until there is no negative feedback query resource library. The optimized retrieval strategy is then output. The search results include search quality, response latency, and result diversity. The search feedback includes browsing time, click frequency, browsing depth, and content extraction volume.

[0017] In this embodiment, the method for matching and querying the resource library includes: The user query statement is obtained and standardized, and then input into a pre-trained language encoder to obtain the original semantic vector; the user query statement is a colloquial search expression; the standardization process includes word segmentation, stop word filtering, and part-of-speech tagging; The user query is input into a pre-trained language model to generate candidate extended queries, and then input into a pre-trained language encoder to obtain candidate semantic vectors. The pre-trained language model embeds a table linking colloquial search terms with standardized domain search terms. The candidate extended queries are standardized domain search terms or search phrases. The original semantic vector is aligned with the candidate semantic vector using a semantic alignment constraint function, and the anchoring score is obtained by semantic anchoring metric on the candidate expanded query. The expression is as follows: ; ; in For semantic alignment constraint functions, For the user's query statement, For the first One candidate extended query, The original semantic vector, For the first candidate semantic vectors As a balance factor, For Jensen-Shannon divergence, for The probability distribution of the original semantic vector along the semantic dimension. for The probability distribution of candidate expanded queries in the semantic dimension. Expanding the query for candidates Anchored rating, For bandwidth parameters; Candidate extended queries with anchor scores greater than the anchor threshold are selected as first-level extended queries; Calculate the representation similarity between the first-level extended query vector and the representation content feature vector of each resource library, and select the resource libraries whose representation similarity is greater than the first representation similarity threshold; the content features are in multi-hot encoding form, including resource category, data format, resource purpose and applicable scenario; In actual evaluation, the use of colloquial queries input by users in natural language was considered. Taking the search term "how to prevent concrete from spoiling easily in seawater" as an example, we performed word segmentation, stop word filtering, and part-of-speech tagging. After word segmentation, we obtained the word sequence "how," "make," "concrete," "in," "seawater," "inside," "not," "easy," and "spoil." After removing "how," "make," "in," "inside," "not," and "easy" through stop word filtering, we retained the core search terms. Part-of-speech tagging confirmed that "concrete" is a noun (the main building material), "seawater" is a noun (corrosive environment), and "spoil" is a verb. This yielded a standardized query statement. "Concrete, seawater, bad"; Standardize query statements Using a thesaurus of keywords in the field of construction engineering as a pre-input, five candidate extended queries (normalized domain keywords) are generated using a pre-trained language model: "chloride ion diffusion coefficient", "resistance to sulfate attack", "durability of concrete in marine environment", "carbonation depth", and "steel corrosion". Candidate extended queries and standardized query statements are input into a pre-trained language encoder to obtain candidate semantic vectors and original semantic vectors. Anchoring alignment is performed through a semantic alignment constraint function with a balance factor of 0.5, a bandwidth parameter of 1, and an anchoring threshold of 0.65. The anchoring scores of each candidate extended query are calculated to be 0.685, 0.412, 0.792, 0.202, and 0.644. Based on an anchoring threshold of 0.6, "chloride ion diffusion coefficient", "concrete durability in marine environment", and "steel corrosion" are selected as first-level extended queries. Feature vectors representing the content of each resource library are extracted. Taking reference library A (academic literature / text / research reference / marine engineering), reference library B (standards / structured tables / compliance review / construction engineering), and reference library C (patent technology / text + image / technological innovation / materials science) as examples, the cosine similarity between the first-level extended query vector and the feature vector representing the content is calculated. Among them, the cosine similarity between "chloride ion diffusion coefficient" and resource libraries A / B / C is 0.82 / 0.45 / 0.38, and the cosine similarity between "marine environment concrete resistance" and "chloride ion diffusion coefficient" is 0.82 / 0.45 / 0.38. The cosine similarity between "durability" and resource libraries A / B / C is 0.91 / 0.756 / 0.52, and the cosine similarity between "steel corrosion" and resource libraries A / B / C is 0.78 / 0.4 / 0.85. According to the first characterization similarity threshold of 0.7, "chloride ion diffusion coefficient" matches resource library A, "marine environment concrete durability" matches resource libraries A and B, and "steel corrosion" matches resource libraries A and C. That is, resource libraries A / B / C are all used as query resource libraries. Similarly, reference libraries D and reference library E are matched.

[0018] In this embodiment, the method for obtaining a three-level extended query includes: Fine-grained expansion variants are obtained by performing multi-perspective semantic expansion on each first-level expansion query using a semantic diversity enhancement function. Based on the expansion query optimization objective, these fine-grained expansion variants are selected as second-level expansion queries, expressed as follows: ; ; in For first-level extended queries The multi-perspective semantic extension results for semantic neighborhood The Query a neighbor's address For first-level extended queries Secondary extended queries, For first-level extended queries The Among the fine-grained extended variants, To estimate the relevance score for the retrieval, for and semantic alignment loss, For correlation-diversity balance weights, Alignment penalty coefficient; The first-level and second-level extended queries are hierarchically merged to obtain the third-level extended query; the third-level extended query includes a corresponding extended query weight; In actual evaluation, multi-perspective semantic expansion is performed on each first-level extended query to obtain a semantic diversity metric, and semantic neighborhoods are selected based on the semantic diversity metric. The semantic neighbors within the query are used as fine-grained expansion variants; among a series of fine-grained expansion variants, the one with the largest value is selected as the secondary expansion query by optimizing the objective function of the expansion query; among them, the retrieval relevance prediction score is used. Indicates the use of fine-grained extended variants A standard search was performed in the query resource library to obtain several search results. The average similarity between all search results and the original semantic vector was calculated. The expression for hierarchical fusion is: ; in This is the result of hierarchical fusion, i.e., a three-level expanded query set. This represents the number of queries that can be expanded to the first level. The weight of each level-three extended query is determined by both semantic anchoring score and relevance score, expressed as follows: ; in For third-level extended queries Extended query weights For semantic anchoring-relevance balance weights, For third-level extended queries Search relevance prediction score, For third-level extended queries Anchored score; Multi-perspective semantic expansion was performed on the first-level extended queries ("chloride ion diffusion coefficient", "marine environment concrete durability" and "steel corrosion"). For each first-level extended query, its nearest semantic neighbors were retrieved in the semantic space to form a semantic neighborhood (the number of semantic neighbors was 3 / 4 / 3, corresponding to the number of fine-grained extended variants). The diversity measures were calculated using the semantic diversity enhancement function, and the values ​​were 0.116 / 0.116 / 0.121. Using a relevance-diversity balance weight of 0.8 and an alignment penalty coefficient of 0.4, taking the fine-grained extended variants of the primary extended query "chloride ion diffusion coefficient" ("chloride ion diffusion coefficient determination method", "Fick's law chloride ion diffusion", "concrete chloride ion permeability coefficient") as an example, the estimated search relevance scores are calculated to be 0.78 / 0.72 / 0.85, and the semantic alignment loss is calculated to be 0.52 / 0.68 / 0.35. The corresponding optimization objective values ​​are calculated to be 0.439 / 0.327 / 0.56. 3. Select "concrete chloride ion permeability coefficient" as the secondary extended query. Similarly, determine the remaining secondary extended queries as "concrete corrosion mechanism in splash zone" and "electrochemical detection of steel corrosion". Combine the first-level and second-level extended queries to obtain the tertiary extended queries. The semantic anchoring-relevance balance weight is 0.6. Calculate the extended query weights of the three first-level extended queries as 0.731 / 0.827 / 0.690, and the extended query weights of the three second-level extended queries as 0.772 / 0.810 / 0.724.

[0019] In this embodiment, the method for generating the initial retrieval strategy includes: The third-level extended query is input into the pre-trained language model encoder to obtain a semantic embedding vector. The semantic centroid closest to the semantic embedding vector is retrieved in the embedding space, and the query intent diffusion is calculated, expressed as: ; in For the set of semantic embedding vectors for query intent diffusion, for, for, To expand the number of queries to level three, For third-level extended queries The semantic centroid that is closest in the embedding space. For bandwidth parameters; Perform topic clustering on the document sets of each query resource library to obtain topic clusters and the corresponding document proportions, and calculate resource structure entropy; Calculate the query-resource repository global semantic similarity between the third-level extended query and each query repository. Concatenate the query intent diffusion, resource structure entropy, and query-resource repository global semantic similarity into a feature vector. Output the query-resource coupling degree through a learnable coupling degree network, expressed as: ; in For three-level extended queries and query resource database Query-resource coupling This is the coupling weight vector. To query the global semantic similarity of the resource repository, To query the resource library Resource structure entropy, It is the bias vector; Based on the query-resource coupling degree between the three-level extended query and each query resource library, select the retrieval strategy for the corresponding query resource library, and combine all retrieval strategies to generate the initial retrieval strategy; In actual evaluation, the greater the query intent diffusion, the more diffuse and ambiguous the query intent is; the higher the resource structure entropy, the more dispersed the topic distribution of the resource library is, and the greater the difficulty of retrieval; the query-resource coupling degree characterizes the matching coupling strength between the query and the resource library. The higher the value, the more concentrated the query intent and the more regular the resource library structure is, which is suitable for adopting a high-precision strategy. The lower the value, the more necessary it is to enable a high-recall strategy. Query - Global semantic similarity of the repository For all third-level extended queries Semantic embedding vectors and query resource library The mean cosine similarity of the feature vectors representing the content; Set the bandwidth parameter to 0.1 and the number of queries for the third-level extended range. The query intent diffusion is calculated to be 0.251, the resource structure entropy of each query resource repository is calculated, and the query-resource coupling degree is calculated based on the global semantic similarity between each third-level extended query and each query resource repository. Taking "Resource Repository A" as an example (resource structure entropy is 1.03), the coupling degree weight vector is taken as [-0.8, -0.6, 3.0]. T With a bias of -0.3, the query-resource coupling degree between the third-level extended query and "Resource Library A" is calculated to be 0.808 (corresponding to a global semantic similarity of 0.85 between the query and the resource library). Based on the mapping relationship between the query-resource coupling degree and the candidate strategies, the retrieval strategy for "Resource Library A" is selected as "dense retrieval" (70%) + "sparse retrieval" (30%), and re-ranking is enabled. Similarly, the query-resource coupling degree of the remaining query resource libraries is calculated to select the retrieval strategy. The mapping relationship between query-resource coupling degree and candidate strategies is as follows: when the query-resource coupling degree is [0, 0.4], sparse retrieval (40%) + graph retrieval (60%) is used; when it is (0.4, 0.6], sparse retrieval (60%) + graph retrieval (40%) is used; when it is (0.6, 0.75], dense retrieval (50%) + sparse retrieval (50%) is used; when it is (0.75, 1], dense retrieval (70%) + sparse retrieval (30%) is used; and reordering is enabled when the query-resource coupling degree is greater than 0.5.

[0020] In this embodiment, the method for determining the adaptive screening threshold includes: Execute the initial search strategy to obtain search results and search feedback, and calculate the search result score and search feedback coefficient respectively, as shown in the following expressions: ; in To query the retrieval feedback coefficient of the resource library, For browsing time, Standard browsing time, For click frequency, Standard click frequency, , For feedback weights, This is the hierarchy sensitivity coefficient. To browse levels, For memory sensitivity, Content extraction volume; The product of the search result score and the search feedback coefficient is taken as the search utility. The maximum search utility value in all query resource databases is taken as the utility benchmark. The search utility is normalized to obtain the normalized search utility. The difference between 1 and the normalized search utility is taken as the instantaneous regret value. All instantaneous regret values ​​are accumulated to obtain the cumulative regret value. The adaptive filtering threshold is calculated, and the expression is: ; in For adaptive filtering thresholds, This represents the average retrieval utility across all query databases. For regret adjustment coefficient, This is the cumulative regret value for all queried repositories. This represents the highest cumulative regret value in history. It is a smoothing factor; In practical evaluation, normalized depreciation cumulative gain is used to measure retrieval quality (reflecting the relevance and ranking quality of retrieval results to query intent). Response latency is measured by the time interval from when the retrieval request is sent to when the result is returned, and is normalized by the maximum tolerable latency. Result diversity is measured by the average cosine distance among the top-10 documents in the retrieval results, expressed as: ； ; ; ; in for The search quality of the query resource library, To reduce the accumulated gain, The cumulative gain of loss under ideal sorting. For the first The artificial relevance rating of each search result. Response latency is the time interval between sending a retrieval request and receiving a result. For maximum tolerable delay, For the sake of diversity of results, The cosine distance of the search results For the search results, the first The semantic vector of a document; The search result score is obtained by averaging and weighting search quality, response latency, and result diversity. The above three-level extended query is used to perform "dense search" (70%) + "sparse search" (30%) on "Resource Library A", and reordering is enabled to obtain the search results (maximum tolerable latency). The actual response time was 1200ms, with a target of 2000ms. The search quality was calculated to be 0.88, response latency 0.4, and result diversity 0.72. Search feedback (browsing time 35s, click frequency 8 times, browsing level 4, content extraction volume 12 (based on content block count)) was also calculated. An average weighted fusion yielded a search result score of 0.736. Using a standard browsing time of 60s, a standard click frequency of 10 times, feedback weights of 0.35 / 0.65, a level sensitivity coefficient of 0.3, and a memory sensitivity coefficient of 0.15, the search feedback coefficient for "Resource Library A" was calculated to be 1.491, and the search utility to be 1.098. Similarly, the retrieval utility of all query resource databases is calculated, and the maximum value of 1.098 is taken as the utility benchmark and normalized. The instantaneous regret values ​​of each query resource database are calculated and accumulated to obtain the cumulative regret value of 1.684. The average retrieval utility is calculated to be 0.482. The historical maximum cumulative regret value of 2.5 and the regret adjustment coefficient of 0.5 are taken, and the adaptive filtering threshold is calculated to be 0.32.

[0021] In this embodiment, the method for obtaining feedback content includes: Query resource repositories whose retrieval utility is greater than the adaptive filtering threshold are classified as positive feedback query resource repositories, and those whose retrieval utility is less than the threshold are classified as negative feedback query resource repositories. Calculate the representation similarity between each positive feedback query resource library. Two positive feedback query resource libraries with a representation similarity greater than the second representation similarity threshold are regarded as a centralized resource pair, and otherwise, they are regarded as a decentralized resource pair. The feedback content is formed by taking the intersection of the concentrated resources and the non-intersection of the dispersed resources and the characteristic representation of the content. The feedback content is then mapped to the third-level extended query, and the third-level extended query is updated. In actual evaluation, taking the division of "resource library A / B / C" as an example, the corresponding retrieval utility is 1.098 / 0.329 / 0.018 respectively. "Resource library A / B" is divided into positive feedback query resource library, and "resource library C" is divided into negative feedback query resource library. Similarly, "resource library D" is divided into positive feedback query resource library and "resource library E" is divided into negative feedback query resource library. Calculate the representation similarity between each query resource repository (multi-hot encoding of representation content features to obtain representation content feature vectors, and calculating the cosine similarity between each pair of vectors). Set the second representation similarity threshold to 0.5. Taking "Resource Repository A" and "Resource Repository B" as an example, the representation similarity is calculated to be 0.671, classifying them as a concentrated resource pair. The intersection of the representation content features of the two query resource repositories is: data format / text, resource purpose / research reference, applicable scenario / structural durability. Taking "Resource Repository A" and "Resource Repository D" as an example, the representation similarity is calculated to be 0.38, classifying them as a dispersed resource pair. The non-intersection of the representation content features of the two query resource repositories is: resource category / academic resources - patent technology, resource purpose / research reference - technological innovation. Similarly, similarity calculations are performed on the remaining positive feedback query resource repositories, and feedback content is extracted. The feedback content is encoded into a semantic vector and semantically matched with the six extended queries in the original three-level extended query set (cosine similarity threshold of 0.7). The extended queries and content tags of the original three-level extended queries are: "chloride ion diffusion coefficient" - materials science / research reference, "concrete durability in marine environment" - marine engineering / structural durability, "steel corrosion" - structural durability / technological innovation, "concrete chloride ion permeability coefficient" - materials science / structural durability, "concrete corrosion mechanism in splash zone" - marine engineering / structural durability / research reference, and "electrochemical detection of steel corrosion" - technological innovation / research reference. Calculate the cosine similarity between the feedback content tag vector and the tag vector of the original level-three extended query (using multi-hot encoding). If a match is found, the feedback content and the original level-three extended query are considered to be the same feedback content. Otherwise, the feedback content is used as the updated content. Calculate the cosine similarity between the tag vectors of all the above feedback content and the original level-three extended query. Only the feedback content [steel corrosion, patent technology, technological innovation] failed to match. This content tag is semantically mapped to the standardized search term "steel corrosion technology" as the level-three extended query.

[0022] In this embodiment, the method for iteratively updating the retrieval strategy of the negative feedback query resource library includes: The query intent diffusion is recalculated based on the updated level 3 extended query. The representation similarity between the updated level 3 extended query and the negative feedback query resource library is calculated. The query-resource coupling degree of each feedback query resource library is updated accordingly. The retrieval strategy of the negative feedback query resource library is readjusted based on the query-resource coupling degree. If the level 3 extended query is not updated, directly fine-tune the proportion of the retrieval strategy used in the negative feedback query resource library; The new retrieval strategy is implemented to obtain the retrieval results and retrieval feedback of the negative feedback query resource library. The retrieval utility of the negative feedback query resource library is updated. If the updated retrieval utility is greater than the original adaptive filtering threshold, the corresponding negative feedback query resource library is divided into a positive feedback query resource library and the third-level extended query is updated. Otherwise, the original division results are maintained. If, in the same round of iteration, no negative feedback query resource library is classified as a positive feedback query resource library, then the proportion of the retrieval strategy used in the negative feedback query resource library is continuously fine-tuned until at least one new positive feedback query resource library is added, and then the third-level extended query is updated. Repeat the previous iteration update steps until there is no negative feedback querying the resource library, and output the retrieval strategy adopted by each positive feedback querying the resource library to form an optimized retrieval strategy; In the actual evaluation, the query intent diffusion was recalculated to be 0.22 based on the updated level 3 extended query (significantly lower than before the update, because "steel corrosion patent technology" is highly matched with the resource category and technological innovation application of the patent database). Taking the retrieval strategy update of "Resource Library C" as an example, the average semantic similarity between the updated level 3 extended query and each level 3 extended query and each query resource library was 0.72. The corresponding query-resource coupling degree was calculated to be 0.707. The mapping determined the retrieval strategy to be dense retrieval (50%) + sparse retrieval (50%), and re-sorting was enabled. Similarly, the retrieval strategy of "Resource Library E" was adjusted. A new retrieval strategy is implemented for the negative feedback query resource library, and the corresponding comprehensive retrieval utility is calculated. Taking "Resource Library C" as an example, the retrieval utility calculation result is 0.685 (the filtering threshold of the same batch of query resource libraries remains unchanged, so it is greater than the original adaptive filtering threshold of 0.32). "Resource Library C" is adjusted to a positive feedback query resource library. At this time, only "Resource Library E" is a negative feedback query resource library. In the next iteration, the representation similarity between "Resource Library C" and other positive feedback query resource libraries is calculated to update the feedback content. At this time, the feedback content is mapped to the third-level extended query of the previous iteration. No new third-level extended queries are added. The proportion of the retrieval strategy used by "Resource Library E" (negative feedback query resource library) is directly fine-tuned (adjusted towards the strategy direction corresponding to the high coupling degree interval, that is, gradually increasing the dense retrieval weight and reducing the graph retrieval weight), adjusted to dense 40%+ sparse 60%, and re-ranking is enabled. The new retrieval strategy was applied to "Resource Library E" again, and the corresponding comprehensive retrieval utility was calculated to be 0.295. Therefore, it was still judged as a negative feedback query resource library. Since there was no new positive feedback query resource library in this iteration, the retrieval strategy of "Resource Library E" was fine-tuned again to dense retrieval (50%) + sparse retrieval (50%). After implementing the new retrieval strategy, the corresponding comprehensive retrieval utility is calculated to be 0.305 (less than the original adaptive filtering threshold of 0.32). There are no new positive feedback query resource libraries in this iteration. Continue to fine-tune the retrieval strategy for "Resource Library E" (negative feedback query resource library), using dense retrieval (50%) + sparse retrieval (50%), and enable reordering. Execute the new retrieval strategy for "Resource Library E" again, and the corresponding comprehensive retrieval utility is calculated to be 0.45 (greater than the original adaptive filtering threshold of 0.32). "Resource Library E" is classified as a positive feedback query resource library. At this time, all 5 query resource libraries are classified as positive feedback query resource libraries. Output the corresponding retrieval strategy as the optimized retrieval strategy.

[0023] Secondly, an adaptive retrieval strategy optimization system based on artificial intelligence includes: The query resource library matching module is used to obtain user query statements, process and filter the corpus to obtain first-level extended queries, and calculate the representation similarity between the first-level extended queries and each resource library to match the query resource library. Extended query acquisition module: used to perform multi-perspective semantic expansion on the first-level extended query to obtain the second-level extended query, and perform hierarchical fusion to obtain the third-level extended query; The retrieval strategy generation module is used to calculate the query intent diffusion of the three-level extended query and the resource structure entropy of each query resource library. It calculates the query-resource coupling degree based on the query intent diffusion, resource structure entropy and representation similarity, and generates the initial retrieval strategy based on the query-resource coupling degree. The query resource database partitioning module executes the initial retrieval strategy to obtain retrieval results and retrieval feedback, calculates the retrieval utility and cumulative regret value of each query resource database, determines the adaptive filtering threshold based on the retrieval utility and cumulative regret value, and partitions the query resource database into positive feedback query resource databases and negative feedback query resource databases based on the adaptive filtering threshold. The iterative update module is used to perform characterization analysis on each positive feedback query resource library to obtain feedback content, update the third-level extended query with the feedback content, update the retrieval strategy and retrieval utility of the negative feedback query resource library based on the updated third-level extended query, adjust the negative feedback query resource library to a positive feedback query resource library based on the updated retrieval utility, iteratively update the retrieval strategy of the negative feedback query resource library until there is no more negative feedback query resource library, and output the optimized retrieval strategy.

[0024] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An artificial intelligence-based adaptive search strategy optimization method, characterized by, Includes the following steps: S1. Obtain user query statements, process and filter the corpus to obtain first-level extended queries, and calculate the representation similarity between the first-level extended queries and each resource library to match the query resource library. S2. Perform multi-perspective semantic expansion on the first-level extended query to obtain the second-level extended query, and perform hierarchical fusion to obtain the third-level extended query; S3. Calculate the query intent diffusion of the three-level extended query and the resource structure entropy of each query resource library. Calculate the query-resource coupling degree based on the query intent diffusion, resource structure entropy and representation similarity. Generate an initial retrieval strategy based on the query-resource coupling degree. S4. Execute the initial search strategy to obtain search results and search feedback, calculate the search utility and cumulative regret value of each query resource database, and determine the adaptive filtering threshold based on the search utility and cumulative regret value. S5. Based on the adaptive filtering threshold, the query resource library is divided into positive feedback query resource library and negative feedback query resource library. The positive feedback query resource library is characterized and analyzed to obtain the feedback content. The feedback content is used to update the third-level extended query. S6. Update the retrieval strategy and retrieval utility of the negative feedback query resource library based on the updated level 3 extended query. Adjust the negative feedback query resource library to a positive feedback query resource library based on the updated retrieval utility. Iteratively update the retrieval strategy of the negative feedback query resource library until there is no negative feedback query resource library. Output the optimized retrieval strategy. The search results include search quality, response latency, and result diversity. The search feedback includes browsing time, click frequency, browsing depth, and content extraction volume. 2.The method of claim 1, wherein, The method for matching and querying resource libraries includes: The user query statement is obtained and standardized, and then input into a pre-trained language encoder to obtain the original semantic vector; the user query statement is a colloquial search expression; the standardization process includes word segmentation, stop word filtering, and part-of-speech tagging; The user query is input into a pre-trained language model to generate candidate extended queries, and then input into a pre-trained language encoder to obtain candidate semantic vectors. The pre-trained language model embeds a table linking colloquial search terms with standardized domain search terms. The candidate extended queries are standardized domain search terms or search phrases. The original semantic vector is aligned with the candidate semantic vector using a semantic alignment constraint function, and the anchoring score is obtained by semantic anchoring metric on the candidate expanded query. The expression is as follows: ； ； in For semantic alignment constraint functions, For the user's query statement, For the first One candidate extended query, The original semantic vector, For the first candidate semantic vectors As a balance factor, For Jensen-Shannon divergence, for The probability distribution of the original semantic vector along the semantic dimension. for The probability distribution of candidate expanded queries in the semantic dimension. Expanding the query for candidates Anchored rating, For bandwidth parameters; Candidate extended queries with anchor scores greater than the anchor threshold are selected as first-level extended queries; Calculate the representation similarity between the first-level extended query vector and the representation feature vectors of each resource library, and select the resource libraries whose representation similarity is greater than the first representation similarity threshold. The content features are in multi-hot encoding form, including resource category, data format, resource purpose and applicable scenario.

3. The adaptive retrieval strategy optimization method based on artificial intelligence according to claim 1, characterized in that, The method for obtaining a three-level extended query includes: Fine-grained expansion variants are obtained by performing multi-perspective semantic expansion on each first-level expansion query using a semantic diversity enhancement function. Based on the expansion query optimization objective, these fine-grained expansion variants are selected as second-level expansion queries, expressed as follows: ； ； in For first-level extended queries The multi-perspective semantic extension results for semantic neighborhood The Query a neighbor's address For first-level extended queries Secondary extended queries, For first-level extended queries The Among the fine-grained extended variants, To estimate the relevance score for the retrieval, for and semantic alignment loss, For correlation-diversity balance weights, Alignment penalty coefficient; The first-level and second-level extended queries are hierarchically merged to obtain the third-level extended query; the third-level extended query includes a corresponding extended query weight.

4. The adaptive retrieval strategy optimization method based on artificial intelligence according to claim 1, characterized in that, The method for generating the initial retrieval strategy includes: The third-level extended query is input into the pre-trained language model encoder to obtain a semantic embedding vector. The semantic centroid closest to the semantic embedding vector is retrieved in the embedding space, and the query intent diffusion is calculated, expressed as: ； in For the set of semantic embedding vectors for query intent diffusion, for, for, To expand the number of queries to level three, For third-level extended queries The semantic centroid that is closest in the embedding space. For bandwidth parameters; Perform topic clustering on the document sets of each query resource library to obtain topic clusters and the corresponding document proportions, and calculate resource structure entropy; Calculate the query-resource repository global semantic similarity between the third-level extended query and each query repository. Concatenate the query intent diffusion, resource structure entropy, and query-resource repository global semantic similarity into a feature vector. Output the query-resource coupling degree through a learnable coupling degree network, expressed as: ； in For three-level extended queries and query resource database Query-resource coupling This is the coupling weight vector. To query the global semantic similarity of the resource repository, To query the resource library Resource structure entropy, It is the bias vector; Based on the query-resource coupling degree between the three-level extended query and each query resource library, the retrieval strategy for the corresponding query resource library is selected, and all retrieval strategies are combined to generate the initial retrieval strategy.

5. The adaptive retrieval strategy optimization method based on artificial intelligence according to claim 1, characterized in that, The method for determining the adaptive screening threshold includes: Execute the initial search strategy to obtain search results and search feedback, and calculate the search result score and search feedback coefficient respectively, as shown in the following expressions: ； in To query the retrieval feedback coefficient of the resource library, For browsing time, Standard browsing time, For click frequency, Standard click frequency, , For feedback weights, This is the hierarchy sensitivity coefficient. To browse levels, For memory sensitivity, Content extraction volume; The product of the search result score and the search feedback coefficient is taken as the search utility. The maximum search utility value in all query resource databases is taken as the utility benchmark. The search utility is normalized to obtain the normalized search utility. The difference between 1 and the normalized search utility is taken as the instantaneous regret value. All instantaneous regret values ​​are accumulated to obtain the cumulative regret value. The adaptive filtering threshold is calculated, and the expression is: ； in For adaptive filtering thresholds, This represents the average retrieval utility across all query databases. For regret adjustment coefficient, This is the cumulative regret value for all queried repositories. This represents the highest cumulative regret value in history. This is a smoothing factor.

6. The adaptive retrieval strategy optimization method based on artificial intelligence according to claim 1, characterized in that, The method for obtaining feedback content includes: Query resource repositories whose retrieval utility is greater than the adaptive filtering threshold are classified as positive feedback query resource repositories, and those whose retrieval utility is less than the threshold are classified as negative feedback query resource repositories. Calculate the representation similarity between each positive feedback query resource library. Two positive feedback query resource libraries with a representation similarity greater than the second representation similarity threshold are regarded as a centralized resource pair, and otherwise, they are regarded as a decentralized resource pair. The feedback content is formed by taking the intersection of the centralized resource pairs representing the content features and the non-intersection of the decentralized resource pairs representing the content features. The feedback content is then mapped to the third-level extended query, and the third-level extended query is updated.

7. The adaptive retrieval strategy optimization method based on artificial intelligence according to claim 1, characterized in that, The method for iteratively updating the retrieval strategy of the negative feedback query resource library includes: The query intent diffusion is recalculated based on the updated level 3 extended query. The representation similarity between the updated level 3 extended query and the negative feedback query resource library is calculated. The query-resource coupling degree of each feedback query resource library is updated accordingly. The retrieval strategy of the negative feedback query resource library is readjusted based on the query-resource coupling degree. If the level 3 extended query is not updated, directly fine-tune the proportion of the retrieval strategy used in the negative feedback query resource library; The new retrieval strategy is implemented to obtain the retrieval results and retrieval feedback of the negative feedback query resource library. The retrieval utility of the negative feedback query resource library is updated. If the updated retrieval utility is greater than the original adaptive filtering threshold, the corresponding negative feedback query resource library is divided into a positive feedback query resource library and the third-level extended query is updated. Otherwise, the original division results are maintained. If, in the same round of iteration, no negative feedback query resource library is classified as a positive feedback query resource library, then the proportion of the retrieval strategy used in the negative feedback query resource library is continuously fine-tuned until at least one new positive feedback query resource library is added, and then the third-level extended query is updated. Repeat the previous iteration update steps until there is no negative feedback querying the resource library, then output the retrieval strategy used by each positive feedback query to form an optimized retrieval strategy.

8. An artificial intelligence-based adaptive retrieval strategy optimization system, used to execute the method according to any one of claims 1-7, characterized in that, include: The query resource library matching module is used to obtain user query statements, process and filter the corpus to obtain first-level extended queries, and calculate the representation similarity between the first-level extended queries and each resource library to match the query resource library. Extended query acquisition module: used to perform multi-perspective semantic expansion on the first-level extended query to obtain the second-level extended query, and perform hierarchical fusion to obtain the third-level extended query; The retrieval strategy generation module is used to calculate the query intent diffusion of the three-level extended query and the resource structure entropy of each query resource library. It calculates the query-resource coupling degree based on the query intent diffusion, resource structure entropy and representation similarity, and generates the initial retrieval strategy based on the query-resource coupling degree. The query resource database partitioning module executes the initial retrieval strategy to obtain retrieval results and retrieval feedback, calculates the retrieval utility and cumulative regret value of each query resource database, determines the adaptive filtering threshold based on the retrieval utility and cumulative regret value, and partitions the query resource database into positive feedback query resource databases and negative feedback query resource databases based on the adaptive filtering threshold. The iterative update module is used to perform characterization analysis on each positive feedback query resource library to obtain feedback content, update the third-level extended query with the feedback content, update the retrieval strategy and retrieval utility of the negative feedback query resource library based on the updated third-level extended query, adjust the negative feedback query resource library to a positive feedback query resource library based on the updated retrieval utility, iteratively update the retrieval strategy of the negative feedback query resource library until there is no more negative feedback query resource library, and output the optimized retrieval strategy.

Images