Retrieval rearrangement method, retrieval method and related device

Abstract

The invention discloses a retrieval rearrangement method, a retrieval method and a related device, and relates to the technical field of information retrieval, and the retrieval rearrangement method comprises the following steps: obtaining a preliminary retrieval result of a query question; determining the question type of the query question in the business dimension to obtain a target question type; obtaining a rearrangement scoring strategy matched with the target question type, and obtaining a target rearrangement scoring strategy; according to a target rearrangement scoring strategy, scoring a plurality of pieces of data contained in the preliminary retrieval result to obtain target scores corresponding to the plurality of pieces of data; and sorting the plurality of pieces of data according to the target scores corresponding to the plurality of pieces of data. Through the retrieval rearrangement method disclosed by the invention, the retrieval rearrangement result which better meets the requirements of the user can be obtained.

Description

Technical Field

[0001] This application relates to the field of information retrieval technology, and in particular to a retrieval rearrangement method, a retrieval method, and related apparatus. Background Technology

[0002] Retrieval reordering is a key technology in the field of information retrieval, widely used in search engines, recommendation systems, question-answering systems, and intelligent customer service. The core objective of retrieval reordering is to retrieve data relevant to the query from massive amounts of data, obtain preliminary search results, and then optimize and rank these preliminary results to improve accuracy and user experience.

[0003] The current retrieval reordering scheme involves first obtaining preliminary retrieval results for the query (i.e., retrieving data related to the query from the database), and then using a specific reordering strategy (such as a keyword matching-based reordering strategy, a keyword weight-based reordering strategy, or a large model-based reordering strategy) to reorder the preliminary retrieval results for the query.

[0004] While current search reordering schemes can reorder the initial search results, the reordering effect is poor and cannot meet user needs. Summary of the Invention

[0005] In view of this, this application provides a retrieval rearrangement method, a retrieval method, and related apparatus to solve the problem that existing retrieval rearrangement schemes are ineffective and cannot meet user needs. The technical solution is as follows:

[0006] A retrieval rearrangement method, comprising:

[0007] Obtain preliminary search results for the query question;

[0008] Determine the question type of the query in the business dimension to obtain the target question type;

[0009] Obtain a reordering scoring strategy that is compatible with the target problem type to obtain the target reordering scoring strategy;

[0010] According to the target reordering scoring strategy, the data contained in the preliminary search results are scored respectively to obtain the target scores corresponding to the data.

[0011] The data are sorted according to the target scores corresponding to the data.

[0012] In one possible implementation, the preliminary search results include several data entries, each corresponding to a relevance score;

[0013] The step of scoring several data entries contained in the preliminary search results according to the target reordering scoring strategy includes:

[0014] Obtain the relevance score range of the preliminary search results;

[0015] By categorizing the relevance scoring range, several scoring intervals are obtained for the preliminary search results;

[0016] For each data point whose relevance score falls within the same scoring range, a score is assigned to that data point according to the target re-ranking scoring strategy.

[0017] In one possible implementation, the target reordering scoring strategy includes: several scoring dimensions in a general domain, and weights corresponding to the several scoring dimensions in the general domain that are adapted to the target problem type.

[0018] In one possible implementation, obtaining the reordering scoring strategy adapted to the target problem type includes:

[0019] From the pre-configured re-ranking scoring strategies for several problem types, obtain the re-ranking scoring strategy pre-configured for the target problem type, and obtain the re-ranking scoring strategy adapted to the target problem type, wherein the several problem types are problem types in the business dimension.

[0020] In one possible implementation, obtaining the reordering scoring strategy adapted to the target problem type includes:

[0021] Obtain a pre-configured basic re-ranking scoring strategy, which includes several scoring dimensions in a general domain and basic weights corresponding to the several scoring dimensions in the general domain.

[0022] Determine the scoring dimensions associated with the target problem type;

[0023] Increase the weight of the scoring dimensions related to the target problem type in the basic reordering scoring strategy;

[0024] The weighted reordering scoring strategy is determined as the reordering scoring strategy that is suitable for the target problem type.

[0025] In one possible implementation, the target reordering scoring strategy includes several scoring dimensions and weights corresponding to the several scoring dimensions.

[0026] The step of scoring the data according to the target reordering scoring strategy includes:

[0027] The data is scored according to the scoring dimensions to obtain the scores corresponding to the data in each of the scoring dimensions.

[0028] Based on the weights corresponding to the aforementioned scoring dimensions, the scores of the data corresponding to the aforementioned scoring dimensions are weighted and summed to obtain the weighted sum score of the data.

[0029] The weighted sum of the data points is combined with the tier score corresponding to the scoring range in which the data points' relevance score falls to obtain the target score for the data points. The tier score is determined based on the upper limit of the relevance scoring range and the weights corresponding to the various scoring dimensions.

[0030] In one possible implementation, scoring the data item across the several scoring dimensions to obtain the score corresponding to the data item across each of the several scoring dimensions includes:

[0031] Obtain a pre-configured re-ranking scoring quantization template, which includes several scoring dimensions in a general domain, as well as scoring elements and scoring methods for each scoring dimension;

[0032] For each scoring dimension in the target re-ranking scoring strategy, the scoring elements and scoring method for that scoring dimension are determined according to the re-ranking scoring quantification template.

[0033] Based on the scoring elements and scoring method of this scoring dimension, the data is scored to obtain the score corresponding to this data in this scoring dimension.

[0034] In one possible implementation, the scoring dimensions of the general domain include some or all of the following scoring dimensions:

[0035] The scoring dimensions include relevance scoring, abstract scoring, title similarity scoring, timeliness scoring, source site scoring, author scoring, content quality scoring, website authority scoring, and keyword overlap scoring.

[0036] In one possible implementation, obtaining several scoring intervals for the preliminary search results by categorizing the relevance scoring range includes:

[0037] By classifying the relevance scoring range according to the first classification method or the second classification method, several scoring intervals of the preliminary search results are obtained.

[0038] The first grading method involves dividing the scoring range to be graded according to a preset grading width to obtain several scoring intervals with the same span. The second grading method involves dividing the scoring range to be graded with the goal of balancing the amount of data corresponding to different scoring intervals to obtain several scoring intervals. The amount of data corresponding to a scoring interval is the number of data whose relevance scores are located in that scoring interval.

[0039] In one possible implementation, the relevance scores corresponding to the several data points included in the preliminary search results are normalized relevance scores, and the relevance scoring range is the normalized relevance scoring range.

[0040] The step of dividing the relevance scoring range according to a first or second grading method to obtain several scoring intervals for the preliminary search results includes:

[0041] By directly classifying the normalized relevance scoring range according to the first classification method or the second classification method, several scoring intervals of the preliminary search results are obtained.

[0042] Alternatively, the normalized relevance scores corresponding to the several data points included in the preliminary search results can be restored to the unnormalized relevance scores. Based on the unnormalized relevance scores corresponding to the several data points included in the preliminary search results, the unnormalized scoring range corresponding to the normalized scoring range can be determined. The unnormalized scoring range can be divided into several scoring intervals of the preliminary search results by using either the first or second grading method.

[0043] A second aspect of this application provides a retrieval method, comprising:

[0044] Get the query question;

[0045] By employing any of the above-mentioned retrieval rearrangement methods, the retrieval rearrangement results corresponding to the query question can be obtained.

[0046] Based on the retrieval rearrangement results corresponding to the query question, the final retrieval result corresponding to the query question is determined.

[0047] A third aspect of this application provides an electronic device, comprising at least one processor and a memory connected to the processor, wherein:

[0048] The memory is used to store computer programs;

[0049] The processor is used to execute the computer program so that the electronic device can implement the steps of any of the above-described retrieval rearrangement methods, or implement the steps of the above-described retrieval method.

[0050] A fourth aspect of this application provides a computer storage medium carrying one or more computer programs that, when executed by an electronic device, enable the electronic device to implement the steps of any of the above-described retrieval rearrangement methods, or to implement the steps of the above-described retrieval method.

[0051] The fifth aspect of this application provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to perform the steps of any of the above-described retrieval rearrangement methods, or to perform the steps of the above-described retrieval method.

[0052] By employing the above technical solution, the retrieval reordering method provided in this application, after obtaining preliminary retrieval results, first determines the question type of the query in the business dimension to obtain the target question type, then obtains a reordering scoring strategy adapted to the target question type, thus obtaining the target reordering scoring strategy. Next, based on the target reordering scoring strategy, scores are applied to several data entries included in the preliminary retrieval results to obtain the target scores corresponding to each data entry. Finally, based on the target scores corresponding to each data entry, the data entries in the preliminary retrieval results are sorted. Considering that user intent and needs may differ in different business scenarios, the retrieval reordering method provided in this application analyzes the question type of the query in the business dimension, and then adopts a reordering scoring strategy adapted to that question type to score the data entries in the preliminary retrieval results, thereby reordering them. The retrieval reordering method provided in this application is applicable to general domains and can provide results that better meet user needs, thus improving user experience. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0054] Figure 1 A schematic flowchart illustrating the retrieval and rearrangement method provided in this application embodiment;

[0055] Figure 2 This is a flowchart illustrating how a target score is obtained by scoring several data items in the preliminary search results according to a target reordering scoring strategy provided in this application embodiment.

[0056] Figure 3This is a schematic diagram illustrating the process of restoring the normalized correlation score to the unnormalized correlation score, as provided in an embodiment of this application.

[0057] Figure 4 A schematic diagram illustrating the rearrangement scoring method provided in an embodiment of this application;

[0058] Figure 5 This is a schematic diagram of the retrieval and rearrangement device provided in an embodiment of this application. Detailed Implementation

[0059] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0060] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0061] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0062] Current search ranking schemes include keyword matching-based, keyword weight-based, large-model-based, and link analysis-based schemes, among others. The inventors of this case, through research on these schemes, found that none of them consider the business scenario. Furthermore, they all simply shuffle the initial relevance search results and then rank them based on a new logic. These schemes produce poor ranking results and fail to meet user needs.

[0063] In view of this, this application provides a search reordering method with better performance. The search reordering method provided by this application will be described in the following embodiments.

[0064] Please see Figure 1 The diagram illustrates a flowchart of a retrieval rearrangement method provided in an embodiment of this application. This method may include:

[0065] Step S101: Obtain preliminary search results for the query question.

[0066] The preliminary search results for the query question include several data points related to the query question. Each data point in the preliminary search results has a relevance score, which is the relevance score between the corresponding data and the query question.

[0067] In this embodiment, the process of obtaining preliminary search results for a query question includes: determining the relevance score between the query question and the data in the search database; sorting the data in the search database in descending order of relevance score; taking the data whose relevance score to the query question is greater than a preset score threshold as the data related to the query question, i.e., the preliminary search results for the query question; or taking the top N data (i.e., top N data) as the data related to the query question, i.e., the preliminary search results for the query question. The value of N can be set according to the actual situation.

[0068] There are several ways to determine the relevance score between the query question and the data in the retrieved database. For example, the query question and each piece of data in the retrieved database can be input into a relevance scoring model (such as a deep learning model) to score them, so as to obtain the relevance score between the query question and each piece of data in the retrieved database.

[0069] Optionally, considering the large amount of data in the retrieval database, it is extremely time-consuming to score the relevance between the query and each piece of data in the retrieval database based on the relevance scoring model. Before scoring based on the relevance scoring model, the data in the retrieval database can be coarsely sorted and scored to filter out most of the irrelevant data, and the remaining data can be scored based on the relevance scoring model.

[0070] This embodiment is not limited to using the scoring method described above. For example, the data in the retrieval database can be vectorized in advance to obtain the vector representation of the data in the retrieval database. After obtaining the query question, the query question is vectorized to obtain the vector representation of the query question. The similarity between the vector representation of the query question and the vector representation of the data in the retrieval database is calculated (for example, by calculating the inner product or by obtaining the similarity through an MLP layer), and this is used as the relevance score. In this way, the relevance score between the query question and the data in the retrieval database can be obtained.

[0071] It should be noted that the above-described method for determining the relevance score between the query question and the data in the retrieved database is merely an example, and other methods for determining the relevance score between the query question and the data in the retrieved database are also applicable to this application.

[0072] Optionally, considering that model-based relevance scoring strategies may have error issues, a simple text relevance scoring (such as calculating the similarity between the query question and the data titles in the retrieved database) can be introduced on the basis of the model-based relevance scoring strategy. The relevance scores of the two aspects can be merged, and the merged relevance score can be used as the final relevance score.

[0073] Step S102: Determine the problem type of the query in the business dimension to obtain the target problem type.

[0074] Multiple question types (such as knowledge, information, and overview) can be set at the business dimension. The question type of the query question is determined from the set multiple question types, thus obtaining the question type of the query question at the business dimension, i.e., the target question type.

[0075] In one possible implementation, a problem classification model can be trained using query problem samples labeled with problem types (such as knowledge-based query problem samples, information-based query problem samples, and summary-based query problem samples labeled with problem types). Then, the query problem to be classified is input into the problem classification model for classification to obtain the problem type of the query problem in the business dimension.

[0076] In another possible implementation, the query question and multiple question types defined in the business dimension can be input into a large model, allowing the large model to analyze the query question and determine the question type from the multiple question types.

[0077] Step S103: Obtain the reordering scoring strategy that matches the target problem type to obtain the target reordering scoring strategy.

[0078] The target reordering scoring strategy includes several scoring dimensions and the corresponding weights for each scoring dimension.

[0079] In one possible implementation, the target reordering scoring strategy may include several scoring dimensions in a general domain, as well as the weights corresponding to each of the several scoring dimensions in the general domain. In this implementation, the weights of the target reordering scoring strategy are weights that are adapted to the target problem type.

[0080] In another possible implementation, the target reordering scoring strategy may include scoring dimensions related to the target problem type from several scoring dimensions in the general domain, as well as the weights corresponding to the scoring dimensions related to the target problem type.

[0081] Step S104: Based on the target reordering scoring strategy, score the data in the preliminary search results respectively to obtain the target scores corresponding to the data in the preliminary search results.

[0082] After determining the reordering scoring strategy that is suitable for the target question type, i.e. the target reordering scoring strategy, each piece of data in the preliminary search results is scored according to the target reordering scoring strategy to obtain the target scores corresponding to the several pieces of data in the preliminary search results.

[0083] Step S105: Sort the data in the preliminary search results according to the target scores corresponding to the data in the preliminary search results.

[0084] After obtaining the target scores corresponding to the several data points contained in the preliminary search results, the several data points contained in the preliminary search results can be sorted in descending order of target scores to obtain the search reordering results corresponding to the query question.

[0085] The retrieval reordering method provided in this application, after obtaining preliminary retrieval results, first determines the question type of the query in the business dimension to obtain the target question type. Then, it obtains a reordering scoring strategy adapted to the target question type, thus obtaining the target reordering scoring strategy. Next, according to the target reordering scoring strategy, it scores several data items included in the preliminary retrieval results to obtain the target scores corresponding to each of the several data items included in the preliminary retrieval results. Finally, it sorts the several data items included in the preliminary retrieval results according to the target scores corresponding to each of the several data items included in the preliminary retrieval results. Considering that users' intentions and needs may differ in different business scenarios, the retrieval reordering method provided in this application analyzes the question type of the query in the business dimension, and then adopts a reordering scoring strategy adapted to the question type to score the several data items included in the preliminary retrieval results, and then reorders them. The retrieval reordering method provided in this application is applicable to general fields and can provide results that are more in line with user needs, thereby improving the user experience.

[0086] In some embodiments of this application, the process of "step S103: obtaining a reordering scoring strategy that is adapted to the target problem type and obtaining the target reordering scoring strategy" is described.

[0087] Obtain a reordering scoring strategy that is compatible with the target problem type. There are multiple ways to implement the target reordering scoring strategy. This embodiment provides the following two optional implementation methods.

[0088] Obtain the reordering scoring strategy that matches the target problem type, resulting in the first implementation of the target reordering scoring strategy:

[0089] From the pre-configured re-ranking scoring strategies for several problem types in the business dimension, obtain the re-ranking scoring strategy pre-configured for the target problem type, and obtain the re-ranking scoring strategy adapted to the target problem type.

[0090] Several question types can be defined at the business dimension, and a re-ranking scoring strategy can be pre-configured for each question type. After determining the question type of the query question at the business dimension, i.e., the target question type, the re-ranking scoring strategy configured for the target question type can be obtained from the re-ranking scoring strategies pre-configured for each question type. The re-ranking scoring strategy configured for the target question type is the re-ranking scoring strategy adapted to the target question type.

[0091] For example, the question types set in the business dimension include knowledge type, information type, and summary type. Re-ranking scoring strategies are pre-configured for knowledge type, information type, and summary type respectively. Assuming that the question type of the query question is determined to be information type, the re-ranking scoring strategy configured for information type is obtained. The re-ranking scoring strategy configured for information type is the re-ranking scoring strategy adapted to the query question.

[0092] In one possible implementation, the pre-configured reordering scoring strategy for any problem type may include: several scoring dimensions in a general domain, and weights corresponding to the various scoring dimensions in the general domain that are adapted to the problem type.

[0093] Each piece of data retrieved from the database may include, but is not limited to, some or all of the following information: abstract, content, title, author, source site, publication date, website information, etc.

[0094] The scoring dimensions in the general domain may include, but are not limited to, some or all of the following scoring dimensions: relevance scoring dimension, abstract scoring dimension, title similarity scoring dimension, keyword overlap scoring dimension, timeliness scoring dimension, source site scoring dimension, website authority scoring dimension, author scoring dimension, content quality scoring dimension, etc.

[0095] For example, several scoring dimensions in a general domain include relevance scoring dimensions, abstract scoring dimensions, title similarity scoring dimensions, timeliness scoring dimensions, source site scoring dimensions, author scoring dimensions, content quality scoring dimensions, and website authority scoring dimensions. The pre-configured re-ranking scoring strategy for any question type can include these scoring dimensions and the weights corresponding to these scoring dimensions, wherein the weights corresponding to the scoring dimensions related to the question type are relatively high.

[0096] In another possible implementation, the pre-configured re-ranking scoring strategy for any question type may include: a number of scoring dimensions in a general domain that are relevant to that question type, and the weights corresponding to those scoring dimensions. In this implementation, the more relevant a scoring dimension is to the question type, the higher its weight can be.

[0097] For example, several scoring dimensions in the general domain include summary scoring, title similarity scoring, timeliness scoring, source site scoring, author scoring, content quality scoring, and website authority scoring. The target question type is news, and the scoring dimensions related to news can include timeliness scoring (the core value of news lies in its freshness, and the newer the content, the more important it is usually) and website authority scoring (news published by authoritative websites is more credible). Therefore, the pre-configured re-ranking scoring strategy for news can include timeliness scoring and website authority scoring, as well as the weights corresponding to timeliness scoring and website authority scoring.

[0098] This embodiment also provides a second implementation method for obtaining a reordering scoring strategy that is adapted to the target problem type, thus obtaining the target reordering scoring strategy:

[0099] Step a1: Obtain the pre-configured basic reordering scoring strategy.

[0100] The basic reordering scoring strategy includes several scoring dimensions in the general domain, as well as the basic weights corresponding to each of these scoring dimensions.

[0101] Step a2: Determine the scoring dimensions related to the target question type.

[0102] In one possible implementation, the scoring dimensions related to the target question type can be determined as follows: a correspondence between question types and related scoring dimensions is pre-defined. After obtaining the target question type, the relevant scoring dimensions corresponding to the target question type can be determined based on the pre-defined correspondence between question types and related scoring dimensions.

[0103] For example, several question types in the business dimension include information, knowledge, and review. The relevant scoring dimensions for information queries are pre-defined, including timeliness and website authority. For knowledge queries, the relevant scoring dimensions are defined as content quality and website authority. For review queries, the relevant scoring dimensions are defined as content quality and authorship. Assuming the target question type is a review, then based on the pre-defined correspondence between question types and relevant scoring dimensions, the relevant scoring dimensions for the target question type can be determined to be content quality and authorship. In other words, the scoring dimensions related to the target question type are content quality and authorship.

[0104] In addition to determining the scoring dimensions related to the target question type based on the pre-defined correspondence between question types and related scoring dimensions, this embodiment also provides another optional implementation method, namely, using a large model to determine the scoring dimensions related to the target question type. Specifically, the query question, the target question type, and the scoring dimensions of the general domain can be input into the large model, allowing the large model to filter out the scoring dimensions related to the target question type from the scoring dimensions of the general domain based on the query question and the target question type.

[0105] Step a3: Increase the weight of the scoring dimensions related to the target problem type in the basic reordering scoring strategy.

[0106] After identifying the scoring dimensions relevant to the target question type, the weights of the scoring dimensions relevant to the target question type in the basic reordering scoring strategy can be increased.

[0107] For example, if the target question type is information, and the scoring dimensions related to information are timeliness and website authority, then the weights of the timeliness and website authority dimensions in the basic re-ranking scoring strategy will be increased respectively.

[0108] In one possible implementation, increasing the weight of the scoring dimension related to the target question type can be achieved by adding a set value.

[0109] Understandably, there may be multiple scoring dimensions related to the target question type. In one possible implementation, the same value can be added to different scoring dimensions. For example, if the target question type is news, and the scoring dimensions related to news are timeliness and website authority, then the weights corresponding to timeliness and website authority can each be increased by 0.5. In another possible implementation, different values ​​can be added to the weights corresponding to different scoring dimensions based on their relevance to the target question type. A relatively larger value can be added to the weights corresponding to scoring dimensions that are more relevant to the target question type. For example, if the target question type is news, and the scoring dimensions related to news are timeliness and website authority, considering that timeliness is more relevant to news, the weight corresponding to timeliness can be increased by 0.5, and the weight corresponding to website authority can be increased by 0.3.

[0110] It should be noted that this embodiment does not limit the way of increasing the weight to adding a set value. For example, it can also be multiplied by a preset multiple. Other ways to increase the weight are also applicable to this application.

[0111] Step a4: Determine the weighted re-ranking scoring strategy as a re-ranking scoring strategy that is suitable for the target problem type.

[0112] By increasing the weights of the scoring dimensions related to the target problem type in the basic reordering scoring strategy, a reordering scoring strategy adapted to the target problem type is obtained, namely the target reordering scoring strategy.

[0113] In some embodiments of this application, the implementation process of "step S104: according to the target reordering scoring strategy, score the several data items contained in the preliminary search results respectively to obtain the target scores corresponding to the several data items contained in the preliminary search results respectively" is described.

[0114] In one possible implementation, such as Figure 2 As shown, the process of scoring several data points in the preliminary search results according to the target reordering scoring strategy to obtain the target score for each of the several data points in the preliminary search results may include:

[0115] Step S201: Obtain the relevance score range of the preliminary search results.

[0116] For example, if the relevance score of each data point in the preliminary search results is within the range of [0, 1], then the relevance score range is [0, 1].

[0117] Step S202: By classifying the relevance scoring range, several scoring intervals of the preliminary search results are obtained.

[0118] The principle for categorizing the relevance scoring range is to ensure that the relevance scores of data in the same category are very similar, and that the relevance ranking results are acceptable regardless of the position of any two data in the same category.

[0119] There are several ways to divide the relevance scoring range into several scoring zones to obtain preliminary search results. This embodiment provides the following four optional implementation methods.

[0120] First implementation method:

[0121] By directly categorizing the relevance scoring range according to the first categorization method, several scoring intervals with the same span are obtained for the preliminary search results. The first categorization method involves categorizing the scoring range to be categorized according to a preset categorization width.

[0122] The span between each scoring zone obtained by directly dividing the relevance scoring range according to the first grading method is the same, and the span between each scoring zone is the preset grading width.

[0123] For example, if the relevance scoring range is [0, 1] and the preset grading width is 0.05, then by directly grading the relevance scoring range according to the first grading method, the following scoring intervals can be obtained: [0, 0.05), [0.05, 0.1), ..., [0.85, 0.9), [0.9, 0.95), [0.95, 1), 1.

[0124] It should be noted that the preliminary search results contain data related to the query question. These data typically have relatively high relevance scores to the query question. This means that some intervals obtained by categorizing the relevance score ranges may be invalid. For example, if the relevance score ranges are categorized as [0, 0.05), [0.05, 0.1), ..., [0.85, 0.9), [0.9, 0.95), [0.95, 1), and 1, the minimum relevance score among the data points in the preliminary search results is 0.85, and the maximum score is 0.99. The relevance scores of the data points in the preliminary search results fall within the three scoring intervals [0.85, 0.9), [0.9, 0.95), and [0.95, 1). Therefore, these three scoring intervals are valid scoring intervals, and the others are invalid scoring intervals. The scoring intervals in the preliminary search results refer to the valid scoring intervals.

[0125] Of course, the relevance scoring range can also be determined based on the minimum and maximum relevance scores of the data points contained in the preliminary search results. For example, if the minimum relevance score of the data points contained in the preliminary search results is 0.85 and the maximum score is 0.99, then [0.85, 1) can be determined as the relevance scoring range.

[0126] The second implementation method:

[0127] By directly categorizing the relevance scoring range according to the second categorization method, several scoring intervals are obtained for the preliminary search results. The second categorization method aims to balance the amount of data corresponding to different scoring intervals by categorizing the scoring range to be categorized, resulting in several scoring intervals. The amount of data corresponding to each scoring interval is the number of data points whose relevance scores fall within that interval.

[0128] It should be noted that the relevance scores of the several data points in the initial search results are usually normalized scores. For example, when performing relevance scoring based on a deep learning model, the last layer of the model typically performs a round of normalization. Common normalization methods include sigmoid-based normalization and softmax-based normalization. In these normalized scores, the relevance scores corresponding to data points highly relevant to the query question are usually clustered together; for example, relevance scores close to 1 tend to cluster together. Figure 3 As shown, correlation scores close to 1 include 0.99330715, 0.99752738, 0.99908895, and 0.99998330. These correlation scores are all concentrated around 1, and direct grading cannot separate them. In view of this, the embodiments of this application provide the following two grading methods.

[0129] The third implementation method:

[0130] The relevance scores (normalized relevance scores) of several data points in the preliminary search results are restored to their unnormalized relevance scores, thus obtaining the unnormalized relevance scores of the several data points in the preliminary search results. Based on the unnormalized relevance scores of the several data points in the preliminary search results, the unnormalized scoring range corresponding to the relevance scoring range is determined. The unnormalized scoring range is then divided into several scoring intervals of the preliminary search results by classifying the unnormalized scoring range according to the first classification method.

[0131] When restoring the normalized relevance score corresponding to any data point in the preliminary search results, the unnormalized relevance score can be obtained by reverse calculation based on the normalization calculation method.

[0132] like Figure 3 As shown, restoring the correlation score of 0.99752738 yields 6, restoring the correlation score of 0.99908895 yields 7, restoring the correlation score of 0.99966465 yields 8, restoring the correlation score of 0.99987661 yields 9, and restoring the correlation score of 0.99995460 yields 10. Figure 3 It can be seen that after restoring the correlation scores clustered around 1, the correlation scores were effectively split.

[0133] The fourth implementation method:

[0134] The relevance scores (normalized relevance scores) of several data points in the preliminary search results are restored to their unnormalized relevance scores, thus obtaining the unnormalized relevance scores of several data points in the preliminary search results. Based on the unnormalized relevance scores of several data points in the preliminary search results, the unnormalized scoring range corresponding to the relevance scoring range is determined. The unnormalized scoring range is then divided into several scoring intervals for the preliminary search results by using the second grading method.

[0135] Step S203: For each data point whose relevance score (normalized relevance score or unnormalized relevance score) is located in the same scoring zone, score the data point according to the target re-ranking scoring strategy.

[0136] In order to achieve personalized ranking while ensuring relevance, this application introduces a tiered strategy and scores within each tier according to the target re-ranking scoring strategy.

[0137] The above embodiment mentions that the target reordering scoring strategy includes several scoring dimensions and corresponding weights for each scoring dimension. For any data D whose relevance score falls within the same scoring range... i (The i-th data in the preliminary search results), according to the target re-ranking scoring strategy, for this data D i The scoring process may include:

[0138] Step b1: Analyze the data D across several scoring dimensions. i Scoring was performed to obtain data D. i The scores corresponding to several scoring dimensions.

[0139] In one possible implementation, the data D is evaluated across several scoring dimensions. i Scoring was performed to obtain data D. i The process of scoring corresponding to several scoring dimensions can include:

[0140] Step b11: Obtain the pre-configured rearrangement scoring quantization template.

[0141] The reordered scoring quantification template includes several scoring dimensions in a general domain, as well as the scoring elements and scoring methods corresponding to each of these dimensions. Table 1 below shows an example of the reordered scoring quantification template.

[0142] Table 1. Examples of pre-configured rearrangement scoring quantization templates

[0143]

[0144] It should be noted that the scoring dimensions shown above are only examples. In actual applications, more scoring dimensions may be included, such as website preference scoring dimensions. In actual use, the reordering scoring quantification template can be supplemented, adjusted and optimized to meet the actual application needs.

[0145] Step b12: For each scoring dimension in the target re-ranking scoring strategy, determine the scoring elements and scoring method for that scoring dimension based on the re-ranking scoring quantification template, and then apply the scoring elements and scoring method to the data D. i Scoring was performed to obtain data D. i The score corresponding to this scoring dimension.

[0146] For example, one scoring dimension in the target reordering scoring strategy is the relevance scoring dimension, for data D. i When scoring the relevance scoring dimension, the scoring elements and scoring method of the relevance scoring dimension are determined according to the pre-configured re-ranking scoring quantization template. Based on the re-ranking scoring quantization template shown in Table 1 above, the scoring elements of the relevance scoring dimension can be determined as the preliminary relevance score. The scoring method is to take the smaller value between the two boundary values ​​of the scoring interval where the relevance score lies. Thus, data D is obtained. i The relevance score is used to obtain data D. i The scoring interval where the relevance score lies, take data D i The smaller of the two boundary values ​​of the scoring interval in which the relevance score lies is taken as data D. i The score corresponding to the relevance scoring dimension.

[0147] For example, one scoring dimension in the target reordering scoring strategy is the summary scoring dimension, for data D i When scoring the abstract, the scoring elements and scoring method for the abstract scoring dimension are determined according to the pre-configured rearrangement scoring quantization template. Based on the rearrangement scoring quantization template shown in Table 1 above, the scoring element for the abstract scoring dimension is determined to be abstract length, and the scoring method is tanh(abstract length / 10000). Therefore, data D is obtained. i Given the abstract length L, calculate tanh(L / 10000) to obtain the data D. i The score corresponding to the summary scoring dimension.

[0148] For example, one scoring dimension in the target reordering scoring strategy is the title similarity scoring dimension, for data D iWhen scoring the title similarity dimension, the scoring elements and scoring methods for the title similarity dimension are determined based on the pre-configured re-ranking scoring quantification template. According to the re-ranking scoring quantification template shown in Table 1 above, the scoring elements for the title similarity dimension are the query question and the title of the data to be scored. The scoring method can be any of the following: calculating Jaccard similarity, calculating edit distance, calculating string similarity (lcs), or calculating the hit ratio of keywords in the query question in the title. Then, the query question and data D are obtained. i The title uses one of the scoring methods mentioned above to evaluate the query question and data D. i Scoring is done on the title, for example, based on the query question and data D. i Calculate the Jaccard similarity of the titles to obtain data D. i The score corresponding to the title similarity scoring dimension.

[0149] For example, one scoring dimension in the target reordering scoring strategy is the timeliness scoring dimension, for data D. i When scoring the timeliness dimension, the scoring elements and scoring method of the timeliness scoring dimension are determined according to the pre-configured re-ranking scoring quantification template. Based on the re-ranking scoring quantification template shown in Table 1 above, the scoring element for the timeliness scoring dimension is determined to be the publication time of the data to be scored. The scoring method is to calculate tanh based on the ranking position of the publication time of the data to be scored among the publication times of all data to be scored within the file. Therefore, the data D is determined. i The publication time of data D is its sorted position among the publication times of all data to be scored within the file. i If the release time is ranked second, then calculate tanh(2) to obtain the data D. i The score corresponding to the timeliness scoring dimension.

[0150] For example, one scoring dimension in the target re-ranking scoring strategy is the source site scoring dimension, for data D i When scoring the source site scoring dimension, the scoring elements and scoring method of the source site scoring dimension are determined according to the pre-configured re-ranking scoring quantification template. Based on the re-ranking scoring quantification template shown in Table 1 above, the scoring elements of the source site scoring dimension are determined as the source sites of the data to be scored. The scoring method involves determining the importance level of the source sites and then determining the score based on the importance level. This process is then used to obtain data D. i The source site, determine data D i The importance level of the source site, according to data D i The importance level of the source site is determined by data D. iThe scores correspond to the scoring dimensions of the source site. It should be noted that important sites within the domain can be pre-selected, and each selected site can be ranked according to its importance, with a corresponding score assigned to each rank. This is then applied to data D. i When scoring based on the source site scoring dimension, data D can be determined. i The importance level of the source site, and thus the data D i The score corresponding to the importance level of the source site is determined as data D. i The score corresponding to the scoring dimension of the source site.

[0151] For example, one scoring dimension in the target re-ranking scoring strategy is the author scoring dimension. Based on a pre-configured re-ranking scoring quantization template, the scoring elements and scoring method for the author scoring dimension are determined. According to the re-ranking scoring quantization template shown in Table 1 above, the scoring elements for the author scoring dimension are determined to be the authors of the data to be scored, and the scoring method is to determine the author's rank and then determine the score based on the author's rank. Therefore, the data D is then... i When scoring based on the author rating dimension, data D is obtained. i The author, identified data D i The author's ranking, based on data D i The author's ranking data D i Scores are awarded based on the author rating dimension. It should be noted that information from several authors can be collected, including factors such as follower count and article publication volume. Based on this information, authors are categorized and assigned a corresponding score for each category. This data is then used to score the authors. i When scoring based on the author's rating dimension, determine the data D. i The author's level will be the data D i The score corresponding to the author's level is determined as data D. i The score corresponding to the author's rating dimension.

[0152] For example, one scoring dimension in the target re-ranking scoring strategy is the content quality scoring dimension, which is applied to data D. i When scoring the author rating dimension, the scoring elements and scoring methods for the content quality scoring dimension are determined based on the pre-configured re-ranking scoring quantification template. According to the re-ranking scoring quantification template shown in Table 1 above, the scoring elements for the content quality scoring dimension are the title and body content of the data to be scored. The scoring method involves scoring from multiple aspects, including body format standardization, title-body relevance, compression rate (content duplication), and content fluency. Then, the scores from each aspect are combined (e.g., weighted summation) to obtain data D. i The title and body content, based on data D i The title and body content are scored according to the scoring method described above, resulting in data D. iThe score corresponding to the content quality scoring dimension.

[0153] For example, one scoring dimension in the target re-ranking scoring strategy is the website authority scoring dimension, which is applied to data D. i When scoring the website authority rating dimension, the scoring elements and scoring method of the website authority rating dimension are determined according to the pre-configured re-ranking scoring quantification template. Based on the re-ranking scoring quantification template shown in Table 1 above, the scoring element of the website authority rating dimension can be determined as the source website, and the scoring method is to determine the authority level of the source website and determine the score based on the authority level of the source website. Thus, data D is obtained. i The source website, confirming data D i The authority level of the source website, according to data D i The authority level of the source website determines the data D. i The score corresponds to the website authority rating dimension. It should be noted that several authoritative websites can be pre-acquired, categorized into levels of authority, and a corresponding score assigned to each level. This is then applied to data D. i When scoring a website's authority, data D can be determined. i The authority level of the source website will determine the data D i The score corresponding to the authority level of the source website is determined as Data D. i The score corresponding to the website authority rating dimension.

[0154] Step b2: Apply the weights corresponding to the scoring dimensions to the data D. i The data D is obtained by weighted summation of the scores corresponding to each of the several dimensions. i The score is obtained by weighted summation.

[0155] like Figure 4 As shown, after obtaining the weights corresponding to several scoring dimensions, a weighted summation method is used to sum the data D. i The scores are fused according to several dimensions.

[0156] Step b3, transfer data D i The weighted sum of the scores and data D i The gear positions are merged to obtain data D. i The corresponding target score.

[0157] In obtaining data D i After weighted summation and scoring, the data D is... i The weighted sum of the scores and data D i The gear positions are merged to obtain data D. i The corresponding target score, such as Figure 4 As shown.

[0158] Among them, data D i The gears are divided into data D i The corresponding score range for the relevance score.

[0159] Data D i The weighted sum of the scores and data D i The method of merging gear positions can be, but is not limited to, direct summation.

[0160] Considering that sorting by weighted summation might result in lower-level data appearing before higher-level data, this embodiment introduces a tier score to avoid this situation. The tier score corresponding to the higher-level scoring range is greater than the tier score corresponding to the lower-level scoring range. For example, the tier score corresponding to [0.95, 1] ​​is greater than the tier score corresponding to [0.9, 0.95).

[0161] The following section uses the target re-ranking scoring strategy, which includes relevance scoring, summary scoring, timeliness scoring, source site scoring, author scoring, content quality scoring, and website authority scoring, as an example to introduce how to determine the score levels.

[0162] In one possible implementation, the upper limit of the total score for the scoring dimensions other than the relevance scoring dimension can be calculated based on the upper limit of the relevance scoring range (e.g., 1) and the weights corresponding to the other scoring dimensions besides the relevance scoring dimension. Based on the upper limit of the total score, the tier scores corresponding to several scoring intervals of the preliminary search results can be determined, with higher tier scores corresponding to higher tier scores.

[0163] The process of calculating the upper limit of the total score for the scoring dimensions other than the relevance scoring dimension, based on the upper limit of the relevance scoring range and the weights corresponding to the other scoring dimensions, includes: weighting the weights corresponding to the other scoring dimensions with the upper limit of the score and summing the weights to obtain the upper limit of the total score for the other scoring dimensions.

[0164] Through the above-described re-sorting and scoring process, the target scores corresponding to several data points in the preliminary search results can be obtained.

[0165] After obtaining the target scores corresponding to the several data points contained in the preliminary search results, the several data points contained in the preliminary search results can be sorted according to the target scores corresponding to the several data points contained in the preliminary search results to obtain the search reordering results corresponding to the query question.

[0166] Optionally, the rearrangement strategy provided in this application can be used for multiple rounds of rearrangement. For example, the rearrangement strategy provided in this application can be used for the first round of rearrangement based on the preliminary search results, and the rearrangement strategy provided in this application can be used for the second round of rearrangement based on the results of the first round of rearrangement, and so on.

[0167] Based on the retrieval rearrangement method provided in the above embodiments, this application also provides a retrieval method, which may include:

[0168] Step c1: Obtain the query question.

[0169] Step c2: Using the retrieval rearrangement method provided in the above embodiments, obtain the retrieval rearrangement results corresponding to the query question.

[0170] Step c3: Based on the retrieval rearrangement results corresponding to the query question, determine the final retrieval results corresponding to the query question.

[0171] For example, the topM results (where M is less than the total number of data contained in the reordered search results) in the search reordered results can be used as the final search results corresponding to the query question, or the results in the search reordered results where the target score is greater than a preset score threshold can be used as the final search results corresponding to the query question.

[0172] The retrieval method provided in this embodiment can obtain retrieval results that meet user needs, resulting in a better user experience.

[0173] This application also provides a retrieval rearrangement device, such as... Figure 5 As shown, the retrieval rearrangement device may include: a relevance retrieval unit 501 and a retrieval result rearrangement unit 502.

[0174] The relevance retrieval unit 501 is used to obtain preliminary retrieval results for the query question.

[0175] The search result rearrangement unit 502 is used to rearrange the preliminary search results for the query question.

[0176] The retrieval result rearrangement unit may include: a question type identification unit 5021, a rearrangement scoring strategy acquisition unit 5022, a rearrangement scoring unit 5023, and a sorting unit 5024.

[0177] The problem type identification unit 5021 is used to determine the problem type of the query problem in the business dimension and obtain the target problem type.

[0178] The reordering scoring strategy acquisition unit 5022 is used to acquire a reordering scoring strategy that is compatible with the target problem type, thereby obtaining the target reordering scoring strategy.

[0179] The rearrangement scoring unit 5023 is used to score several data items contained in the preliminary search results according to the target rearrangement scoring strategy, so as to obtain the target scores corresponding to the several data items contained in the preliminary search results.

[0180] The sorting unit 5024 is used to sort the data in the preliminary search results according to the target scores corresponding to the data in the preliminary search results.

[0181] In one possible implementation, the target reordering scoring strategy includes: several scoring dimensions in a general domain, and weights corresponding to the target problem type for each of the several scoring dimensions in the general domain.

[0182] In one possible implementation, the preliminary search results contain several data points, each corresponding to a relevance score. The re-ranking and scoring unit 5023, according to the target re-ranking and scoring strategy, scores each of the several data points in the preliminary search results, including:

[0183] Obtain the relevance score range of the preliminary search results;

[0184] By categorizing the relevance scoring ranges, several scoring intervals are obtained that comprise the preliminary search results.

[0185] For each data point whose relevance score falls within the same scoring range, a score is assigned based on the target re-ranking scoring strategy.

[0186] In one possible implementation, the process by which the reordering scoring strategy acquisition unit 5022 acquires a reordering scoring strategy that is compatible with the target problem type includes:

[0187] From the pre-configured re-ranking scoring strategies for several problem types, obtain the re-ranking scoring strategy pre-configured for the target problem type, and obtain the re-ranking scoring strategy adapted to the target problem type. The several problem types are problem types at the business dimension.

[0188] In another possible implementation, the process by which the reordering scoring strategy acquisition unit 5022 acquires a reordering scoring strategy that is compatible with the target problem type includes:

[0189] Obtain the pre-configured basic re-ranking scoring strategy, which includes several scoring dimensions in the general domain and the basic weights corresponding to each of the scoring dimensions in the general domain.

[0190] Determine the scoring dimensions relevant to the target problem type;

[0191] Increase the weight of the scoring dimensions related to the target problem type in the basic reordering scoring strategy;

[0192] The weighted reordering scoring strategy is determined as the reordering scoring strategy that is suitable for the target problem type.

[0193] In one possible implementation, the target reordering scoring strategy includes several scoring dimensions and corresponding weights for each dimension. The reordering scoring unit 5023 scores each data point whose relevance scores fall within the same scoring range according to the target reordering scoring strategy, including the following steps:

[0194] The data is scored across several scoring dimensions to obtain the scores corresponding to the data across each of the scoring dimensions.

[0195] The scores of the data point are weighted and summed according to the weights corresponding to the various scoring dimensions to obtain the weighted sum score of the data point.

[0196] The weighted sum of the scores for this data point is combined with the tier score corresponding to the scoring range in which the data point's relevance score falls to obtain the target score for this data point. The tier score is determined based on the upper limit of the relevance scoring range and the weights corresponding to several scoring dimensions.

[0197] In one possible implementation, the process by which the reordering scoring unit 5023 scores the data item across several scoring dimensions to obtain the scores corresponding to the data item across those dimensions includes:

[0198] Obtain a pre-configured re-ranking scoring quantization template, which includes several scoring dimensions in a general domain, as well as the scoring elements and scoring methods for each scoring dimension;

[0199] For each scoring dimension in the target re-ranking scoring strategy, the scoring elements and scoring methods for that scoring dimension are determined based on the re-ranking scoring quantification template.

[0200] Based on the scoring elements and scoring method of this scoring dimension, the data is scored to obtain the score corresponding to this data in this scoring dimension.

[0201] In one possible implementation, several scoring dimensions in the general domain include some or all of the following scoring dimensions: relevance scoring dimension, abstract scoring dimension, title similarity scoring dimension, timeliness scoring dimension, source site scoring dimension, author scoring dimension, content quality scoring dimension, website authority scoring dimension, and keyword overlap scoring dimension.

[0202] In one possible implementation, the process by which the reordering scoring unit 5023 obtains several scoring intervals included in the preliminary search results by classifying the relevance scoring range includes:

[0203] By classifying the relevance scoring range according to the first or second classification method, several scoring intervals of the preliminary search results can be obtained.

[0204] The first grading method is to divide the scoring range to be graded according to a preset grading width to obtain several scoring intervals with the same span. The second grading method is to divide the scoring range to be graded with the goal of balancing the amount of data corresponding to different scoring intervals to obtain several scoring intervals. The amount of data corresponding to a scoring interval is the number of data whose relevance scores are located in that scoring interval.

[0205] In one possible implementation, the relevance scores corresponding to several data points in the preliminary search results are normalized relevance scores, and the relevance scoring range is the normalized relevance scoring range. The process by which the re-ranking scoring unit 5023 categorizes the relevance scoring range according to a first categorization method or a second categorization method to obtain several scoring intervals for the preliminary search results includes:

[0206] By directly classifying the normalized relevance scoring range according to the first or second classification method, several scoring intervals of the preliminary search results can be obtained.

[0207] Alternatively, the normalized relevance scores of the data points included in the preliminary search results can be restored to their unnormalized relevance scores. Based on the unnormalized relevance scores of the data points included in the preliminary search results, the unnormalized scoring range corresponding to the normalized scoring range can be determined. The unnormalized scoring range can then be divided into several scoring intervals of the preliminary search results by using either the first or second grading method.

[0208] The retrieval rearrangement device provided in this application embodiment is applicable to general fields and can provide retrieval rearrangement results that are more in line with user needs.

[0209] This application embodiment also provides an electronic device, including at least one processor and a memory connected to the processor, wherein:

[0210] The memory is used to store computer programs;

[0211] The processor is used to execute the computer program so that the electronic device can implement the steps of the retrieval rearrangement method provided in the above embodiments, or implement the steps of the retrieval method provided in the above embodiments.

[0212] This disclosure also provides a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, implement the steps of the retrieval and rearrangement method provided in the above embodiments, or implement the steps of the retrieval method provided in the above embodiments.

[0213] This application also provides a computer program product, including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement the steps of the retrieval rearrangement method provided in the above embodiments, or to implement the steps of the retrieval method provided in the above embodiments.

[0214] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0215] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0216] In the above embodiments, the implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, in the form of a computer program product.

[0217] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A retrieval and rearrangement method, characterized in that, include: Obtain preliminary search results for the query question; Determine the question type of the query in the business dimension to obtain the target question type; Obtain a reordering scoring strategy that is compatible with the target problem type to obtain the target reordering scoring strategy; According to the target reordering scoring strategy, the data contained in the preliminary search results are scored respectively to obtain the target scores corresponding to the data. The data are sorted according to the target scores corresponding to the data.

2. The retrieval and rearrangement method according to claim 1, characterized in that, The preliminary search results contain several data points, each corresponding to a relevance score; The step of scoring several data entries contained in the preliminary search results according to the target reordering scoring strategy includes: Obtain the relevance score range of the preliminary search results; By categorizing the relevance scoring range, several scoring intervals are obtained for the preliminary search results; For each data point whose relevance score falls within the same scoring range, a score is assigned to that data point according to the target re-ranking scoring strategy.

3. The retrieval and rearrangement method according to claim 1, characterized in that, The target reordering scoring strategy includes: several scoring dimensions in a general domain, and weights corresponding to the several scoring dimensions in the general domain that are adapted to the target problem type.

4. The retrieval and rearrangement method according to claim 1, characterized in that, The process of obtaining a re-ranking scoring strategy that is compatible with the target question type includes: From the pre-configured re-ranking scoring strategies for several problem types, obtain the re-ranking scoring strategy pre-configured for the target problem type, and obtain the re-ranking scoring strategy adapted to the target problem type, wherein the several problem types are problem types in the business dimension.

5. The retrieval and rearrangement method according to claim 1, characterized in that, The process of obtaining a re-ranking scoring strategy that is compatible with the target question type includes: Obtain a pre-configured basic re-ranking scoring strategy, which includes several scoring dimensions in a general domain and basic weights corresponding to the several scoring dimensions in the general domain. Determine the scoring dimensions associated with the target problem type; Increase the weight of the scoring dimensions related to the target problem type in the basic reordering scoring strategy; The weighted reordering scoring strategy is determined as the reordering scoring strategy that is suitable for the target problem type.

6. The retrieval and rearrangement method according to claim 2, characterized in that, The target reordering scoring strategy includes several scoring dimensions and the weights corresponding to each of the several scoring dimensions; The step of scoring the data according to the target reordering scoring strategy includes: The data is scored according to the scoring dimensions to obtain the scores corresponding to the data in each of the scoring dimensions. Based on the weights corresponding to the aforementioned scoring dimensions, the scores of the data corresponding to the aforementioned scoring dimensions are weighted and summed to obtain the weighted sum score of the data. The weighted sum of the data points is combined with the tier score corresponding to the scoring range in which the data points' relevance score falls to obtain the target score for the data points. The tier score is determined based on the upper limit of the relevance scoring range and the weights corresponding to the various scoring dimensions.

7. The retrieval and rearrangement method according to claim 6, characterized in that, The process of scoring the data item across the several scoring dimensions to obtain the score corresponding to the data item across each of the several scoring dimensions includes: Obtain a pre-configured re-ranking scoring quantization template, which includes several scoring dimensions in a general domain, as well as scoring elements and scoring methods for each scoring dimension; For each scoring dimension in the target re-ranking scoring strategy, the scoring elements and scoring method for that scoring dimension are determined according to the re-ranking scoring quantification template. Based on the scoring elements and scoring method of this scoring dimension, the data is scored to obtain the score corresponding to this data in this scoring dimension.

8. The retrieval rearrangement method according to claim 3, 5, or 7, characterized in that, The scoring dimensions of the general domain include some or all of the following scoring dimensions: The scoring dimensions include relevance scoring, abstract scoring, title similarity scoring, timeliness scoring, source site scoring, author scoring, content quality scoring, website authority scoring, and keyword overlap scoring.

9. The retrieval and rearrangement method according to claim 2, characterized in that, The step of dividing the relevance scoring range into tiers to obtain several scoring intervals for the preliminary search results includes: By classifying the relevance scoring range according to the first classification method or the second classification method, several scoring intervals of the preliminary search results are obtained. The first grading method involves dividing the scoring range to be graded according to a preset grading width to obtain several scoring intervals with the same span. The second grading method involves dividing the scoring range to be graded with the goal of balancing the amount of data corresponding to different scoring intervals to obtain several scoring intervals. The amount of data corresponding to a scoring interval is the number of data whose relevance scores are located in that scoring interval.

10. The retrieval and rearrangement method according to claim 9, characterized in that, The relevance scores corresponding to the several data points included in the preliminary search results are normalized relevance scores, and the relevance score range is the normalized relevance score range. The step of dividing the relevance scoring range according to a first or second grading method to obtain several scoring intervals for the preliminary search results includes: By directly classifying the normalized relevance scoring range according to the first classification method or the second classification method, several scoring intervals of the preliminary search results are obtained. Alternatively, the normalized relevance scores corresponding to the several data points included in the preliminary search results can be restored to the unnormalized relevance scores. Based on the unnormalized relevance scores corresponding to the several data points included in the preliminary search results, the unnormalized scoring range corresponding to the normalized scoring range can be determined. The unnormalized scoring range can be divided into several scoring intervals of the preliminary search results by using either the first or second grading method.

11. A retrieval method, characterized in that, include: Get the query question; The retrieval rearrangement method as described in any one of claims 1 to 10 is used to obtain the retrieval rearrangement result corresponding to the query question; Based on the retrieval rearrangement results corresponding to the query question, the final retrieval result corresponding to the query question is determined.

12. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is configured to execute the computer program to enable the electronic device to implement the steps of the retrieval rearrangement method as described in any one of claims 1 to 10, or to implement the steps of the retrieval method as described in claim 11.

13. A computer storage medium, characterized in that, The storage medium carries one or more computer programs that, when executed by an electronic device, enable the electronic device to implement the steps of the retrieval rearrangement method as described in any one of claims 1 to 10, or the steps of the retrieval method as described in claim 11.

14. A computer program product, characterized in that, It includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to perform the steps of the retrieval rearrangement method as described in any one of claims 1 to 10, or to perform the steps of the retrieval method as described in claim 11.

Images