Multipath recall method, training method and processing equipment

Abstract

The invention provides a multi-path recall method which is applied to the field of artificial intelligence. The multi-path recall method comprises the following steps that a processing device obtains a word to be retrieved; the processing device obtains M similarities between the to-be-searched word and the M training words; the processing equipment obtains a target weight group in a first weight set according to the training word corresponding to the maximum similarity in the M similarities, the first weight set comprises M weight groups, and each weight group in the M weight groups comprises N weights; the processing equipment inputs the to-be-retrieved word into N single-path recall models to obtain N first recall results; and the processing device fuses the N first recall results according to the target weight group to obtain a second recall result. In the technical scheme provided by the invention, through similarity matching, corresponding weights can be adaptively configured for different single-path recall models, and the recall effect is improved, so that the user experience is improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to multi-channel recall methods, training methods, and processing devices. Background Technology

[0002] The retrieval model is a core technology of search engines. Its main function is to retrieve thousands of relevant documents from a candidate set of hundreds of billions based on the user's search query. These documents are then fed into a downstream ranking model, which sorts the retrieved documents according to their relevance and displays them to the user. In many search engine optimization (SEO) technologies, single-path retrieval models often suffer from performance limitations, resulting in poor retrieval effectiveness. Therefore, search engines frequently employ multi-path retrieval methods, utilizing multiple single-path retrieval models to perform retrieval separately, obtaining multiple retrieval results. These results are then merged to obtain the final retrieval result. For example, the score of a document in each single-path retrieval candidate set can be added to or multiplied by the scores of the same document in other single-path retrieval candidate sets, and the documents are then re-ranked based on the calculated scores.

[0003] In some scenarios, simple addition or multiplication calculations cannot provide good recall, thus reducing the user experience. Summary of the Invention

[0004] This application provides a multi-path recall method, training method, and processing device. Through similarity matching, it can adaptively configure corresponding weights for different single-path recall models, improve recall performance, and thus enhance user experience.

[0005] The first aspect of this application provides a multi-path recall method. The multi-path recall method is applied to a processing device. The multi-path recall method includes the following steps: the processing device obtains a term to be retrieved; the processing device obtains M similarities between the term to be retrieved and M training words, where M is greater than 1, and the M similarities correspond one-to-one with the M training words; the processing device obtains a target weight reassembly from a first weight set based on the training word corresponding to the maximum similarity among the M similarities, the first weight set including M weight reassemblies, each of the M weight reassemblies corresponding one-to-one with the M training words, each weight reassembly including N weights, each of the N weights corresponding one-to-one with N single-path recall models, where N is greater than 1; the processing device inputs the term to be retrieved into the N single-path recall models to obtain N first recall results, each of the N first recall results corresponding one-to-one with the N single-path recall models; the processing device fuses the N first recall results based on the target weight reassembly to obtain a second recall result.

[0006] In one alternative embodiment of the first aspect, the multi-path recall method further includes the following steps: the processing device obtains a second weight set; the processing device replaces the first weight set with the second weight set. By replacing the weight set, the recall effect of the multi-path recall method can be improved, thereby enhancing the user experience.

[0007] In one alternative approach of the first aspect, the multi-path recall method further includes the following step: if the M similarities include the same multiple maximum similarities, then increase the precision of the multiple maximum similarities. By increasing the precision, it is possible to avoid the occurrence of the same maximum similarity, which could lead to failure of the multi-path recall model.

[0008] In one alternative approach of the first aspect, the precision of the M similarities is less than 1 / M. By controlling the precision, the efficiency of the processing device in obtaining the maximum similarity can be improved, the efficiency of returning recall results can be increased, thereby enhancing the user experience.

[0009] In one optional approach of the first aspect, the processing device obtains M similarities between the search term and M training terms by: vectorizing the search term to obtain a first vector; and obtaining M similarities between the first vector and the M vectors, where each of the M vectors corresponds one-to-one with one of the M training terms. By vectorizing the search term and the training terms, the accuracy of similarity calculation can be improved, thereby enhancing recall performance and improving user experience.

[0010] The second aspect of this application provides a training method. The training method is applied to a training device. The multi-path recall method includes the following steps: the training device acquires a target weight set, which includes H weight combinations, each of the H weight combinations including N weights; the training device acquires K training data, each training data corresponding to multiple training words, the K training data corresponding to M training words, where K and M are greater than 1; the training device performs single-path recall on each of the M training words according to N single-path recall models, obtaining N recall results, with each of the N recall results corresponding one-to-one with one of the N single-path recall models; the training device determines the weight combinations in the H weight combinations corresponding to each training word based on the N recall results, and the M training words correspond to the M weight combinations in the H weight combinations.

[0011] In one alternative approach of the second aspect, the training device determines the weighted group among the H weighted groupings corresponding to each training word based on the N recall results, including: the training device acquiring N sets of relevance scores corresponding to the N recall results, each set of relevance scores including K relevance scores, and the K relevance scores corresponding to K training data in a one-to-one correspondence; the training device determining the weighted group among the H weighted groupings corresponding to each training word based on the K relevance scores and the training data corresponding to each training word.

[0012] In one optional approach of the second aspect, K training data include first training data and second training data, with each training word corresponding to the first training data; N relevance score sets include first relevance score sets and second relevance score sets; N single-path recall models include first single-path recall model and second single-path recall model; the first relevance score set is obtained based on the first single-path recall model; the second relevance score set is obtained based on the second single-path recall model; and the K relevance scores include first relevance scores and second relevance scores, with the first relevance score corresponding to the first training data and the second relevance score corresponding to the second training data. The training device determines the weighted reassembly among the H weighted reassemblies corresponding to each training word based on the K relevance scores and the training data corresponding to each training word, including: if in the first relevance score set, the first relevance score is greater than the second relevance score, and in the second relevance score set, the second relevance score is greater than the first relevance score, then the training device selects a target weighted reassembly from the H weighted reassemblies as the weighted reassembly corresponding to each training word, where the weight corresponding to the first single-path recall model in the target weighted reassembly is greater than the weight corresponding to the second single-path recall model.

[0013] In one alternative approach of the second aspect, the training device determines the weighted recombination among the H weighted recombinations corresponding to each training word based on the K relevance scores and the training data corresponding to each training word, including: if in the first relevance score set, the first relevance score is less than the second relevance score, and in the second relevance score set, the second relevance score is less than the first relevance score, then the training device selects the target weighted recombination from the H weighted recombinations as the weighted recombination corresponding to each training word, and the weight corresponding to the first single-path recall model in the target weighted recombination is less than the weight corresponding to the second single-path recall model.

[0014] In one alternative approach of the second aspect, the training device determines the weighting of the H weightings corresponding to each training word based on the K relevance scores and the training data corresponding to each training word, including: if the second relevance score is equal to the first relevance score in the first relevance score set and the second relevance score set, then the training device selects the target weighting as the weighting of each training word from the H weightings, and the weight corresponding to the first single-path recall model in the target weighting is equal to the weight corresponding to the second single-path recall model.

[0015] In one alternative approach to the second aspect, the value of H is between 1000 and 100000. By controlling the value of H, training efficiency can be improved while enhancing recall.

[0016] A third aspect of this application provides a multi-path recall device. The multi-path recall device includes an input module and a calculation module. The input module is used to obtain the term to be retrieved. The calculation module is used to obtain M similarities between the term to be retrieved and M training words. M is greater than 1. Each of the M similarities corresponds one-to-one with one of the M training words. The calculation module is also used to obtain a target weight reassembly in a first weight set based on the training word corresponding to the maximum similarity among the M similarities. The first weight set includes M weight reassemblies. Each of the M weight reassemblies corresponds one-to-one with the M similarities. Each weight reassembly includes N weights. Each of the N weights corresponds one-to-one with one of the N single-path recall models, where N is greater than 1. The calculation module is also used to input the term to be retrieved into the N single-path recall models to obtain N first recall results. Each of the N first recall results corresponds one-to-one with one of the N single-path recall models. The calculation module is also used to fuse the N first recall results based on the target weight reassembly to obtain a second recall result.

[0017] In an alternative approach to the third aspect, the input module is further configured to obtain a second weight set. The calculation module is further configured to replace the first weight set with the second weight set.

[0018] In one alternative approach of the third aspect, if the M similarities include the same multiple maximum similarities, the calculation module is also used to increase the precision of the multiple maximum similarities.

[0019] In one alternative approach of the third aspect, the calculation module is used to obtain M similarities between the term to be retrieved and M training terms, including: the calculation module vectorizing the term to be retrieved to obtain a first vector; the calculation module obtaining M similarities between the first vector and the M vectors, wherein the M vectors correspond one-to-one with the M training terms.

[0020] In one alternative approach to the third aspect, the precision of the M similarities is less than 1 / M.

[0021] A fourth aspect of this application provides a training apparatus. The training apparatus includes an input module and a calculation module. The input module is used to acquire a target weight set. The target weight set includes H weight reassemblies. Each weight reassembly in the H weight reassemblies includes N weights. The input module is also used to acquire K training data, each training data corresponding to multiple training words. The K training data correspond to M training words, where K and M are greater than 1. The calculation module is used to perform single-path recall on each of the M training words according to N single-path recall models, obtaining N recall results. The N recall results correspond one-to-one with the N single-path recall models. The calculation module is also used to determine the weight reassemblies in the H weight reassemblies corresponding to each training word based on the N recall results, where the M training words correspond to the M weight reassemblies in the H weight reassemblies.

[0022] In one optional approach of the fourth aspect, the calculation module is used to determine the weighted group among the H weighted groupings corresponding to each training word based on the N recall results. This includes: the calculation module obtaining N sets of relevance scores corresponding to the N recall results, each set of relevance scores including K relevance scores, and the K relevance scores corresponding to K training data in a one-to-one correspondence; the calculation module determining the weighted group among the H weighted groupings corresponding to each training word based on the K relevance scores and the training data corresponding to each training word.

[0023] In one optional approach of the fourth aspect, K training data include first training data and second training data, with each training word corresponding to the first training data; N relevance score sets include first relevance score sets and second relevance score sets; N single-path recall models include first single-path recall model and second single-path recall model; the first relevance score set is obtained based on the first single-path recall model; the second relevance score set is obtained based on the second single-path recall model; and the K relevance scores include first relevance scores and second relevance scores, with the first relevance score corresponding to the first training data and the second relevance score corresponding to the second training data. The calculation module is used to determine the weighted reassemblies among the H weighted reassemblies corresponding to each training word based on the K relevance scores and the training data corresponding to each training word. This includes: if, in the first relevance score set, the first relevance score is greater than the second relevance score, and in the second relevance score set, the second relevance score is greater than the first relevance score, then the calculation module is used to select a target weighted reassembly from the H weighted reassemblies as the weighted reassembly corresponding to each training word, where the weight corresponding to the first single-path recall model in the target weighted reassembly is greater than the weight corresponding to the second single-path recall model.

[0024] In one optional approach of the fourth aspect, the calculation module is used to determine the weighted recombination among the H weighted recombinations corresponding to each training word based on the K relevance scores and the training data corresponding to each training word, including: if in the first relevance score set, the first relevance score is less than the second relevance score, and in the second relevance score set, the second relevance score is less than the first relevance score, then the calculation module is used to select the target weighted recombination as the weighted recombination corresponding to each training word from the H weighted recombinations, wherein the weight corresponding to the first single-path recall model in the target weighted recombination is less than the weight corresponding to the second single-path recall model.

[0025] In one alternative approach of the fourth aspect, the training device determines the weighted recombination among the H weighted recombinations corresponding to each training word based on the K relevance scores and the training data corresponding to each training word, including: if the second relevance score is equal to the first relevance score in the first relevance score set and the second relevance score set, the calculation module is used to select the target weighted recombination as the weighted recombination corresponding to each training word in the H weighted recombinations, and the weight corresponding to the first single-path recall model in the target weighted recombination is equal to the weight corresponding to the second single-path recall model.

[0026] A fifth aspect of this application provides a processing apparatus. The processing apparatus includes a memory and a processor. The memory stores a program; the processor executes the program stored in the memory. When the program stored in the memory is executed, the processor performs the method described in or in any of the optional methods of the first aspect, or performs the method described in or in any of the optional methods of the second aspect.

[0027] The sixth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed on a processor, causes the processor to perform the methods described in the first aspect, any alternative to the first aspect, the second aspect, or any alternative to the second aspect.

[0028] The seventh aspect of this application provides a computer program product, characterized in that, when the computer program product is run on a processor, the processor executes the method described in the first aspect, any optional mode of the first aspect, the second aspect, or any optional mode of the second aspect.

[0029] An eighth aspect of this application provides a chip. The chip includes a processor and an interface. The processor is used to acquire program instructions or data through the interface; the processor is used to execute program line instructions to implement the methods described in the first aspect, any optional method of the first aspect, the second aspect, or any optional method of the second aspect.

[0030] It should be understood that the beneficial effects of the aforementioned third to eighth aspects can be found in the relevant descriptions in the first or second aspects above, and will not be repeated here. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the structure of the multi-channel recall system provided in the embodiments of this application;

[0032] Figure 2 A flowchart illustrating the training method provided in an embodiment of this application;

[0033] Figure 3A diagram showing the correspondence between training words and training data provided in the embodiments of this application;

[0034] Figure 4 This is a first flowchart illustrating the multi-path recall method provided in an embodiment of this application;

[0035] Figure 5 A second flowchart illustrating the multi-channel recall method provided in this application embodiment;

[0036] Figure 6 This is a first structural schematic diagram of the processing device provided in an embodiment of this application;

[0037] Figure 7 This is a second structural schematic diagram of the processing device provided in an embodiment of this application;

[0038] Figure 8 This is a schematic diagram of the structure of a chip provided in an embodiment of this application. Detailed Implementation

[0039] The term "and / or" in this document describes an association relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three scenarios: A alone, A and B simultaneously, and B alone. The terms "first," "second," or "target," etc., in the specification and claims are used to distinguish different objects, not to describe a specific order of objects. For example, "first recall result" and "second recall result," etc., are used to distinguish different recall results, not to describe a specific order of recall results.

[0040] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner. In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple training words refer to two or more training words, etc.

[0041] First, let me introduce the technical terms used in this application.

[0042] (1) Recall (match): refers to filtering as much correct information as possible from the full set of information based on the search terms.

[0043] (2) Recall model: refers to a neural network model with recall capability.

[0044] (3) Single-path recall model; refers to a single recall model.

[0045] (4) Multi-path recall model: refers to a combined model composed of multiple single-path recall models, which can use multiple single-path recall models to perform single-path recall respectively, and output the results of each single-path recall after merging them.

[0046] Generally, commonly used recall models include sparse models and dense models. Sparse models offer better scalability and interpretability, but suffer from word matching errors, leading to poor recall performance. Therefore, while sparse models can efficiently support indexes of billions of documents, they only fulfill basic recall functions, with poor accuracy in matching retrieved documents with search terms. Dense models, on the other hand, offer better semantic matching performance but lack interpretability and have lower data scalability than sparse models. Therefore, they are only suitable for information retrieval on small datasets and perform poorly for industrial-scale document retrieval.

[0047] To address this, leveraging the complementary capabilities of these two models in semantic matching, precise matching, efficiency, interpretability, and maintainability, a multi-path recall approach can be employed. This involves training a sparse representation model and a dense representation model separately, then ranking the recall results from both models together as the final multi-path recall result. For example, the score of a document in each single-path recall candidate set can be added to or multiplied by the scores of the same document in other single-path recall candidate sets, and the results can be re-ranked based on the calculated scores. However, in some scenarios, simple addition or multiplication calculations may not provide satisfactory recall results, thus degrading the user experience.

[0048] Therefore, this application provides a training method and a multi-path recall method. The training method is applied to a training device, and the multi-path recall method is applied to a processing device. Figure 1 This is a schematic diagram of the structure of a multi-channel recall system provided in an embodiment of this application. Figure 1As shown, the multi-path recall system 100 includes a training device 101 and a processing device 102. The training device 101 and processing device 102 can be any device, equipment, platform, or cluster of devices with computing and processing capabilities. The training device 101 is used to train a multi-path recall model using a training method to obtain M training words and corresponding M weighted recombinations. The processing device 102 is used to configure the multi-path recall model, the M training words, and the corresponding M weighted recombinations. The processing device 102 is used to perform single-path recall on the target word based on the N recall models to obtain N first recall results. The processing device 102 is also used to obtain second recall results based on the M training words, the M weighted recombinations, the target word, and the N first recall results. It should be understood that in practical applications, the training device 101 and the processing device 102 can be the same device. In this case, the processing device 102 is also used to execute the steps executed by the training device 101.

[0049] To facilitate understanding of the technical solutions in the embodiments of this application, some scenarios of multi-path recall using the multi-path recall model in the embodiments of this application will be described below. For example, the multi-path recall model in the embodiments of this application can be applied in a machine learning-based search engine recall scenario. The general process of multi-path recall using the multi-path recall model in a search engine is as follows: the user inputs the query they wish to query, i.e., the search term or the term to be searched; then the search engine obtains multiple index query results through each single-path recall model; subsequently, the results of each index are merged and filtered to obtain relevant documents, which are then input into downstream tasks.

[0050] For example, the multi-path recall model can be applied to search recall scenarios in large search engines, information feeds, or service platforms or applications (Apps), which will be described below.

[0051] In the context of large-scale search applications, "large-scale search" refers to the use of keyword search to display relevant information. For example, processing device 102 matches and displays web page information that the user might be interested in based on their search query. For instance, training device 101 collects web page data from the internet to form a training dataset, and after preprocessing, creates an indexer for this training dataset. When a user enters keywords (i.e., search terms, hereinafter the same) through the search box provided on the search engine page, such as "patent application," the search engine's multi-path recall model retrieves the keyword. Through multi-path recall, the indexer accesses the training dataset, recalling web page information related to the keyword "patent application," such as information from official patent application websites, patent search websites, and patent agency websites. Furthermore, the multi-path recall model assigns relevance scores to the recalled web page information, which are then fused and fed to the downstream fine-ranking model. Finally, by sorting all the recalled web page information according to their corresponding relevance scores, the ranking results are output to the search page for display.

[0052] In information feed applications, information feeds utilize user preferences and other characteristics to display relevant information. For example, if a user's preference characteristics include "music," then after the multi-path recall model obtains the keyword "music," it can perform multi-path recall to retrieve and display data about songs, instruments, and concerts from the database based on information that other users with the same preference characteristic are interested in, such as songs, instruments, and concert information.

[0053] In service platform application scenarios, service platforms can include platforms that specifically provide certain services, such as internet resource search platforms, shopping platforms, and online video media platforms. Service platforms can collect users' keywords or preference characteristics and use a multi-path recall model to retrieve relevant data from corresponding databases for display. For example, if a user's historical keyword input on a shopping platform includes "backpack," the multi-path recall model can retrieve several shopping web pages about backpacks and display them to the user. Similarly, if a user enters keywords such as the name of a TV series or a character name on a webpage provided by an online video media platform, the corresponding multi-path recall model can retrieve video or audio webpages related to those keywords and display them to the user.

[0054] In the context of in-app search, apps can include app stores, video apps, music apps, news apps, browsers, and other applications deployed on the device. These apps can collect user keywords or preference features and retrieve relevant data from corresponding databases using a multi-path retrieval model for display.

[0055] In some examples, multi-path recall models can be applied to business systems such as question-answering systems and recommendation systems to recall relevant data information based on keywords or preference features input by users.

[0056] Figure 2 This is a flowchart illustrating the training method provided in an embodiment of this application. Figure 2 As shown, the training method includes the following steps.

[0057] In step 201, the training device acquires a target weight set, which includes H weight reassemblies. Each of the H weight reassemblies includes N weights. The value of N is the number of single-path recall models in the multi-path recall model. In the following examples, N equals 2. The sum of the N weights equals 1. Each of the N weights represents the score proportion of the corresponding single-path recall model. The value of H depends on the application scenario of the multi-path recall model. For example, if H equals 9, the H weight reassemblies are (0.9, 0.1), (0.8, 0.2), (0.7, 0.3)...(0.2, 0.8) and (0.1, 0.9). Similarly, if H equals 99, the H weight reassemblies are (0.99, 0.01), (0.98, 0.02), (0.97, 0.03)...(0.02, 0.98) and (0.01, 0.99). For example, if H equals 999, the H weighted recombinations are (0.999, 0.001), (0.998, 0.002), (0.997, 0.003)...(0.002, 0.998) and (0.001, 0.999). In practical applications, an excessively large H value will reduce training efficiency, while an excessively small H value will reduce recall. To improve training efficiency while enhancing recall, the value of H can be between 1000 and 100000. In the following examples, we will use H equal to 9 as an example.

[0058] In step 202, the training device acquires K training data, each training data corresponds to multiple training words, and the K training data correspond to M training words.

[0059] Training data can also be referred to as the original data source or the data document to be searched. K training data points are denoted as D. k D k ={Doc1, Doc2, ... Doc k K is greater than 1. In applications based on multi-path recall models, the training data can be diverse. For example, in a large search application, the training data can be web pages. In subsequent examples, we will use K equal to 2 as an example. The training device generates a set of several queries, i.e., multiple training words, based on each training data set. For example, Figure 3 This is a diagram showing the correspondence between training words and training data provided in an embodiment of this application. Figure 3 In the example, the K training data points include training data 1 and training data 2. The training device generates training words 1-3 using training data 1, meaning training data 1 corresponds to training words 1-3. The training device generates training words 4-7 using training data 2, meaning training data 2 corresponds to training words 4-7. The K training data points correspond to M training words (301). M is greater than 1. Figure 3 In the example, M equals 7. In one example, the training device is able to perform inference using a large language model (LLM) to generate training words corresponding to the training data.

[0060] The training device vectorizes M training words, resulting in M ​​vectors. Each of the M vectors corresponds one-to-one with a training word. For example... Figure 3 As shown, the M vectors 302 include vectors 1 through 7. Vector 1 is generated from training word 1, meaning vector 1 corresponds to training word 1. Similarly, the other vectors in the M vectors 302 correspond one-to-one with the other training words in the M training words.

[0061] In step 203, the training device performs single-path recall on each of the M training words according to the N single-path recall models, and obtains N recall results. The N recall results correspond one-to-one with the N single-path recall models.

[0062] The N single-path recall models include a first single-path recall model and a second single-path recall model. The first single-path recall model can be a sparse representation model or a dense representation model. For example, the first single-path recall model is a dense representation model. In this example, the first single-path recall model can include an input layer, a representation layer, and a matching layer. The input layer, also known as the token embedding layer, maps the text to a low-dimensional vector space, transforming it into word vectors that are provided to the representation layer. This text includes training words and K training data from the document library. It should be understood that the training device can vectorize M training words through the input layer to obtain M vectors. The representation layer can be used to represent word vectors using a neural network; representation is the construction from words to sentences. The matching layer can be used to semantically match the vectors output by the representation layer (including the representation vectors corresponding to the training words and the representation vectors corresponding to the K training data) and score the relevance between the training words and the matched training data, obtaining a corresponding relevance score. Relevance refers to the similarity between training words and training data. Therefore, the relevance score represents the degree of matching between training words and training data. Scoring can be achieved by calculating the cosine similarity or Euclidean similarity between the representation vectors corresponding to the training words and the representation vectors corresponding to the training data, or through inner product operations.

[0063] The second single-path recall model can be of the same type as the first single-path recall model, or it can be a different type; this embodiment does not impose any limitations. For example, the first single-path recall model is a sparse representation model, and the second single-path recall model is either a sparse representation model or a dense representation model. For example, the first single-path recall model is an ES keyword recall model, and the second single-path recall model is a VS vectorized two-path recall model.

[0064] The training device performs single-path recall on each of the M training words using N single-path recall models, resulting in N recall results. Each of the N recall results corresponds one-to-one with one of the N single-path recall models. For example, for training word 1, the training device obtains N recall results. These N results include recall result 1 and recall relevance result 2. Each of the N recall results corresponds one-to-one with N sets of relevance scores. Each set of relevance scores includes K relevance scores, and each of the K scores corresponds one-to-one with K training data points. For example, the N sets of relevance scores include a first relevance score set T1 and a second relevance score set T2. T1 is obtained from the first single-path recall model, and T2 is obtained from the second single-path recall model. T1 is shown in Table 1, and T2 is shown in Table 2. The relevance scores in the tables are sorted in descending order. For example, in Table 1, ES1 is greater than ES2. Doc1 is also called training data 1 or the first training data. Doc2 is also called training data 2 or the second training data.

[0065] Training data Correlation score Doc1 ES1 score Doc2 ES2 score

[0066] Table 1

[0067] Training data Correlation score Doc2 VS1_Score Doc1 VS2 Score

[0068] Table 2

[0069] In step 204, the training device determines the weighted recombinations in the H weighted recombinations corresponding to each training word based on the N recall results, and the M weighted recombinations in the H weighted recombinations corresponding to the M training words.

[0070] In step 203 above, the training device obtains N relevance score sets for each training word. The training device can then determine the weighted recombinations among the H weighted recombinations corresponding to each training word based on the K relevance scores and the training data corresponding to each training word. Figure 3 As can be seen from the correspondence, training word 1 is generated from training data 1. Therefore, among the N relevance score sets, when the relevance score of training data 1 is higher than that of training data 2, the recall result corresponding to the relevance score set is more accurate. That is, the single-path recall model that obtains this relevance score set should be assigned a higher weight. Based on this principle, the following describes the different possible scenarios.

[0071] In the first scenario, the first relevance score set T1 is shown in Table 1, and the second relevance score set T2 is shown in Table 2. At this time, the accuracy of the first single-path recall model is higher. Therefore, the training device assigns a higher weight to training word 1 of the first single-path recall model. The training device can then select one weight reassembly from the following H weight reassemblies. In the selected weight reassemblies, the weight of the first single-path recall model is greater than the weight of the second single-path recall model. According to the description of step 201 above, the weight reassemblies that satisfy the above conditions include (0.9, 0.1), (0.8, 0.2), (0.7, 0.3), and (0.6, 0.4). The first value in the weight reassembly represents the weight of the first single-path recall model, and the second value represents the weight of the second single-path recall model. The training device can randomly select one weight reassembly from the above weight reassemblies as the weight reassembly corresponding to training word 1. For example, the training device selects weight reassembly 1 as the weight reassembly corresponding to training word 1. Reassembly 1 is (0.9, 0.1).

[0072] In practical applications, training devices can also select a suitable weighted reassembly from multiple weighted reassemblies that meet certain conditions using certain algorithms. For example, the training device obtains multiple DeltaWs. DeltaW = (X1 × ES1 - X2 × ES2). Here, X1 represents the first value in the weighted reassembly. X2 represents the second value in the weighted reassembly. ES1 and ES2 are the scores shown in Table 1. The number of DeltaWs corresponds one-to-one with the number of weighted reassemblies that meet the above conditions. In the example above, the training device obtains 4 DeltaWs. The 4 DeltaWs correspond to (0.9, 0.1), (0.8, 0.2), (0.7, 0.3), and (0.6, 0.4), respectively. The 4 DeltaWs include DeltaW0 to 3. DeltaW0 is obtained from (0.9, 0.1), that is, DeltaW0 corresponds to (0.9, 0.1). DeltaW0 = (0.9 × ES1 - 0.1 × ES2). The training device selects the weighted combination corresponding to the maximum value among multiple DeltaW as the weighted combination corresponding to training word 1. It should be understood that this application embodiment provides an exemplary description of a specific algorithm for the training device to select a weighted combination from multiple weighted combinations. In practical applications, those skilled in the art can design suitable algorithms to select a weighted combination from multiple weighted combinations according to their needs.

[0073] In the second scenario, the first relevance score set T1 is shown in Table 3, and the second relevance score set T2 is shown in Table 4. In this case, the accuracy of the second single-path recall model is higher. Therefore, the training device assigns a higher weight to training word 1 for the second single-path recall model. At this time, the training device can select one weight reassembly from the following H weight reassemblies. In the selected weight reassemblies, the weight of the second single-path recall model is greater than the weight of the first single-path recall model. According to the description of step 201 above, the weight reassemblies that satisfy the above conditions include (0.4, 0.6), (0.3, 0.7), (0.2, 0.8), and (0.1, 0.9). The first value in the weight reassembly represents the weight of the first single-path recall model, and the second value in the weight reassembly represents the weight of the second single-path recall model. The training device can randomly select one weight reassembly from the above weight reassemblies as the weight reassembly corresponding to training word 1.

[0074] In practical applications, training devices can also select a suitable weighted reassembly from multiple weighted reassemblies that meet certain conditions using certain algorithms. For example, the training device obtains multiple DeltaWs. DeltaW = (X1 × VS1 - X2 × VS2). Where X1 represents the first value in the weighted reassembly. X2 represents the second value in the weighted reassembly. VS1 and VS2 are scores as shown in Table 4. The number of DeltaWs corresponds one-to-one with the number of weighted reassemblies that meet the above conditions. In the example above, the training device obtains 4 DeltaWs. The 4 DeltaWs correspond to (0.4, 0.6), (0.3, 0.7), (0.2, 0.8), and (0.1, 0.9), respectively. The 4 DeltaWs include DeltaW0 to 3. DeltaW0 is obtained from (0.4, 0.6), that is, DeltaW0 corresponds to (0.4, 0.6). DeltaW0 = (0.4 × ES1 - 0.6 × ES2). The training device selects the weighted recombination corresponding to the maximum value among multiple DeltaW as the weighted recombination corresponding to training word 1.

[0075] Training data Correlation score Doc2 ES1 score Doc1 ES2 score

[0076] Table 3

[0077] Training data Correlation score Doc1 VS1_Score Doc2 VS2 Score

[0078] Table 4

[0079] In the third scenario, the first relevance score set T1 is shown in Table 1, and the second relevance score set T2 is shown in Table 4. Alternatively, the first relevance score set T3 is shown in Table 1, and the second relevance score set T2 is shown in Table 2. In this case, the accuracy of the second single-path recall model is the same as that of the first single-path recall model. The training device matches the same weights for training word 1 in both the second and first single-path recall models. In this case, the training device selects the weight recombination (0.5, 0.5) as the weight recombination corresponding to training word 1.

[0080] In the previous example, the training device matched the corresponding weighted recombination for training word 1 based on N recall results. Using a similar method, the training device matched M weighted recombinations for M training words. There is a one-to-one correspondence between the M training words and the M weighted recombinations. For example... Figure 3 As shown, the M weighted recombinations 303 include weighted recombinations 1 to 7. Weighted recombination 1 corresponds to vector 1 and training word 1. Weighted recombination 2 corresponds to vector 2 and training word 2.

[0081] It should be understood that, Figure 2 In the example, the training device is described with N equal to 2, M equal to 7, and H equal to 9. In practical applications, those skilled in the art can design specific values ​​for N or M according to requirements, and the embodiments of this application do not impose specific limitations. After obtaining the correspondence between M weighted recombinations and M vectors, the training device sends the correspondence, N single-path recall models, and K training data to the processing device. The processing device provides search services to the user based on the correspondence, N single-path recall models, and K training data. Alternatively, the training device acts as the processing device to provide search services to the user. The multi-path recall method provided by this application is described below. Alternatively, the training device sends the correspondence between M weighted recombinations and M training words, N single-path recall models, and K training data to the processing device. The processing device provides search services to the user based on the correspondence, N single-path recall models, and K training data.

[0082] Figure 4 This is a first flowchart illustrating the multi-channel recall method provided in an embodiment of this application. Figure 4 As shown, the multi-channel recall method includes the following steps.

[0083] In step 401, the processing device acquires the term to be searched.

[0084] The search term can be text information composed of characters, words, and / or sentences. In some possible implementations, the search term can also be information determined by extracting or recognizing content from images, audio, or video. For example, a user can enter the words "leaves" into the search box on a search page as a search term. Alternatively, a user can use the voice input function in the search box to pronounce "leaves," and the search engine can use voice recognition to identify the word "leaves" from the speech and use it as the search term. Or, for example, a user can upload an image containing a leaf graphic into the search box, and the search engine can use image recognition to identify the word "leaves" from the image and use it as the search term.

[0085] In step 402, the processing device obtains M similarities between the word to be retrieved and M training words.

[0086] This application does not limit the method by which the processing device obtains similarity. For example, the processing device vectorizes the term to be retrieved to obtain a first vector. The processing device obtains M similarities between the first vector and M vectors. The M vectors correspond one-to-one with the M training words. Similarity A1 is obtained by the processing device through the first vector and vector 1, that is, the first vector corresponds to vector 1. For example, Here, B represents the first vector. Emba query represents vector 1. Using a similar method, the processing device can obtain M similarity values.

[0087] The processing device determines the maximum similarity among M similarities. In practical applications, due to the limited precision of the M similarities obtained by the processing device, it is possible that the M similarities may contain the same magnitude of similarity. If the M similarities include multiple identical maximum similarities, the processing device can increase the calculation precision of these multiple maximum similarities. After increasing the calculation precision, the processing device determines the maximum similarity among the multiple maximum similarities. For example, before increasing the calculation precision, the precision of the multiple maximum similarities is 0.0001%. The multiple maximum similarities include a first similarity and a second similarity. Both the first and second similarities are 78.6761%. After increasing the calculation precision, the precision of the multiple maximum similarities is 0.000001%. The first similarity is 78.676155%. The second similarity is 78.676167%. Based on this, the processing device determines the second similarity as the maximum similarity. In this embodiment, the processing device only increases the calculation precision of the multiple maximum similarities, rather than increasing the precision of the M similarities. Therefore, this embodiment can save processing resources.

[0088] In practical applications, when the precision of similarity calculation is too low, the same similarity values ​​may appear repeatedly, increasing the processing efficiency of the processing device and affecting the recall efficiency. Therefore, before increasing the calculation precision, the precision of M similarities can be less than 1 / M. For example, when M is 10000, the precision of the similarity calculation by the processing device is 0.00001 or 0.000001.

[0089] In step 403, the processing device obtains the target weight recombination in the first weight set based on the training word corresponding to the maximum similarity among the M similarities. The first weight set includes M weight recombinations, and each of the M weight recombinations corresponds one-to-one with the M training words. Each weight recombination in the M weight recombinations includes N weights, and each of the N weights corresponds one-to-one with the N single-path recall models.

[0090] For a description of the model with M weighted recombinations, M training words, and N single-path recall, please refer to the aforementioned... Figure 2 The relevant description is as follows: In step 402 above, the processing device determines the maximum similarity among the M similarities. The processing device obtains the target weight reorganization in the first weight set based on the training word corresponding to the maximum similarity. Assume that the training word corresponding to the maximum similarity is training word 2. Training word 2 corresponds to weight reorganization 2. Therefore, weight reorganization 2 is the target weight reorganization corresponding to the maximum similarity.

[0091] In step 404, the processing device inputs the term to be retrieved into N single-path recall models and obtains N first recall results. The N first recall results correspond one-to-one with the N single-path recall models.

[0092] When N equals 2, the N single-path recall models include the first single-path recall model and the second single-path recall model. The N first recall results include first recall result 1 and first recall result 2. First recall result 1 is obtained by the first single-path recall model based on the search term. First recall result 2 is obtained by the second single-path recall model based on the search term.

[0093] In step 405, the processing device merges the N first recall results according to the target weight reorganization to obtain the second recall result.

[0094] In N first-recall results, the same document (or training data) corresponds to its own relevance score, also known as a score. The processing device processes these relevance scores according to the target weight reassembly to obtain a fusion score C. For example, C = ES × F1 + VS × F2. Where ES is the relevance score of the document in the first-recall results, VS is the relevance score of the document in the second-recall results, F1 is the weight of the first single-path recall model in the target weight reassembly, and F2 is the weight of the second single-path recall model in the target weight reassembly.

[0095] Assume the target weights are recombined as (0.3, 0.7). In Table 5, the first column represents the document index, the second column represents the first recall result, the third column represents the second recall result, and the fourth column represents the second recall result. At this point, the document score ranking in the second recall result is the same as the document score ranking in the first recall result 2, meaning the first recall result 2 is close to the optimal recall result.

[0096] Training data ES score VS score Integration score Doc1 0.6 0.9 0.81 Doc2 0.8 0.6 0.66 Doc3 0.4 0.5 0.47

[0097] Table 5

[0098] Assume the target weights are recombined as (0.8, 0.2). In Table 6, the first column represents the document index, the second column represents the first recall result, the third column represents the second recall result, and the fourth column represents the second recall result. At this point, the document score ranking in the second recall result is the same as the document score ranking in the first recall result 1, meaning the first recall result 1 is close to the optimal recall result.

[0099] Training data ES score VS score Integration score Doc1 0.6 0.9 0.66 Doc2 0.8 0.6 0.76 Doc3 0.4 0.5 0.42

[0100] Table 6

[0101] Assume the target weight is recombined as (0.5, 0.5). In Table 7, the first column represents the document index, the second column represents the first recall result, the third column represents the second recall result, and the fourth column represents the second recall result. In this case, the document score ranking in the second recall result is different from the document score ranking in either the first recall result 1 or the first recall result 2. That is, neither the first recall result 1 nor the first recall result 2 is the optimal recall result, and the second recall result obtained by the processing device is the optimal recall result.

[0102] Training data ES score VS score Integration score Doc1 0.6 0.9 0.75 Doc2 0.8 0.6 0.65 Doc3 0.4 0.5 0.55

[0103] Table 7

[0104] It should be understood that Tables 5 through 7 are examples provided in the embodiments of this application. In practical applications, the documents recalled by the two single-path recall models may be different. For example, in Table 7, the second single-path recall model may not recall Doc3. In this case, the VS score corresponding to Doc3 is 0.

[0105] Figure 5 This is a second flowchart illustrating the multi-channel recall method provided in an embodiment of this application. Figure 5As shown, the processing device is configured with M vectors 302, M weighted recombinations 303, and N single-path recall models. The N single-path recall models include a first single-path recall model 505 and a second single-path recall model 506. The processing device receives the term to be retrieved 501, vectorizes it to obtain a first vector 502. The processing device calculates the M similarities between the first vector 502 and the M vectors 302. The processing device determines the vector 503 corresponding to the maximum similarity among the M similarities, and the corresponding weighted recombination 504, i.e., the target weighted recombination. The processing device inputs the term to be retrieved 501 into the first single-path recall model 505 to obtain a first recall result 507. The processing device inputs the term to be retrieved 501 into the second single-path recall model 506 to obtain a first recall result 508. The processing device fuses the first recall result 507 and the first recall result 508 according to the weighted recombination 504 to obtain a second recall result 509.

[0106] In practical applications, the correspondence between M weighted reorganizations and M vectors, or the correspondence between M weighted reorganizations and M training words, can exist in an externally configurable and maintainable form to meet the needs of data verification and optimization. For example, the processing device is also used to receive the second weight set obtained by the processing device and replace the first weight set with the second weight set. As mentioned above... Figure 4 In the example, the processing device is used to receive M1 weighted recombinations and a correspondence 1 of M1 vectors. The processing device is also used to receive M2 weighted recombinations and a correspondence 2 of M2 vectors. The processing device replaces correspondence 1 with correspondence 2.

[0107] The training method and multi-path recall method provided in this application have been described above. The processing device provided in this application is described below.

[0108] Figure 6 This is a first structural schematic diagram of the processing device provided in an embodiment of this application. (See attached diagram.) Figure 6 As shown, the processing device 600 includes an input module 601 and a calculation module 602.

[0109] When the processing device 600 is a multi-channel recall device, the input module 601 is used to obtain the term to be retrieved. The calculation module 602 is used to obtain M similarities between the term to be retrieved and M training words. M is greater than 1. The M similarities correspond one-to-one with the M training words. The calculation module 602 is also used to obtain the target weight reassembly in the first weight set based on the training word corresponding to the maximum similarity among the M similarities. The first weight set includes M weight reassemblies. The M weight reassemblies correspond one-to-one with the M similarities. Each weight reassembly in the M weight reassemblies includes N weights. The N weights correspond one-to-one with the N single-channel recall models, and N is greater than 1. The calculation module 602 is also used to input the term to be retrieved into the N single-channel recall models to obtain N first recall results. The N first recall results correspond one-to-one with the N single-channel recall models. The calculation module 602 is also used to fuse the N first recall results based on the target weight reassembly to obtain a second recall result. A multi-channel recall device can also be called a processing device, for example... Figure 1 The processing device 102 in the middle.

[0110] It should be understood that the description of the processing device 600 is consistent with the foregoing. Figure 4 The descriptions of the multi-channel recall method in the previous text are similar. Therefore, the description of the processing device 600 can be referenced from the aforementioned text. Figure 4 The description of the multi-path recall method is as follows. For example, the input module 601 is also used to obtain a second weight set. The calculation module 602 is also used to replace the first weight set with the second weight set. Furthermore, if the M similarities include the same multiple maximum similarities, the calculation module 602 is also used to increase the precision of the multiple maximum similarities.

[0111] When the processing device 600 is a training device, the input module 601 is used to obtain a target weight set. The target weight set includes H weight reassemblies. Each weight reassembly in the H weight reassemblies includes N weights. The input module 601 is also used to obtain K training data, each training data corresponding to multiple training words. The K training data correspond to M training words, where K and M are greater than 1. The calculation module 602 is used to perform single-path recall on each of the M training words according to N single-path recall models, obtaining N recall results. The N recall results correspond one-to-one with the N single-path recall models. The calculation module 602 is also used to determine the weight reassemblies in the H weight reassemblies corresponding to each training word according to the N recall results, where the M training words correspond to the M weight reassemblies in the H weight reassemblies. The training device can also be called a training equipment, for example... Figure 1 Training equipment 101 in the middle.

[0112] It should be understood that the description of the processing device 600 is consistent with the foregoing. Figure 2 The descriptions of the training methods in the text are similar. Therefore, the description of the processing device 600 can be found in the foregoing. Figure 4The description of the multi-path recall method is as follows. For example, the value of H is between 1000 and 100000. By controlling the value of H, training efficiency can be improved while enhancing recall performance. For another example, the calculation module 602 is used to determine the weighted recombinations among the H weighted recombinations corresponding to each training word based on the N recall results. This includes: the calculation module 602 obtaining N sets of relevance scores corresponding to the N recall results, each set of relevance scores including K relevance scores, and the K relevance scores corresponding to K training data points; the calculation module 602 determining the weighted recombinations among the H weighted recombinations corresponding to each training word based on the K relevance scores and the training data corresponding to each training word.

[0113] Based on the methods in the foregoing embodiments, this application provides a processing device. Figure 7 This is a second structural schematic diagram of the processing device provided in an embodiment of this application. (See attached diagram.) Figure 7 As shown, the processing device 700 includes a memory 701 and a processor 702. The memory 701 is used to store programs; the processor 702 is used to execute the programs stored in the memory; wherein, when the programs stored in the memory are executed, the processor is used to execute the methods in the above embodiments.

[0114] Based on the methods in the foregoing embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods described in the foregoing embodiments.

[0115] Based on the methods in the foregoing embodiments, this application provides a computer program product, characterized in that, when the computer program product is run on a processor, the processor executes the methods in the foregoing embodiments.

[0116] Based on the methods described in the foregoing embodiments, this application also provides a chip. Please refer to... Figure 8 , Figure 8 This is a schematic diagram of a chip structure provided in an embodiment of this application. Figure 8As shown, chip 800 includes one or more processors 801 and interface circuits 802. Optionally, chip 800 may also include a bus 803. The processor 801 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 801 or by instructions in software form. The processor 801 may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods and steps disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor.

[0117] The interface circuit 802 can be used to send or receive data, instructions, or information. The processor 801 can process the data, instructions, or other information received by the interface circuit 802 and send the processed information out through the interface circuit 802. Optionally, the chip 800 also includes a memory, which may include read-only memory and random access memory, and provides operation instructions and data to the processor. A portion of the memory may also include non-volatile random access memory (NVRAM). Optionally, the memory stores executable software modules or data structures, and the processor can execute corresponding operations by calling the operation instructions stored in the memory (which may be stored in the operating system).

[0118] Optionally, the interface circuit 802 can be used to output the execution results of the processor 801. It should be noted that the functions of the processor 801 and the interface circuit 802 can be implemented through hardware design, software design, or a combination of hardware and software; no restrictions are imposed here.

[0119] It should be understood that each step in the above method embodiments can be implemented by hardware logic circuits or software instructions in a processor. It is also understood that the sequence number of each step in the above embodiments does not imply the order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. Furthermore, in some possible implementations, each step in the above embodiments may be selectively executed according to actual circumstances; it may be partially or fully executed, and no limitation is made here.

[0120] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0121] The method steps in the embodiments of this application can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0122] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. 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 can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, 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 can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0123] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A multi-path recall method, characterized in that, include: Get the search term; Obtain M similarities between the term to be retrieved and M training terms, where M is greater than 1, and the M similarities correspond one-to-one with the M training terms; The target weight recombination is obtained from the training word corresponding to the maximum similarity among the M similarities. The first weight set includes M weight recombinations, and each of the M weight recombinations corresponds one-to-one with the M training words. Each weight recombination in the M weight recombinations includes N weights, and each of the N weights corresponds one-to-one with N single-path recall models, where N is greater than 1. The term to be retrieved is input into the N single-path recall models to obtain N first recall results, and the N first recall results correspond one-to-one with the N single-path recall models; The N first recall results are merged according to the target weighting to obtain the second recall result.

2. The multi-channel recall method according to claim 1, characterized in that, The method further includes: Obtain the second weight set; Replace the first weight set with the second weight set.

3. The multi-channel recall method according to claim 1 or 2, characterized in that, The method further includes: If the M similarities include the same multiple maximum similarities, then the accuracy of the multiple maximum similarities is increased.

4. The multi-channel recall method according to any one of claims 1 to 3, characterized in that, The accuracy of the M similarities is less than 1 / M.

5. The multi-channel recall method according to any one of claims 1 to 4, characterized in that, The step of obtaining the M similarities between the term to be retrieved and the M training terms includes: The term to be retrieved is vectorized to obtain a first vector; Obtain the M similarities between the first vector and the M vectors, wherein the M vectors correspond one-to-one with the M training words.

6. A training method, characterized in that, include: Obtain a target weight set, which includes H weight reorganizations, and each of the H weight reorganizations includes N weights; Obtain K training data points, each training data point corresponding to multiple training words, and the K training data points corresponding to M training words, where K and M are greater than 1; Each of the M training words is recalled using N single-path recall models to obtain N recall results, and the N recall results correspond one-to-one with the N single-path recall models. Based on the N recall results, determine the weighted recombination in the H weighted recombinations corresponding to each training word, and the M training words correspond to the M weighted recombinations in the H weighted recombinations.

7. The training method according to claim 6, characterized in that, The step of determining the weighted recombination among the H weighted recombinations corresponding to each training word based on the N recall results includes: Obtain N sets of relevance scores corresponding to the N recall results, each set of relevance scores including K relevance scores, and the K relevance scores correspond one-to-one with the K training data; The weighted recombinations in the H weighted recombinations corresponding to each training word are determined based on the K relevance scores and the training data corresponding to each training word.

8. The training method according to claim 7, characterized in that, The K training data include first training data and second training data, each training word corresponds to the first training data, the N relevance score sets include a first relevance score set and a second relevance score set, the N single-path recall models include a first single-path recall model and a second single-path recall model, the first relevance score set is obtained based on the first single-path recall model, the second relevance score set is obtained based on the second single-path recall model, and the K relevance scores include a first relevance score and a second relevance score, the first relevance score corresponds to the first training data, and the second relevance score corresponds to the second training data; The step of determining the weighted group among the H weighted groupings corresponding to each training word based on the K relevance scores and the training data corresponding to each training word includes: If, in the first set of relevance scores, the first relevance score is greater than the second relevance score, and in the second set of relevance scores, the second relevance score is greater than the first relevance score, then a target weight reorganization is selected from the H weight reorganizations as the weight reorganization corresponding to each training word, and the weight corresponding to the first single-path recall model in the target weight reorganization is greater than the weight corresponding to the second single-path recall model.

9. A multi-channel recall device, characterized in that, It includes an input module and a calculation module, wherein: The input module is used to obtain the term to be searched; The calculation module is used to obtain M similarities between the word to be retrieved and M training words, where M is greater than 1, and the M similarities correspond one-to-one with the M training words; The calculation module is also used to obtain the target weight recombination in the first weight set according to the training word corresponding to the maximum similarity among the M similarities. The first weight set includes M weight recombinations, and the M weight recombinations correspond one-to-one with the M similarities. Each weight recombination in the M weight recombinations includes N weights, and the N weights correspond one-to-one with N single-path recall models, where N is greater than 1. The calculation module is also used to input the term to be retrieved into the N single-path recall models to obtain N first recall results, and the N first recall results correspond one-to-one with the N single-path recall models; The calculation module is also used to fuse the N first recall results according to the target weighting to obtain a second recall result.

10. The multi-channel recall device according to claim 9, characterized in that, The input module is also used to obtain a second weight set; The calculation module is also used to replace the first weight set with the second weight set.

11. The multi-channel recall device according to claim 9 or 10, characterized in that, If the M similarities include multiple identical maximum similarities, the calculation module is further used to increase the accuracy of the multiple maximum similarities.

12. The multi-channel recall device according to any one of claims 9 to 11, characterized in that, The calculation module is used to obtain M similarities between the term to be retrieved and M training terms, including: The calculation module is used to vectorize the term to be retrieved to obtain a first vector; The calculation module is used to obtain the M similarities between the first vector and the M vectors, wherein the M vectors correspond one-to-one with the M training words.

13. A training device, characterized in that, It includes an input module and a calculation module, wherein: The input module is used to obtain a target weight set, which includes H weight reconfigurations, and each of the H weight reconfigurations includes N weights. The input module is also used to acquire K training data, each training data corresponding to multiple training words, and the K training data corresponding to M training words, where K and M are greater than 1; The calculation module is used to perform single-path recall on each of the M training words according to the N single-path recall models to obtain N recall results, and the N recall results correspond one-to-one with the N single-path recall models; The calculation module is further configured to determine the weighted recombination in the H weighted recombinations corresponding to each training word based on the N recall results, and the M training words correspond to the M weighted recombinations in the H weighted recombinations.

14. The training device according to claim 13, characterized in that, The calculation module is used to determine, based on the N recall results, the weighted recombinations among the H weighted recombinations corresponding to each training word, including: The calculation module is used to obtain N sets of relevance scores corresponding to the N recall results. Each set of relevance scores includes K relevance scores, and the K relevance scores correspond one-to-one with the K training data. The calculation module is used to determine the weighted combination among the H weighted combination corresponding to each training word based on the K relevance scores and the training data corresponding to each training word.

15. A processing apparatus, characterized in that, Includes memory and processor, wherein: The memory is used to store programs; The processor is used to execute the program stored in the memory; When the program stored in the memory is executed, the processor is used to execute the multiplex recall method as described in any one of claims 1 to 5 or the training method of the multiplex recall model as described in claims 6 to 8.

Images