Parent slice selection method, electronic devices and program products

Abstract

This invention provides a parent slice selection method, electronic device, and program product. The method includes: obtaining one or more extended questions based on the user's original question; performing parallel searches on the original question and extended questions in multiple preset knowledge bases, and performing mixed keyword and semantic parallel recall for each question to obtain all recalled sub-slices; calling a predetermined re-ranking model to re-rank all sub-slices to obtain the re-ranking ranking and re-ranking score of the sub-slices; the re-ranking weighted score considers the knowledge base ranking weight, question type weight, and recall type weight; traversing all sub-slices to obtain the ID of the corresponding parent slice, calculating the comprehensive score of each parent slice based on multiple dimensions, and ranking each parent slice according to the comprehensive score; and constructing a context based on the recalled parent slices. Using the above technical solution, the recall source and comprehensive relevance are considered, avoiding the one-sidedness of the scoring and the redundancy of the context.

Description

Technical Field

[0001] This invention relates to retrieval enhancement generation technology in the field of artificial intelligence, and in particular to a parent slice selection method, electronic device and program product. Background Technology

[0002] Currently, in Retrieval-augmented Generation (RAG) systems, a "parent-child indexing" strategy is typically used to balance retrieval accuracy and the completeness of the generated context. This involves dividing long documents into larger parent slices, and then further dividing these parent slices into smaller child slices; the parent slices are used for generation, while the child slices are used for retrieval. Retrieval often employs a hybrid retrieval strategy, using keyword recall (such as the BM25 algorithm based on Elasticsearch) and semantic recall (such as similarity search based on vector databases) to improve recall and accuracy. However, post-retrieval reranking has the following drawbacks: 1. One-sided scoring: The highest score of the sub-slice is usually used directly as the score of the parent slice, ignoring the cumulative effect of a parent slice containing multiple polymer slices; 2. Ignoring the source of recall: In hybrid retrieval, the same sub-slice may be recalled by both keyword and semantic methods at the same time. This "double hit" signal is usually ignored and cannot reflect the comprehensive relevance of the content. 3. Context redundancy: If multiple child slices belonging to the same parent slice are retrieved, returning them directly will cause the content of the parent slice to be calculated or displayed repeatedly. Summary of the Invention

[0003] To address the aforementioned problems in the prior art, embodiments of the present invention provide a parent slice selection method, an electronic device, and a program product.

[0004] To achieve the above objectives, on the one hand, a method for optimizing parent slices is provided, including: S1, Based on the user's original question, obtain one or more extended questions related to the original question; S2, perform parallel searches on the original question and the extended question in multiple preset knowledge bases respectively, and perform mixed keyword and semantic parallel recall for each question to obtain all recalled sub-slices; wherein, the information contained in each sub-slice includes: the ID of the parent slice, the knowledge base to which it belongs, the question type associated with the recall, and the recall type; the question type includes: the original question and the extended question; the recall type includes: keyword recall and semantic recall; S3, call the predefined re-ranking model to re-rank all sub-slices, and obtain the re-ranking ranking and re-ranking score of the sub-slices. The information contained in each sub-slice after re-ranking includes: the ID of the parent slice, the knowledge base to which it belongs, the question type associated with the recall, the recall type, the re-ranking ranking, the re-ranking score, and the re-ranking weighted score; where, the re-ranking weighted score = re-ranking score × preset knowledge base ranking weight × preset question type weight × preset recall type weight; S4, iterate through all child slices, get the ID of the corresponding parent slice, calculate the comprehensive score of each parent slice, and sort the parent slices according to the comprehensive score, with the one with the higher score ranked first. Wherein, the overall score of the parent slice = (wf×Sf)+(wc×Sc)+(wp×Sp)+Sd; wf represents the preset frequency weight, wc represents the preset average score weight of the sub-slice, wp represents the preset position weight; Sf represents the frequency score, Sc represents the average score of the sub-slice, Sp represents the position score, and Sd represents the additional score given when the parent slice contains multiple different sub-slices, the additional score being determined according to preset rules. in: Sf = Total number of times the child slice was recalled / Number of parent slices; Sc = Sum of reordered weighted scores of all recalled child slices of the parent slice / Number of child slices recalled by the parent slice; Sp = 1 / the highest reorder position of the child slice among all recalled child slices of the parent slice; S5, construct a context based on the recalled parent slice, the context being used by the large model to answer the user's original question.

[0005] Preferably, in the parent slice selection method, Sd = min(0.05, number of child slices recalled by the parent slice × 0.03).

[0006] Preferably, in the parent slice selection method, wherein, The weight of the problem type corresponding to the original problem is greater than the weight of the problem type corresponding to the extended problem, and the weight of the problem type corresponding to the original problem plus the weight of the problem type corresponding to the extended problem equals 1. The sum of the ranking weights of each knowledge base is 1. The sum of the recall type weights corresponding to keyword recall and semantic recall is 1.

[0007] Preferably, the parent slice selection method extends the original problem through a large model to obtain one or more extended problems.

[0008] Preferably, the parent slice selection method further adds a time decay factor weight when calculating the reordering weighted score, wherein: Re-ranking weighted score = Re-ranking score × Preset knowledge base ranking weight × Preset question type weight × Preset recall type weight × Preset time decay factor weight.

[0009] Preferably, the parent slice selection method further includes: dynamically adjusting the recall type weight based on query features; wherein, for queries containing proper nouns or product models, the recall type weight corresponding to keyword recall is increased; for natural language description questions, the recall type weight corresponding to semantic recall is increased.

[0010] Preferably, the parent slice selection method further includes: configuring the quota of each knowledge base in the parent slice candidate set according to a preset target ratio.

[0011] Preferably, the parent slice selection method sets the quota for each knowledge base in the parent slice candidate set as follows: =max(1, round(N×) )); N is the total number of parent slices to be selected; The target quota for the i-th knowledge base; The target proportion for the i-th knowledge base; round(N× () indicates rounding, where =1.

[0012] On the other hand, an electronic device is also provided, including a memory and a processor, the memory storing at least one program, the at least one program being executed by the processor to implement the steps of the parent slice preference method as described in any of the above.

[0013] In another aspect, a computer program product is also provided, comprising a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the parent slice selection method as described above. The above technical solution has the following technical effects: The parent slice optimization method of this invention combines keyword and semantic parallel recall for each question and tracks the number of times each sub-slice is hit by different recall methods, integrating keyword and vector retrieval results and considering the recall source. Moreover, after the retrieval and recall stage, instead of simply deduplicating and re-ranking by similarity, it performs statistical analysis on all recalled sub-slices belonging to the same parent slice, and calculates the aggregate score of the parent slice based on multiple dimensions including frequency, average sub-slice score, position score, and diversity. This reflects the comprehensive relevance, coverage, and density of the content, and avoids the one-sidedness of the scoring and contextual redundancy. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating a method for selecting the optimal parent slice according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the parallel recall process in one embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0015] To further illustrate the various embodiments, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention, primarily used to illustrate the embodiments and to explain the operating principles of the embodiments in conjunction with the relevant descriptions in the specification. With reference to these drawings, those skilled in the art should be able to understand other possible implementations and the advantages of the present invention. Components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0016] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0017] Example 1: Figure 1 This is a flowchart illustrating a method for selecting the optimal parent slice according to an embodiment of the present invention. Figure 1 The preferred method for the parent slice in this embodiment includes: S1, Based on the user's original question, obtain one or more extended questions related to the original question; in one specific implementation, the original question is expanded through a large model to obtain one or more extended questions; this step can increase the search scope of the questions; In one specific implementation, the prompt words used are as follows: Generate 3 questions that are semantically identical to the following question but are expressed differently: [Original Question] The output format is: {"extended_questions":["extended question 1","extended question 2","extended question 3"]} For example, the original question is "How to back up data?". After the question expansion module, three extended questions can be obtained: ["Data backup method", "Steps to back up data", "How to archive data"].

[0018] S2 performs parallel searches on the original question and extended question in multiple pre-defined knowledge bases, and performs a mixed keyword and semantic parallel recall for each question to obtain all recalled sub-slices. Each sub-slice contains the following information: the ID of the parent slice, the knowledge base it belongs to, the question type associated with the recall, and the recall type. Question types include: original question and extended question; recall types include: keyword recall and semantic recall. The question type associated with the recall indicates which question was used for recall; the recall type indicates which type of recall was used. In this step, parallel recall using keywords and semantics is employed, specifically keyword recall using the ES database and semantic recall using the vector database. Compared to serial retrieval recall, this greatly improves the efficiency of retrieval recall. Finally, the recall results of each recall type are merged to obtain all the recalled sub-slices. Figure 2 A schematic diagram of the parallel recall process is shown.

[0019] For example: There are three knowledge bases: user manual knowledge base, QA knowledge base, and case knowledge base. For the exemplary problem in step S1 above, 6 sub-slices were recalled, of which sub-slice 1 and sub-slice 3 were recalled twice; a total of 4 parent slices were involved. The six sub-slices are: Sub-slice 1: Parent slice A, User Manual Knowledge Base, Original Question, Semantic Recall; Sub-slice 1: Parent slice A, User Manual Knowledge Base, Original Question, Keyword Recall; Sub-slice 2: Parent slice B, User Manual Knowledge Base, Extended Questions, Keyword Recall; Sub-slice 3: Parent slice C, QA knowledge base, original question, semantic recall; Sub-slice 3: Parent slice C, QA knowledge base, extended questions, keyword recall; Sub-slice 4: Parent slice D, case knowledge base, extended questions, semantic recall; S3, invoke the predefined re-ranking model to re-rank all sub-slices, and obtain the re-ranking ranking and re-ranking score of the sub-slices. The information contained in each sub-slice after re-ranking includes: the ID of the parent slice, the knowledge base to which it belongs, the question type associated with the recall, the recall type, the re-ranking ranking, the re-ranking score, and the re-ranking weighted score; where, the re-ranking weighted score = re-ranking score × preset knowledge base ranking weight × preset question type weight × preset recall type weight; in a specific implementation, the re-ranking model is the rerank model.

[0020] For example: Assume the user manual knowledge base ranking weight is 0.5, the QA knowledge base ranking weight is 0.3, the case study knowledge base ranking weight is 0.2, the weight of the original question is 0.7, the weight of the extended question is 0.3, the weight of semantic recall is 0.6, and the weight of keyword recall is 0.4. Wherein, the user manual knowledge base ranking weight + QA knowledge base ranking weight + case study knowledge base ranking weight = 1; the weight of the original question + the weight of the extended question = 1; the weight of semantic recall + the weight of keyword recall = 1; and the weight of the original question is greater than the weight of the extended question. Based on the above, the sub-slice information obtained is as follows: Sub-slice 1: Parent slice A, User manual knowledge base, original question, semantic recall, 5, 0.61, 0.1281 (via 0.61) 0.5 0.7 (0.6 obtained); Sub-slice 1: Parent slice A, User manual knowledge base, original question, keyword recall, 6, 0.61, 0.0854 (via 0.61) 0.5 0.7 (0.4 obtained); Sub-slice 2: Parent slice B, User Manual Knowledge Base, Extended Questions, Keyword Recall, 1, 0.91, 0.0546 (via 0.91) 0.5 0.3 (0.4 obtained); Sub-slice 3: Parent slice C, QA knowledge base, original question, semantic recall, 2, 0.81, 0.10206 (via 0.81) 0.3 0.7 (0.6 obtained); Sub-slice 3: Parent slice C, QA knowledge base, extended questions, keyword recall, 3, 0.81, 0.02916 (via 0.81) 0.3 0.3 (0.4 obtained); Sub-slice 4: Parent slice D, Case knowledge base, Extended questions, Semantic recall, 4, 0.71, 0.02556 (via 0.71) 0.2 0.3 0.6 (obtained).

[0021] S4, iterate through all child slices, get the ID of the corresponding parent slice, calculate the comprehensive score of each parent slice, and sort the parent slices according to the comprehensive score, with the one with the higher score ranked first. Wherein, the overall score of the parent slice = (wf×Sf)+(wc×Sc)+(wp×Sp)+Sd; wf represents the preset frequency weight, wc represents the preset average score weight of the sub-slice, and wp represents the preset position weight; Sf represents the frequency score, Sc represents the average score of the sub-slice, Sp represents the position score, and Sd represents the additional score given when the parent slice contains multiple different sub-slices, the additional score being determined according to preset rules; in one specific implementation, Sd = min(0.05, number of sub-slices recalled by the parent slice × 0.03); Sd is used for diversity rewards. If the parent slice contains multiple different sub-slices, it means that multiple parts of the paragraph corresponding to the parent slice are related to the problem, so an additional score is given to the parent slice, that is, the cumulative effect in this case is considered; preferably, the above wf, wc, wp, and Sd are configurable to achieve flexible adjustment of scoring measurement. in: Sf = Total number of times child slices were recalled / Number of parent slices; Sf reflects the breadth of content coverage; Sc = Sum of reordering weighted scores of all recalled child slices of the parent slice / Number of child slices recalled by the parent slice; Sc reflects the average quality of the content; Sp = 1 / the highest reorder position of the child slice among all recalled child slices of the parent slice; Sp reflects the original preference for reordering. This step obtains statistical information about the parent slice from all recalled sub-slices to analyze the focus parent slice; in one specific implementation, the statistical information of the parent slice includes: parent_id { child_appearances: The total number of times a subslice is recalled (this variable spans all question types and recall types, and includes the number of unique subslices). child_count: The number of child slices recalled (this variable represents the number of duplicate child slices). total_child_score: The sum of the reordered weighted scores of all recalled child slices. rerank_min_position: The first position of the sub-slice after reordering. }

[0022] parent_id represents the ID of the parent slice.

[0023] For example: In the example above, the parent slice information is as follows: Parent slice A{ child_appearances: 2 child_count: 1 total_child_score: 0.1281+0.0854=0.2135 rerank_min_position: 5 } Parent slice B{ child_appearances: 1 child_count: 1 total_child_score: 0.0546 rerank_min_position: 1 } Parent slice C{ child_appearances: 2 child_count: 2 total_child_score: 0.10206+0.02916=0.13122 rerank_min_position: 2 } Parent slice D{ child_appearances: 1 child_count: 1 total_child_score: 0.02556 rerank_min_position: 4 }

[0024] For example, setting wf=0.5, wc=0.3, wp=0.2, the comprehensive score Sfinal for each parent slice is calculated based on the parent slice information collected in the previous step: Parent slice A: Sf=2 / 4=0.5; Sc=0.2135 / 1=0.2135; Sp=1 / 5=0.2; Sd=min(0.05,1 0.03) = 0.03; Sfinal=0.5 0.5 + 0.3 0.2135+0.2 0.2 + 0.03 = 0.38405; Parent slice B: Sf=1 / 4=0.25; Sc=0.0546 / 1=0.0546; Sp=1 / 1=1; Sd=min(0.05,1 0.03) = 0.03; Sfinal=0.5 0.25 + 0.3 0.0546 + 0.2 1 + 0.03 = 0.37138; Parent slice C: Sf=1 / 4=0.25; Sc=0.13122 / 2=0.06561; Sp=1 / 2=0.5; Sd=min(0.05,2 0.03) = 0.05; Sfinal=0.5 0.25 + 0.3 0.06561 + 0.2 0.5 + 0.05 = 0.294683 Parent slice D: Sf=2 / 4=0.5; Sc=0.02556 / 1=0.02556; Sp=1 / 4=0.25; Sd=min(0.05,1 0.03) = 0.03; Sfinal=0.5 0.5 + 0.3 0.02556 + 0.2 0.25 + 0.03 = 0.337668.

[0025] Based on the comprehensive scores of the parent slices, the parent slices are sorted, with the higher score ranked first and the lower score ranked last. In this example, the order is as follows: parent slice A (0.38405), parent slice B (0.37138), parent slice D (0.337668), and parent slice C (0.294683).

[0026] S5: Construct a context based on the recalled parent slices. This context is used by the large model to answer the user's original question.

[0027] Compared with the prior art, the advantages of the embodiments of the present invention are as follows: Hybrid retrieval signal fusion: Existing solutions typically process keyword and vector retrieval results separately. The solution in this embodiment tracks the number of times each sub-slice is hit by different recall methods. When a sub-slice is hit by both keyword and semantic recall (i.e., double hit), it indicates that its relevance to the query is higher and it will receive additional weight in the scoring. Aggregated scoring: Existing solutions often directly take the maximum value. The solution in this embodiment introduces the dimensions of Frequency and Diversity. For example, if a parent slice has three child slices with a score of 0.8, and another has only one child slice with a score of 0.85, existing solutions would choose the latter, while this solution would choose the former because the former contains more information. 3. Sub-slice weight accumulation mechanism: The solution of this embodiment of the invention accumulates the scores of all related sub-slices by weighting (knowledge base ranking weight, question type weight, recall type weight) rather than simply taking the maximum value or average value, so as to more accurately reflect the comprehensive relevance of the parent slice; Multi-dimensional scoring formula: It comprehensively considers four dimensions: frequency, score, location, and diversity, and allows for flexible adjustment of the scoring strategy through configurable weighting coefficients. Example 2: Building upon Example 1, the re-ranking weighted score in step S3 incorporates a time decay factor. For time-sensitive knowledge bases, this allows newer documents to receive a bonus in the scoring. In this case: Re-ranking weighted score = Re-ranking score × Preset knowledge base ranking weight × Preset question type weight × Preset recall type weight × Preset time decay factor weight.

[0028] Example 3: Based on any of the embodiments described above, the recall type weight in step S3 is dynamically adjusted according to the query features. Specifically, for queries containing proper nouns or product models, the recall type weight corresponding to keyword recall is increased; for natural language descriptions, the recall type weight corresponding to semantic recall is increased.

[0029] Example 4: Based on any of the embodiments described above, a knowledge base sharding balancing strategy is added in step S4. In multi-knowledge base scenarios, to avoid the final context being entirely concentrated in a single knowledge base, one specific implementation configures the proportion or quota of each knowledge base in the parent slice candidate set, thereby ensuring that each knowledge base can contribute to the recalled data. Specifically, let the total number of parent slices to be selected be N, and the target proportion of the i-th knowledge base be... , =1, then the target quota for the i-th knowledge base can be set as : =max(1, round(N×) )); Where round(N× () indicates rounding. This embodiment can first sort the scores by parent slice aggregation within each knowledge base and then select the top... There are several parent slices; if there are not enough candidate parent slices for a certain knowledge base, its unused quotas are redistributed to other knowledge bases according to the score from high to low, so as to ensure that the context window is fully utilized and that each knowledge base is covered.

[0030] Example 5: The present invention also provides an electronic device, such as... Figure 3 As shown, the device includes a processor 301, a memory 302, a bus 303, and a computer program stored in the memory 302 and executable on the processor 301. The processor 301 includes one or more processing cores. The memory 302 is connected to the processor 301 via the bus 303. The memory 302 is used to store program instructions. When the processor executes the computer program, it implements the steps in the above-described method embodiment of Embodiment 1 of the present invention.

[0031] Furthermore, as an executable solution, the electronic device can be a computer unit, which can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. The computer unit may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above-described structure of the computer unit is merely an example and does not constitute a limitation on the computer unit. It may include more or fewer components, or combine certain components, or use different components. For example, the computer unit may also include input / output devices, network access devices, buses, etc., and this embodiment of the invention does not limit this.

[0032] Furthermore, as an executable solution, the processor 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, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The processor is the control center of the computer unit, connecting various parts of the entire computer unit via various interfaces and lines.

[0033] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the computer unit by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital card (SD card), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0034] Example 6: The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps described above.

[0035] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A method for optimizing parent slices, characterized in that, include: S1, Based on the user's original question, obtain one or more extended questions related to the original question; S2, perform parallel searches on the original question and the extended question in multiple preset knowledge bases respectively, and perform mixed keyword and semantic parallel recall for each question to obtain all recalled sub-slices; wherein, the information contained in each sub-slice includes: the ID of the parent slice, the knowledge base to which it belongs, the question type associated with the recall, and the recall type; the question type includes: the original question and the extended question; the recall type includes: keyword recall and semantic recall; S3, call the predefined re-ranking model to re-rank all sub-slices, and obtain the re-ranking ranking and re-ranking score of the sub-slices. The information contained in each sub-slice after re-ranking includes: the ID of the parent slice, the knowledge base to which it belongs, the question type associated with the recall, the recall type, the re-ranking ranking, the re-ranking score, and the re-ranking weighted score; where, the re-ranking weighted score = re-ranking score × preset knowledge base ranking weight × preset question type weight × preset recall type weight; S4, iterate through all child slices, get the ID of the corresponding parent slice, calculate the comprehensive score of each parent slice, and sort the parent slices according to the comprehensive score, with the one with the higher score ranked first. Wherein, the overall score of the parent slice = (wf×Sf)+(wc×Sc)+(wp×Sp)+Sd; wf represents the preset frequency weight, wc represents the preset average score weight of the sub-slice, wp represents the preset position weight; Sf represents the frequency score, Sc represents the average score of the sub-slice, Sp represents the position score, and Sd represents the additional score given when the parent slice contains multiple different sub-slices, the additional score being determined according to preset rules. in: Sf = Total number of times the child slice was recalled / Number of parent slices; Sc = Sum of reordered weighted scores of all recalled child slices of the parent slice / Number of child slices recalled by the parent slice; Sp = 1 / the highest reorder position of the child slice among all recalled child slices of the parent slice; S5, construct a context based on the recalled parent slice, the context being used by the large model to answer the user's original question.

2. The method for selecting the optimal parent slice according to claim 1, characterized in that, Sd=min(0.05, number of child slices recalled by the parent slice × 0.03).

3. The method for selecting the optimal parent slice according to claim 1, characterized in that, The weight of the problem type corresponding to the original problem is greater than the weight of the problem type corresponding to the extended problem, and the weight of the problem type corresponding to the original problem plus the weight of the problem type corresponding to the extended problem equals 1. The sum of the ranking weights of each knowledge base is 1. The sum of the recall type weights corresponding to keyword recall and semantic recall is 1.

4. The method for selecting the optimal parent slice according to claim 1, characterized in that, The original problem is extended by a large model, resulting in one or more extended problems.

5. The method for selecting the preferred parent slice according to claim 1, characterized in that, A time decay factor is also added when calculating the re-ranked weighted score, where: Re-ranking weighted score = Re-ranking score × Preset knowledge base ranking weight × Preset question type weight × Preset recall type weight × Preset time decay factor weight.

6. The method for selecting the optimal parent slice according to claim 1, characterized in that, Also includes: The recall type weights are dynamically adjusted based on query characteristics; for queries containing proper nouns or product models, the recall type weights corresponding to keyword recalls are increased. For problems involving natural language description, increase the weight of the recall type corresponding to semantic recall.

7. The method for selecting the optimal parent slice according to claim 1, characterized in that, Also includes: Configure the quota for each knowledge base in the parent slice candidate set according to the preset target ratio.

8. The method for selecting the optimal parent slice according to claim 7, characterized in that, Set the quota for each knowledge base in the parent slice candidate set to: =max(1,round(N× )); N is the total number of parent slices to be selected; The target quota for the i-th knowledge base; The target proportion for the i-th knowledge base; round(N×) () indicates rounding, where =1.

9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing at least one program, the at least one program being executed by the processor to implement the steps of the parent slice preference method as described in any one of claims 1 to 8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the parent slice preference method as described in any one of claims 1 to 8.

Images