Data processing method and apparatus, electronic device, and computer program product

Abstract

The present disclosure relates to the technical field of artificial intelligence, and disclosed are a data processing method and apparatus, an electronic device, and a computer program product. The method comprises: on the basis of target question information, executing a multi-channel knowledge recall operation in a target knowledge base to obtain a plurality of pieces of initial reply information, wherein the multi-channel knowledge recall operation at least comprises: a sparse vector-based knowledge recall operation and a dense vector-based knowledge recall operation; if the plurality of pieces of initial reply information satisfy a preset condition, performing sorting processing on the plurality of pieces of initial reply information to obtain a plurality of pieces of processed initial reply information; and on the basis of the plurality of pieces of processed initial reply information, determining target reply information corresponding to the target question information.

Description

Data processing methods and apparatus, electronic equipment and computer program products

[0001] Cross-reference

[0002] This disclosure claims priority to Chinese Patent Application No. 2024117070556, filed on November 26, 2024, entitled "Data Processing Method and Apparatus, Electronic Device and Computer Program Product", the entire contents of which are incorporated herein by reference. Technical Field

[0003] This disclosure relates to the field of artificial intelligence technology, and more specifically, to a data processing method and apparatus, electronic equipment, and computer program product. Background Technology

[0004] The goal of intelligent question answering systems is to solve various problems and confusions for users. Traditional intelligent question answering systems generally use search engine technology to match information in the knowledge base based on keyword-based inverted index retrieval. Then, the retrieved relevant information is input into a generative big data model, which then makes a response based on the retrieved information. However, keyword-based inverted index retrieval has the problem of inaccurate retrieval when the semantics change.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This disclosure provides a data processing method and apparatus, electronic device, and computer program product to at least solve the technical problem that retrieving answer information for related questions by keyword matching results in relatively low accuracy of the answer information.

[0007] According to one aspect of the present disclosure, a data processing method is provided, comprising: performing a multi-path knowledge retrieval operation in a target knowledge base based on target question information to obtain multiple initial response information, wherein the multi-path knowledge retrieval operation includes at least: a sparse vector-based knowledge retrieval operation and a dense vector-based knowledge retrieval operation; sorting the multiple initial response information when the multiple initial response information meets preset conditions to obtain processed multiple initial response information; and determining the target response information corresponding to the target question information based on the processed multiple initial response information.

[0008] Furthermore, before performing a multi-way knowledge retrieval operation in the target knowledge base based on the target question information to obtain multiple initial response information, the method further includes: acquiring multiple knowledge information; if the format information corresponding to the target knowledge information is in text format, then performing multi-level block processing on the target knowledge information to obtain processed target knowledge information, wherein the target knowledge information is any one of the multiple knowledge information; and obtaining the target knowledge base based on the processed target knowledge information and the knowledge information other than the text format knowledge information among the multiple knowledge information.

[0009] Furthermore, after performing multi-level block processing on the target knowledge information to obtain processed target knowledge information, the method further includes: if the block granularity corresponding to the processed target knowledge information is document granularity, then constructing index information based on the document title information corresponding to the processed target knowledge information, wherein the index information is used to match with the target question information to perform a knowledge retrieval operation; if the block granularity corresponding to the processed target knowledge information is paragraph granularity, then constructing the index information based on the paragraph title information corresponding to the processed target knowledge information; if the block granularity corresponding to the processed target knowledge information is block granularity, then constructing the index information based on the question information that the processed target knowledge information can answer.

[0010] Furthermore, when constructing index information based on the document title information corresponding to the processed target knowledge information, the method further includes: constructing first prompt information based on the document title information, wherein the first prompt information is used to instruct the generative big model to summarize the content corresponding to the processed target knowledge information; processing the first prompt information through the generative big model to obtain overview information corresponding to the processed target knowledge information; constructing key-value pairs based on the overview information corresponding to the processed target knowledge information and the document title information, and storing the key-value pairs in the target knowledge base, wherein, for document-level knowledge information, during knowledge retrieval operations, the overview information in the key-value pairs is used at least to constitute the plurality of initial response information.

[0011] Furthermore, if the granularity of the processed target knowledge information is block-level, then constructing the index information based on the question information that can be solved by the processed target knowledge information includes: constructing second prompt information based on the processed target knowledge information, wherein the second prompt information is used to instruct the generative big model to output the question information that can be solved by the processed target knowledge information; processing the second prompt information through the generative big model to obtain the question information; and constructing the index information based on the question information.

[0012] Furthermore, after acquiring multiple pieces of knowledge information, the method further includes: if the format information corresponding to the target knowledge information is in table format, then determining whether the size of the target knowledge information is greater than a target threshold; if the size of the target knowledge information is less than or equal to the target threshold, then converting the format of the target knowledge information to obtain unstructured text information; and constructing index information based on the unstructured text information, wherein the index information is used to match the target question information to perform a knowledge retrieval operation.

[0013] Further, based on the target question information, a multi-path knowledge retrieval operation is performed in the target knowledge base to obtain multiple initial response information, including: based on the sparse vector corresponding to the target question information, a sparse vector knowledge retrieval operation is performed in the target knowledge base to obtain a first retrieval result; based on the dense vector corresponding to the target question information, a dense vector knowledge retrieval operation is performed in the target knowledge base to obtain a second retrieval result; based on the keyword corresponding to the target question information, a keyword knowledge retrieval operation is performed in the target knowledge base to obtain a third retrieval result; based on the target question information, a knowledge retrieval operation is performed in the target knowledge base using the maximum boundary correlation algorithm to obtain a fourth retrieval result; and based on the first retrieval result, the second retrieval result, the third retrieval result, and the fourth retrieval result, the multiple initial response information is obtained.

[0014] Further, based on the sparse vector corresponding to the target question information, performing a sparse vector knowledge retrieval operation in the target knowledge base to obtain a first retrieval result includes: vectorizing the target question information to obtain a first sparse vector; obtaining a second sparse vector corresponding to the index information in the target knowledge base; and performing a sparse vector knowledge retrieval operation in the target knowledge base based on the first sparse vector and the second sparse vector to obtain the first retrieval result.

[0015] Further, based on the first sparse vector and the second sparse vector, a knowledge retrieval operation is performed on the sparse vector in the target knowledge base to obtain the first retrieval result, which includes: calculating the similarity between the first sparse vector and the second sparse vector, and determining the target index from the index information in the target knowledge base based on the similarity; performing a knowledge retrieval operation on the target knowledge base based on the target index to obtain the key-value pair corresponding to the target index and the knowledge information corresponding to the target index; and obtaining the first retrieval result based on the key-value pair corresponding to the target index and / or the knowledge information corresponding to the target index.

[0016] Furthermore, before sorting the multiple initial response information to obtain processed multiple initial response information, the method further includes: obtaining multiple first recall scores corresponding to the initial response information among the multiple initial response information, wherein the multiple first recall scores are obtained when performing a multi-way knowledge recall operation in the target knowledge base; determining a matching value corresponding to the initial response information among the multiple initial response information based on the multiple first recall scores, wherein the matching value is used to characterize the degree of matching between the multiple initial response information and the target question information; and determining whether the multiple initial response information meets the preset condition based on the matching value.

[0017] Further, determining the matching value corresponding to the initial response information among the multiple initial response information based on the multiple first recall scores includes: normalizing the multiple first recall scores to obtain a normalized first recall score; calculating the similarity between the target question information and the directory information corresponding to the initial response information among the multiple initial response information to obtain a second recall score; calculating the similarity between the target question information and the initial response information among the multiple initial response information to obtain a third recall score; calculating the similarity between the target question information and the title information corresponding to the initial response information among the multiple initial response information to obtain a fourth recall score; and calculating the matching value based on the normalized first recall score, the second recall score, the third recall score, and the fourth recall score.

[0018] Further, determining whether the plurality of initial response information satisfies the preset conditions includes: determining whether the matching value satisfies the preset requirements; if the matching value satisfies the preset requirements, then the plurality of initial response information satisfies the preset conditions.

[0019] Furthermore, after determining whether the matching value meets the preset requirements, the method further includes: if the matching value does not meet the preset requirements, generating multiple similar question information based on the target question information through the generative large model; performing a multi-way knowledge retrieval operation in the target knowledge base based on the multiple similar question information again to obtain multiple initial response information corresponding to the secondary knowledge retrieval, and obtaining the processed initial response information based on the multiple initial response information corresponding to the secondary knowledge retrieval.

[0020] According to another aspect of the embodiments of this disclosure, a data processing method is also provided, comprising: acquiring target question information uploaded by a client; performing a multi-path knowledge retrieval operation in a target knowledge base based on the target question information in a cloud server to obtain multiple initial response information, wherein the multi-path knowledge retrieval operation includes at least: a sparse vector-based knowledge retrieval operation and a dense vector-based knowledge retrieval operation; sorting the multiple initial response information when the multiple initial response information meets preset conditions to obtain processed multiple initial response information; determining the target response information corresponding to the target question information based on the processed multiple initial response information; and returning the target response information to the client.

[0021] According to another aspect of the embodiments of this disclosure, a data processing apparatus is also provided, comprising: a first execution component, configured to perform a multi-path knowledge retrieval operation in a target knowledge base based on target question information to obtain a plurality of initial response information, wherein the multi-path knowledge retrieval operation includes at least: a knowledge retrieval operation based on sparse vectors and a knowledge retrieval operation based on dense vectors; a first processing component, configured to sort the plurality of initial response information when the plurality of initial response information meets preset conditions to obtain a plurality of processed initial response information; and a first determining component, configured to determine the target response information corresponding to the target question information based on the plurality of processed initial response information.

[0022] Furthermore, the device further includes: a first acquisition component, configured to acquire multiple knowledge information before performing a multi-way knowledge retrieval operation in the target knowledge base based on the target question information to obtain multiple initial response information; a block-segmentation component, configured to perform multi-level block processing on the target knowledge information if the format information corresponding to the target knowledge information is in text format, to obtain processed target knowledge information, wherein the target knowledge information is any one of the multiple knowledge information; and a second determination component, configured to obtain the target knowledge base based on the processed target knowledge information and the knowledge information other than the text format knowledge information among the multiple knowledge information.

[0023] Furthermore, the device further includes: a first construction component, configured to, after performing multi-level block processing on the target knowledge information to obtain processed target knowledge information, if the block granularity corresponding to the processed target knowledge information is document granularity, then construct index information based on the document title information corresponding to the processed target knowledge information, wherein the index information is used to match with the target question information to perform a knowledge retrieval operation; a second construction component, configured to, if the block granularity corresponding to the processed target knowledge information is paragraph granularity, then construct the index information based on the paragraph title information corresponding to the processed target knowledge information; and a third construction component, configured to, if the block granularity corresponding to the processed target knowledge information is block granularity, then construct the index information based on the question information that the processed target knowledge information can answer.

[0024] Furthermore, the device further includes: a fourth construction component, configured to construct first prompt information based on the document title information when constructing index information based on the document title information corresponding to the processed target knowledge information, wherein the first prompt information is used to instruct the generative big model to summarize the content corresponding to the processed target knowledge information; a second processing component, configured to process the first prompt information through the generative big model to obtain overview information corresponding to the processed target knowledge information; and a fifth construction component, configured to construct key-value pairs based on the overview information corresponding to the processed target knowledge information and the document title information, and store the key-value pairs in the target knowledge base, wherein, for document-level knowledge information, during knowledge retrieval operations, the overview information in the key-value pairs is at least used to constitute the plurality of initial response information.

[0025] Furthermore, the third construction component includes: a first construction sub-component, configured to construct second prompt information based on the processed target knowledge information, wherein the second prompt information is used to instruct the generative big model to output question information that can be solved by the processed target knowledge information; a first processing sub-component, configured to process the second prompt information through the generative big model to obtain the question information; and a second construction sub-component, configured to construct the index information based on the question information.

[0026] Furthermore, the device further includes: a judgment component, configured to, after acquiring multiple pieces of knowledge information, determine whether the size of the target knowledge information is greater than a target threshold if the format information corresponding to the target knowledge information is in table format; a conversion component, configured to convert the target knowledge information to unstructured text information if the size of the target knowledge information is less than or equal to the target threshold; and a fifth construction component, configured to construct index information based on the unstructured text information, wherein the index information is used to match the target question information to perform a knowledge retrieval operation.

[0027] Further, the execution component includes: a first operation subcomponent, configured to perform a sparse vector knowledge recall operation in the target knowledge base based on the sparse vector corresponding to the target question information, to obtain a first recall result; a second operation subcomponent, configured to perform a dense vector knowledge recall operation in the target knowledge base based on the dense vector corresponding to the target question information, to obtain a second recall result; a third operation subcomponent, configured to perform a keyword knowledge recall operation in the target knowledge base based on the keyword corresponding to the target question information, to obtain a third recall result; a fourth operation subcomponent, configured to perform a knowledge recall operation in the target knowledge base based on the target question information using a maximum boundary correlation algorithm, to obtain a fourth recall result; and a first determination subcomponent, configured to obtain the plurality of initial response information based on the first recall result, the second recall result, the third recall result, and the fourth recall result.

[0028] Further, the first operation sub-component includes: a processing component, configured to perform vectorization processing on the target question information to obtain a first sparse vector; an acquisition component, configured to acquire a second sparse vector corresponding to the index information in the target knowledge base; and an execution component, configured to perform a sparse vector knowledge retrieval operation in the target knowledge base based on the first sparse vector and the second sparse vector to obtain the first retrieval result.

[0029] Further, the execution component includes: a calculation subcomponent, configured to calculate the similarity between the first sparse vector and the second sparse vector, and determine a target index from the index information in the target knowledge base based on the similarity; an operation subcomponent, configured to perform a knowledge retrieval operation in the target knowledge base based on the target index, to obtain the key-value pair corresponding to the target index and the knowledge information corresponding to the target index; and a determination subcomponent, configured to obtain the first retrieval result based on the key-value pair corresponding to the target index and / or the knowledge information corresponding to the target index.

[0030] Furthermore, the device further includes: a second acquisition component, configured to acquire multiple first recall scores corresponding to the initial response information among the multiple initial response information before sorting the multiple initial response information to obtain processed multiple initial response information, wherein the multiple first recall scores are obtained when performing a multi-way knowledge recall operation in the target knowledge base; a third determination component, configured to determine a matching value corresponding to the initial response information among the multiple initial response information based on the multiple first recall scores, wherein the matching value is used to characterize the degree of matching between the multiple initial response information and the target question information; and a third processing component, configured to determine whether the multiple initial response information meets the preset conditions based on the matching value.

[0031] Further, the third determining component includes: a second processing subcomponent, configured to normalize the plurality of first recall scores to obtain a normalized first recall score; a first calculation subcomponent, configured to perform similarity calculation based on the directory information corresponding to the target question information and the initial response information among the plurality of initial response information to obtain a second recall score; a second calculation subcomponent, configured to perform similarity calculation based on the target question information and the initial response information among the plurality of initial response information to obtain a third recall score; a third calculation subcomponent, configured to perform similarity calculation based on the title information corresponding to the target question information and the initial response information among the plurality of initial response information to obtain a fourth recall score; and a fourth calculation subcomponent, configured to calculate the matching value based on the normalized first recall score, the second recall score, the third recall score, and the fourth recall score.

[0032] Furthermore, the third processing component includes: a judgment sub-component, configured to judge whether the matching value meets a preset requirement; and a third processing sub-component, configured to, if the matching value meets the preset requirement, then the plurality of initial response information meets the preset condition.

[0033] Furthermore, the device further includes: a generation component, configured to, after determining whether the matching value meets the preset requirements, if the matching value does not meet the preset requirements, generate multiple similar question information based on the target question information through the generative large model; and a second execution component, configured to again execute the step of performing a multi-way knowledge retrieval operation in the target knowledge base based on the multiple similar question information, to obtain multiple initial response information corresponding to the secondary knowledge retrieval, and to obtain the processed initial response information based on the multiple initial response information corresponding to the secondary knowledge retrieval.

[0034] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the storage medium is located to perform the data processing method described in any one of the above embodiments.

[0035] According to another aspect of the present invention, an electronic device is also provided, including a memory storing an executable program; and a processor configured to run the program, wherein the program executes the data processing method described in any of the preceding embodiments.

[0036] According to another aspect of the present invention, a computer program product is also provided, the computer program product comprising a stored computer program that, when executed by a processor, implements the data processing method described in any one of the preceding embodiments.

[0037] In this embodiment, a multi-path knowledge retrieval operation is performed in the target knowledge base based on the target question information to obtain multiple initial response information. The multi-path knowledge retrieval operation includes at least two methods: a sparse vector-based knowledge retrieval operation and a dense vector-based knowledge retrieval operation. If the multiple initial response information meets preset conditions, the multiple initial response information is sorted to obtain processed initial response information. Based on the processed initial response information, the target response information corresponding to the target question information is determined. By performing a multi-path knowledge retrieval operation in the target knowledge base, the coverage of knowledge retrieval is improved, thereby enhancing the accuracy of answering questions. This solves the technical problem that retrieving response information for related questions using keyword matching results in relatively low accuracy of the response information. Attached Figure Description

[0038] The accompanying drawings, which are included to provide a further understanding of this disclosure and form part of this disclosure, illustrate exemplary embodiments of the present disclosure and are used to explain the disclosure, but do not constitute an undue limitation of the disclosure. In the drawings:

[0039] Figure 1 is a hardware structure block diagram of a computer terminal provided according to Embodiment 1 of this disclosure;

[0040] Figure 2 is a flowchart of a data processing method provided according to Embodiment 1 of this disclosure;

[0041] Figure 3 is a schematic diagram of the intelligent agent architecture provided according to Embodiment 1 of this disclosure;

[0042] Figure 4 is a block diagram provided according to Embodiment 1 of this disclosure;

[0043] Figure 5 is a flowchart of the data processing method provided according to Embodiment 2 of this disclosure;

[0044] Figure 6 is a schematic diagram of a data processing apparatus provided according to Embodiment 3 of the present disclosure;

[0045] Figure 7 is a structural block diagram of an electronic device provided according to Embodiment 4 of this disclosure. Detailed Implementation

[0046] To enable those skilled in the art to better understand the present disclosure, the technical solutions of the present disclosure will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present disclosure, and not all embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present disclosure.

[0047] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or components is not necessarily limited to those steps or components explicitly listed, but may include other steps or components not explicitly listed or inherent to such processes, methods, products, or devices.

[0048] First, some nouns or terms that appear in the description of the embodiments of this disclosure shall be interpreted as follows:

[0049] Generative large models are machine learning algorithms that can autonomously generate text. Unlike traditional supervised learning, generative large models do not require manual input of label information. Instead, they can automatically learn text generation patterns from large amounts of data, thereby generating new text with similar structure, syntax, and semantics.

[0050] Vector retrieval: The information to be retrieved is first vectorized into a multi-dimensional vector based on a vector model, and similar information is retrieved from the vector database based on vector similarity.

[0051] Dense vector: A vector in which most of the pointers are non-zero values.

[0052] Sparse vector: A vector in which most of the elements are zero or close to zero, and only a few elements are meaningful non-zero values.

[0053] Maximum Boundary Relevance Algorithm: This algorithm aims to rank documents by calculating the relevance between query text and documents, in order to find documents that are more relevant to the query from a large number of documents. The query usually refers to a request or question.

[0054] Token: A word element. Word elements are obtained from a text using word segmentation methods, which can be understood as similar to word segmentation.

[0055] Jaccard Distance: This is an indicator used to measure the difference between two sets. The smaller the Jaccard distance, the higher the similarity between the two sets.

[0056] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0057] According to embodiments of this disclosure, a data processing method is also provided. It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowcharts, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0058] The method embodiment provided in this disclosure can be executed in a mobile terminal, computer terminal, or similar computing device. Figure 1 shows a hardware structure block diagram of a computer terminal (or mobile device) for implementing a data processing method. As shown in Figure 1, the computer terminal (or mobile device) 10 may include a processor set 102 (the processor set 102 may include, but is not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA, and the processor set 102 may include a processor set, shown in Figure 1 as 102a, 102b, ..., 102n), a memory 104 for storing data, and a transmission component 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a Universal Serial Bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that the structure shown in Figure 1 is only illustrative and does not limit the structure of the above-described electronic device. For example, the computer terminal 10 may also include more or fewer components than shown in Figure 1, or have a different configuration than shown in Figure 1.

[0059] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuitry are generally referred to herein as "data processing circuitry". This data processing circuitry may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuitry may be a single, independent processing component, or may be integrated, in whole or in part, into any other element within the computer terminal 10 (or mobile device). As involved in embodiments of this disclosure, the data processing circuitry serves as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0060] The memory 104 can be used to store software programs and components of application software, such as program instructions / data storage devices corresponding to the data processing method in this embodiment of the present disclosure. The processor 102 executes various functional applications and data processing by running the software programs and components stored in the memory 104, thereby implementing the aforementioned data processing method. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0061] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a radio frequency (RF) component used for wireless communication with the Internet.

[0062] The display may be a touchscreen LCD display that allows the user to interact with the user interface of the computer terminal 10 (or mobile device).

[0063] Under the aforementioned operating environment, this disclosure provides a data processing method as shown in Figure 2. Figure 2 is a flowchart of the data processing method according to Embodiment 1 of this disclosure. The data processing method includes:

[0064] Step S201: Based on the target question information, perform a multi-way knowledge retrieval operation in the target knowledge base to obtain multiple initial response information. The multi-way knowledge retrieval operation includes at least: a knowledge retrieval operation based on sparse vectors and a knowledge retrieval operation based on dense vectors.

[0065] Optionally, the data processing method provided in Embodiment 1 of this disclosure can be applied to an Agent architecture (intelligent architecture, which generally refers to the way and structure of designing and building intelligent agents in the field of artificial intelligence; an intelligent agent is an intelligent entity that can perceive the environment, make decisions, and execute actions). It can receive target question information input by the user through the front-end interface of the Agent architecture, and then perform multi-way knowledge retrieval operations in the target knowledge base through the multi-way retrieval coarse-ranking module in the Agent architecture to obtain multiple initial response information. It should be noted that the multi-way knowledge retrieval operations include at least: sparse vector-based knowledge retrieval operations and dense vector-based knowledge retrieval operations.

[0066] For example, sparse vector-based knowledge retrieval operations involve converting target question information into sparse vectors. In a sparse vector, each dimension corresponds to a word in the vocabulary, and the vector values ​​typically reflect the importance of the word in the document. In an alternative embodiment, sparse vectors can be computed automatically using a text vectorization model, such as text-embedding-v3. Retrieval is then performed based on the sparse vectors using a retrieval algorithm (e.g., a nearest neighbor search algorithm).

[0067] It should be noted that knowledge recall operations based on dense vectors are similar to those based on sparse vectors, and will not be elaborated further here.

[0068] The accuracy of knowledge retrieval can be effectively improved by using knowledge retrieval operations based on sparse vectors and knowledge retrieval operations based on dense vectors.

[0069] Step S202: If multiple initial response messages meet preset conditions, sort the multiple initial response messages to obtain multiple processed initial response messages.

[0070] Optionally, after obtaining multiple initial response information corresponding to multi-path knowledge retrieval, it can be first determined whether the multiple initial response information meets preset conditions. If the preset conditions are met, the retrieved knowledge can be sorted by the re-ranking module in the Agent architecture to obtain the processed multiple initial response information. For example, the re-ranking module compares the semantic relevance between the user question and the retrieved knowledge at the semantic level, and then sorts the retrieved knowledge according to the semantic relevance. It should be noted that the preset conditions may be that the matching degree between multiple initial response information and the target question information exceeds a threshold.

[0071] Step S203: Based on the processed initial response information, determine the target response information corresponding to the target question information.

[0072] Optionally, the generative large model in the Agent architecture outputs the target response information corresponding to the target question information based on the processed initial response information.

[0073] In summary, when performing knowledge retrieval based on target question information, compared with keyword matching in existing technologies, sparse vector-based knowledge retrieval operations and dense vector-based knowledge retrieval operations can effectively improve the knowledge coverage and accuracy of knowledge retrieval, thereby achieving the technical effect of improving the accuracy of answering questions.

[0074] To improve the accuracy of knowledge retrieval, in the data processing method provided in Embodiment 1 of this disclosure, before performing multi-path knowledge retrieval operations in the target knowledge base based on the target question information to obtain multiple initial response information, the method further includes: acquiring multiple knowledge information; if the format information corresponding to the target knowledge information is in text format, then performing multi-level block processing on the target knowledge information to obtain processed target knowledge information, wherein the target knowledge information is any one of the multiple knowledge information; and obtaining the target knowledge base based on the processed target knowledge information and the knowledge information other than the text format knowledge information among the multiple knowledge information.

[0075] Optionally, based on the technical field corresponding to the Agent architecture, multiple document knowledge items within that technical field are collected, i.e., the aforementioned multiple knowledge information items. Then, the format type of the knowledge information is determined, such as whether it is text or table format. If the target knowledge information is in text format, it undergoes multi-level segmentation, dividing it into different granularities, such as document granularity, paragraph granularity, and block granularity. This increases the number of segments, and the semantic cohesion also varies, ensuring coverage of different semantic levels. After obtaining the processed target knowledge information, a target knowledge base is constructed using the processed target knowledge information and other knowledge information (excluding text format) from the multiple knowledge information items.

[0076] It should be noted that knowledge information in tabular format can be directly stored in the target knowledge base. Alternatively, the size of the knowledge information in tabular format can be used to determine whether to split it. If it is too large, it can be split into multiple smaller tables and then the split tables can be stored in the target knowledge base.

[0077] By segmenting knowledge information into different granularities, it is possible to ensure coverage of different semantic levels, thereby improving the accuracy of knowledge retrieval.

[0078] To improve the efficiency of knowledge retrieval, in the data processing method provided in Embodiment 1 of this disclosure, after multi-level block processing of the target knowledge information to obtain processed target knowledge information, the method further includes: if the block granularity corresponding to the processed target knowledge information is document granularity, then index information is constructed based on the document title information corresponding to the processed target knowledge information, wherein the index information is used to match with target question information to perform knowledge retrieval operation; if the block granularity corresponding to the processed target knowledge information is paragraph granularity, then index information is constructed based on the paragraph title information corresponding to the processed target knowledge information; if the block granularity corresponding to the processed target knowledge information is block granularity, then index information is constructed based on the question information that can be answered by the processed target knowledge information.

[0079] Optionally, after segmenting the target knowledge information to obtain processed target knowledge information, to avoid the need to match the question with the entire knowledge information during knowledge retrieval, data augmentation can be performed on the processed target knowledge information. This includes the following steps: If the segment granularity corresponding to the processed target knowledge information is document-level, then index information is constructed based on the document title information corresponding to the processed target knowledge information, i.e., the document title information is used as the index information, and the index information and question information are directly matched during subsequent knowledge retrieval. If the segment granularity corresponding to the processed target knowledge information is paragraph-level, then index information is constructed based on the paragraph title information corresponding to the processed target knowledge information, i.e., the paragraph title information is used as the index information. If the segment granularity corresponding to the processed target knowledge information is block-level, then index information is constructed based on the question information that the processed target knowledge information can answer, i.e., the question information is used as the index information.

[0080] Furthermore, when constructing index information at the document granularity level based on the document title information corresponding to the processed target knowledge information, the method further includes: constructing first prompt information based on the document title information, wherein the first prompt information is used to instruct the generative big model to summarize the content corresponding to the processed target knowledge information; processing the first prompt information through the generative big model to obtain the overview information corresponding to the processed target knowledge information; constructing key-value pairs based on the overview information and document title information corresponding to the processed target knowledge information, and storing the key-value pairs in the target knowledge base, wherein, for document-granular knowledge information, during the knowledge retrieval operation, the overview information in the key-value pairs is used to constitute at least a plurality of initial response information, wherein, for document-granular knowledge information, during the knowledge retrieval operation, the overview information in the key-value pairs is used to constitute at least the plurality of initial response information.

[0081] When constructing index information, for document level (i.e., target knowledge information at the document granularity), a special prompt (i.e., the first prompt information mentioned above) is constructed based on the obtained document title information. Then, the content corresponding to the processed target knowledge information is summarized by a generative large model to obtain a summary of no more than 512 characters (i.e., the overview information mentioned above), which contains document-level information. Then, the overview information corresponding to the processed target knowledge information and the document title information are used to construct key-value pairs so that multiple initial response information can be formed by the overview information in the key-value pairs during subsequent knowledge retrieval operations.

[0082] It should be noted that for Paragraph level (i.e., target knowledge information at the paragraph level), a special prompt can also be constructed based on the obtained document title information. Then, a generative large model can be used to summarize the corresponding content of the processed target knowledge information to obtain a summary of no more than 512 characters, which makes it more focused on paragraph-level representation.

[0083] For chunk level (i.e., block-level target knowledge information), 1-2 core questions that users might ask are obtained through generative large models (i.e., questions that can be answered by the processed target knowledge information mentioned above), and the core questions are used as indexes for the block-level target knowledge information.

[0084] By designing a special prompt through the above steps to obtain relevant summaries or questions, three levels of data augmentation information are constructed. The obtained data augmentation information (i.e., index, overview information, and question information) can then be stored in the target knowledge base in the form of key-value pairs. For the document level and paragraph level, the key is the index (i.e., document title or paragraph title), and the value is the overview information. For the block level, the key is the index (i.e., question information), and the value is the processed target knowledge information itself.

[0085] Compared to knowledge retrieval based solely on knowledge information and question information, knowledge retrieval using the aforementioned index information and question information can effectively improve the efficiency of knowledge retrieval. Furthermore, based on the index information, the summary information corresponding to the relevant knowledge information can be used as the retrieved knowledge, thereby improving the accuracy of subsequent large-scale model responses.

[0086] When indexing a question, it is often difficult to summarize the key semantics of the knowledge information. Therefore, in the data processing method provided in Embodiment 1 of this disclosure, if the granularity of the processed target knowledge information is block-level, the index information is constructed based on the question information that can be answered by the processed target knowledge information. This includes: constructing second prompt information based on the processed target knowledge information, wherein the second prompt information is used to instruct the generative big model to output the question information that can be answered by the processed target knowledge information; processing the second prompt information through the generative big model to obtain the question information; and constructing the index information based on the question information.

[0087] Optionally, based on the processed target knowledge information, a second prompt is constructed. It should be noted that the second prompt is used to indicate to the generative large model that the processed target knowledge information can answer certain questions; that is, it outputs 1-2 core questions that users might ask (i.e., questions that the processed target knowledge information can answer). For example, the second prompt might be "What questions can xx knowledge answer?". Then, index information is constructed based on the questions corresponding to the processed target knowledge information. For example, after filtering out stop words from the question information, the filtered question information is used as the index.

[0088] By constructing a knowledge information index at the block level through questions, the core of the knowledge information can be more accurately summarized, thereby improving the accuracy of subsequent knowledge retrieval.

[0089] Document knowledge information often contains tabular data. Existing technologies either directly index tables or summarize large models before indexing. Direct indexing leads to inaccurate retrieval due to the structured data being vectorized. Conversely, summarizing large models often results in the loss of detailed information from the tables, making it difficult to retrieve knowledge when users ask specific questions. Therefore, in the data processing method provided in Embodiment 1 of this disclosure, after acquiring multiple pieces of knowledge information, the method further includes: if the format information corresponding to the target knowledge information is in tabular format, determining whether the size of the target knowledge information is greater than a target threshold; if the size of the target knowledge information is less than or equal to the target threshold, converting the target knowledge information to obtain unstructured text information; and constructing index information based on the unstructured text information.

[0090] Optionally, if the target knowledge information is in tabular format, the system first determines whether the size of the target knowledge information exceeds a target threshold, for example, whether the size of the target knowledge information exceeds 512k. If the size of the target knowledge information is less than or equal to the target threshold, the system directly converts the tabular knowledge information into unstructured text information. Finally, the system constructs index information based on the unstructured text information, thus determining the unstructured text information as the index information.

[0091] If the size of the target knowledge information exceeds the target threshold, the target knowledge information is split into tables. It's important to note that during splitting, each split table must retain its corresponding header information. After obtaining the split tables, they are converted into unstructured text knowledge, and then an index is built.

[0092] For example, the target knowledge information shown in Table 1:

[0093] Table 1

[0094] After converting the tabular knowledge information into unstructured text information, it is represented as "Operation: yundun-sas:RollbackSuspEventQuaraFile Access Level: Read Resource Type: *All Resources acs:yundun-sas:\\*::\\*Associated Operation: None".

[0095] By transforming structured tabular knowledge into unstructured text knowledge and then building an index, it is possible to cover detailed information in the table and improve the accuracy of subsequent knowledge retrieval.

[0096] To address the issue of low accuracy in knowledge retrieval based on keywords in existing technologies, the data processing method provided in Embodiment 1 of this disclosure performs a multi-path knowledge retrieval operation in a target knowledge base based on target question information to obtain multiple initial response information. This includes: performing a sparse vector knowledge retrieval operation in the target knowledge base based on the sparse vector corresponding to the target question information to obtain a first retrieval result; performing a dense vector knowledge retrieval operation in the target knowledge base based on the dense vector corresponding to the target question information to obtain a second retrieval result; performing a keyword knowledge retrieval operation in the target knowledge base based on the keywords corresponding to the target question information to obtain a third retrieval result; performing a knowledge retrieval operation in the target knowledge base based on the target question information using the maximum boundary correlation algorithm to obtain a fourth retrieval result; and obtaining multiple initial response information based on the first, second, third, and fourth retrieval results.

[0097] Optionally, performing multi-path knowledge retrieval operations in the target knowledge base includes: performing sparse vector knowledge retrieval operations in the target knowledge base based on the sparse vector corresponding to the target question information to obtain the first retrieval result. For example, sparse vector retrieval can be based on traditional information retrieval models, such as TF-IDF (Term Frequency-Inverse Document Frequency, a commonly used weighting technique for information retrieval and data mining), representing text as sparse feature vectors. In this representation, each dimension corresponds to a word in the vocabulary, and the value of the vector usually reflects the importance of the word in the document. The advantage of sparse vector retrieval is its high computational efficiency and ease of implementation, but it may not be able to fully capture the deep semantics of the text. Alternatively, sparse vectors can be automatically calculated using a text vectorization model, such as text-embedding-v3. Then, a retrieval algorithm, such as the nearest neighbor search algorithm HNSW, can be used to perform retrieval based on the sparse vectors to obtain the aforementioned first retrieval result.

[0098] The second recall result is obtained by performing a dense vector knowledge recall operation in the target knowledge base based on the dense vector corresponding to the target question information. That is, the target question information is converted into a dense vector, and then the dense vector corresponding to the target question information and the dense vector corresponding to the index are matched by a retrieval algorithm to obtain the above-mentioned second recall result.

[0099] The third recall result is obtained by performing a keyword knowledge retrieval operation in the target knowledge base based on the keywords corresponding to the target question information. That is, the keywords corresponding to the target question information are extracted, and then the keywords corresponding to the target question information are matched with the index in the target knowledge base to obtain the above-mentioned third recall result.

[0100] The fourth recall result is obtained by performing a knowledge retrieval operation in the target knowledge base using the Maximum Marginal Relevance (MMR) algorithm based on the target question information. The MMR algorithm aims to balance document relevance and redundancy by selecting documents with a high MMR value. The MMR algorithm ranks documents by calculating the similarity between the query text (i.e., the target question information mentioned above) and the searched documents. This ensures that the selected documents are highly relevant to the query topic. In addition to relevance, MMR also considers the similarity between previously selected results and newly selected results to avoid selecting overly similar documents, thereby increasing the diversity of the results.

[0101] Finally, based on the results of the first recall, the second recall, the third recall, and the fourth recall, multiple initial response information is obtained, that is, the response information from the first recall, the second recall, the third recall, and the fourth recall results is used as multiple initial response information.

[0102] By using multi-path retrieval, including dense vector retrieval, sparse vector retrieval, keyword retrieval, and MMR retrieval, we can more comprehensively cover the details of knowledge information and improve the accuracy of knowledge retrieval.

[0103] To improve the efficiency of knowledge retrieval, in the data processing method provided in Embodiment 1 of this disclosure, a sparse vector knowledge retrieval operation is performed in the target knowledge base based on the sparse vector corresponding to the target question information to obtain a first retrieval result. This includes: vectorizing the target question information to obtain a first sparse vector; obtaining a second sparse vector corresponding to the index information in the target knowledge base; and performing a sparse vector knowledge retrieval operation in the target knowledge base based on the first and second sparse vectors to obtain the first retrieval result.

[0104] Optionally, performing sparse vector knowledge retrieval operations in the target knowledge base includes: vectorizing the target question information to obtain a first sparse vector, for example, representing text as a sparse feature vector based on a traditional information retrieval model, such as TF-IDF. Another example is automatically calculating sparse vectors using a text vectorization model, such as text-embedding-v3. Then, obtaining the second sparse vector corresponding to the index information in the target knowledge base. It should be noted that the sparse vector corresponding to the index can be generated using a sparse vector model during the construction phase of the target knowledge base and stored in the target knowledge base. Finally, performing sparse vector knowledge retrieval operations in the target knowledge base based on the first and second sparse vectors yields the aforementioned first retrieval result.

[0105] By performing knowledge retrieval using the second sparse vector corresponding to the index information and the first index vector corresponding to the target question information, the efficiency of knowledge retrieval can be effectively improved compared to the existing technology of matching the target question information with the document information itself.

[0106] To improve the efficiency of obtaining response information, in the data processing method provided in Embodiment 1 of this disclosure, a knowledge recall operation is performed on the target knowledge base based on a first sparse vector and a second sparse vector to obtain a first recall result. This includes: calculating the similarity between the first sparse vector and the second sparse vector, and determining the target index from the index information in the target knowledge base based on the similarity; performing a knowledge recall operation on the target knowledge base based on the target index to obtain the key-value pairs corresponding to the target index and the knowledge information corresponding to the target index; and obtaining the first recall result based on the key-value pairs corresponding to the target index and / or the knowledge information corresponding to the target index.

[0107] Optionally, a similarity score is calculated between the first and second sparse vectors. Then, based on the similarity score, a target index is determined from the index information in the target knowledge base. For example, the top five indexes ranked by similarity score are selected as the target index. Next, knowledge retrieval is performed in the target knowledge base based on the target index to obtain the key-value pairs and knowledge information corresponding to the target index. Finally, based on the key-value pairs and / or knowledge information corresponding to the target index, a first retrieval result is obtained, which is the value and / or knowledge information in the key-value pairs corresponding to the target index.

[0108] It should be noted that if the knowledge information corresponding to the target index is at the document granularity, it is necessary to determine whether the size of the knowledge information exceeds the threshold (it should be noted that this threshold can be determined based on the size of knowledge that the generative large model can process).

[0109] If the size of the knowledge information exceeds the threshold, the first recall result is obtained based on the overview information in the key-value pair corresponding to the target index. If the size of the knowledge does not exceed the threshold, the key-value pair corresponding to the target index and the knowledge information corresponding to the target index are used as the first recall result.

[0110] If the knowledge information corresponding to the target index is at other granularities, such as paragraph granularity or block granularity, then the knowledge information corresponding to the target index will be used as the first recall result.

[0111] By selecting target indices based on similarity and obtaining recall results based on the key-value pairs and / or knowledge information corresponding to the target indices, the accuracy of knowledge recall is improved.

[0112] Because different strategies are used to recall knowledge, the score calculation methods for each strategy are different, and the score ranges are inconsistent, making it impossible to directly sort and filter based on the scores. Therefore, in the data processing method provided in Embodiment 1 of this disclosure, before sorting the multiple initial response information to obtain the processed multiple initial response information, the method further includes: obtaining multiple first recall scores corresponding to the initial response information in the multiple initial response information, wherein the multiple first recall scores are obtained when performing multi-way knowledge recall operations in the target knowledge base; determining the matching value corresponding to the initial response information in the multiple initial response information based on the multiple first recall scores, wherein the matching value is used to characterize the degree of matching between the multiple initial response information and the target question information; and judging whether the multiple initial response information meets the preset conditions based on the matching value.

[0113] Optionally, after obtaining multiple initial response information corresponding to multi-way knowledge retrieval, the retrieved knowledge can be sorted using the re-ranking module in the intelligent agent architecture to obtain the processed multiple initial response information. Since dense vector, sparse vector, full-text keyword retrieval, and MMR retrieval strategies are used in the multi-way retrieval stage, the score calculation methods of different retrieval strategies are inconsistent, resulting in different score ranges that cannot be directly compared. Therefore, after obtaining multiple first recall scores corresponding to the initial response information among the multiple initial response information, the matching value corresponding to the initial response information among the multiple initial response information is determined based on the multiple first recall scores. For example, all scores are normalized to the range of 0 to 1, and then the normalized score is used as the matching value. Finally, based on the matching value, it is determined whether the multiple initial response information meets the preset conditions. For example, if the matching value is higher than a threshold, the multiple initial response information meets the preset conditions; if the matching value is lower than the threshold, the multiple initial response information does not meet the preset conditions. The matching value can accurately assess whether the multiple initial response information always answers the target question information, improving the accuracy of subsequent responses to the target question information.

[0114] To improve the accuracy and rationality of the matching value, in the data processing method provided in Embodiment 1 of this disclosure, determining the matching value corresponding to the initial response information among multiple initial response information based on multiple first recall scores includes: normalizing the multiple first recall scores to obtain a normalized first recall score; calculating the similarity between the target question information and the directory information corresponding to the initial response information among the multiple initial response information to obtain a second recall score; calculating the similarity between the target question information and the initial response information among the multiple initial response information to obtain a third recall score; calculating the similarity between the target question information and the title information corresponding to the initial response information among the multiple initial response information to obtain a fourth recall score; and calculating the matching value based on the normalized first recall score, second recall score, third recall score, and fourth recall score.

[0115] Optionally, multiple first recall scores can be normalized to obtain a normalized first recall score. It should be noted that the first recall score includes the recall score corresponding to sparse vectors, the recall score corresponding to dense vectors, the recall score corresponding to MMR recall, and the recall score for keyword recall. For example, the recall score corresponding to dense vectors is the first factor score: dense vector recall factor score (dense_vector_recall_factor_score). Because the dense vector score itself is in the range of 0 to 1, it does not need to be normalized. The recall score for keyword recall is the second factor score: the full-text recall factor score (fulltext_recall_factor_score). At this point, the current recalled knowledge score needs to be normalized. The normalization formula is as follows, where i is the current knowledge index: fulltext_recall_factor_score = (all_docs_score[i]['fulltext'] - fulltext_min_score) / (fulltext_max_score - fulltext_min_score) if (fulltext_max_score - fulltext_min_score > 0), where all_docs_score[i]['fulltext'] is the score corresponding to the i-th index, fulltext_min_score is the minimum score of all knowledge obtained by the full-text keyword recall strategy, and fulltext_max_score is the maximum score of all knowledge obtained by the full-text keyword recall strategy.

[0116] The recall score corresponding to the sparse vector is the third factor score: the sparse vector recall factor score (sparse_vector_recall_factor_score). The normalization formula is as follows: sparse_vector_recall_factor_score = (all_docs_score[i]['sparse'] - sparse_min_score) / (sparse_max_score - sparse_min_score) if (sparse_max_score - sparse_min_score > 0). `all_docs_score[i]['sparse']` represents the score corresponding to the i-th index, `sparse_max_score` is the maximum score obtained from all knowledge points using the sparse vector recall strategy, and `sparse_min_score` is the minimum score obtained from all knowledge points using the sparse vector recall strategy.

[0117] The recall score corresponding to MMR recall is the fourth factor score: MMR recall factor score. Since MMR is calculated based on dense vector retrieval, the dense vector score itself is in the range of 0 to 1, so it does not need to be normalized and can be used directly.

[0118] After obtaining the normalized first recall score, a fifth factor score can be calculated based on the similarity between the target question information and the directory information corresponding to the initial responses from multiple initial response information: the document directory factor score (i.e., the second recall score mentioned above). In an optional embodiment, the Jaccard Distance between the keywords of the current question and the segmented document directory can be calculated as the second recall score. It should be noted that the keywords can be extracted using a large model, and some open-source tools such as jieba can be used for segmentation. For example, if the user's question is "How to detect database baseline verification", the segmented result would be {'baseline', 'database', 'detection'}.

[0119] A sixth factor score can be calculated based on the similarity between the target question information and the initial response information from multiple initial response information: the content text factor score (i.e., the third recall score mentioned above). Similarly, the Jaccard Distance between the current question and the indexed content can be used as the third recall score mentioned above.

[0120] A seventh factor score can be calculated based on the similarity between the target question information and the title information corresponding to the initial responses from multiple initial responses: the title enhancement factor score (i.e., the fourth recall score mentioned above). Similarly, the Jaccard Distance between the current question and the title can be used as the fourth recall score. It should be noted that a large model can be used to enhance the title information, adding additional auxiliary information. For example, if the first-level title of the current knowledge is head1 = "Manage Server" and the second-level title is head2 = "View Server Information", the title enhancement could be title_enhance = "The following is part of the content from the document 'Manage Server'".

[0121] After obtaining the normalized first recall score, second recall score, third recall score, and fourth recall score, the total score is obtained by weighted summation of all recall scores, thus obtaining the matching value.

[0122] By using a multi-factor filtering algorithm, the quality of knowledge recall data is further improved on the basis of coarse recall, thereby improving the accuracy of knowledge recall.

[0123] To further improve the quality of knowledge retrieval, in the data processing method provided in Embodiment 1 of this disclosure, determining whether multiple initial response information meets preset conditions based on the matching value includes: determining whether the matching value meets preset requirements; if the matching value meets the preset requirements, then the multiple initial response information meets the preset conditions. If the matching value does not meet the preset requirements, then multiple similar question information is generated based on the target question information using a generative large model; the step of performing multi-path knowledge retrieval operation in the target knowledge base based on the multiple similar question information is executed again to obtain multiple initial response information corresponding to the secondary knowledge retrieval, and the processed initial response information is obtained based on the multiple initial response information corresponding to the secondary knowledge retrieval.

[0124] Optionally, after obtaining the matching value, it can be determined whether the matching value meets a preset requirement. For example, the preset requirement could be that the larger value in the matching value must be greater than a matching threshold, or that the smaller value in the matching value must be greater than the matching threshold. If the matching value meets the preset requirement, the multiple initial response information is directly sorted according to the matching value to obtain multiple processed initial response information. If the matching value does not meet the preset requirement, a secondary retrieval is performed. Multiple similar question information is generated based on the target question information using a generative large model. The step of performing multi-way knowledge retrieval operation in the target knowledge base based on the multiple similar question information is performed again to obtain multiple initial response information corresponding to the secondary knowledge retrieval. The processed initial response information is then obtained based on the multiple initial response information corresponding to the secondary knowledge retrieval.

[0125] By using multiple similar questions for secondary retrieval, the coverage of recalled knowledge can be increased, thereby further improving the quality of knowledge retrieval.

[0126] In an optional embodiment, a typical interaction agent architecture for a large-model-based security knowledge question-answering system is shown in Figure 3. The architecture includes: a multi-turn dialogue module, responsible for rewriting the current question based on historical dialogue data to enrich the question information; an intent recognition module, responsible for identifying the intent of the current user's question, determining the specific intent (chat, question-answering, tool invocation), and executing different logic for different intents; a multi-way retrieval coarse ranking module, responsible for knowledge retrieval; a re-ranking module, which filters and selects high-quality knowledge due to the large volume and varying quality of retrieved knowledge and the different score calculation methods of different retrieval strategies; a response strategy routing module, which uses different response strategies to respond to questions based on the quality of the retrieved knowledge; a post-processing module, which controls whether the response content is streamed or batch-output and can control whether to include links to knowledge within the answer; a tool interface, including various toolsets that need to be called; and a knowledge base (vector library) module, which stores knowledge and involves operations such as knowledge slicing and indexing.

[0127] The optimizations disclosed herein revolve around the multi-way retrieval coarse-ranking module, the re-ranking module, and the knowledge base (vector library) module. The optimization of the knowledge base (vector library) module is as follows: Traditional knowledge base construction involves segmenting document information based on delimiters and controlling the number of tokens in each segment. An index is then built based on these segments, configured for vectorized or keyword-based retrieval. However, because this method generates an index from original document fragments, the semantics are often formal and do not represent the full picture of the article. User questions are often colloquial, concise, and lacking in information, making effective retrieval matching difficult and resulting in poor knowledge retrieval quality. The optimization methods at this stage include constructing semantic-level segments and data augmentation.

[0128] The text is divided into document-level, paragraph-level, and block-level granularity, as shown in Figure 4. This differentiation increases the number of blocks while also varying the semantic cohesion, ensuring coverage of different semantic levels. By interacting with a large model and designing special prompts to obtain relevant summaries or questions, data augmentation information at three levels is constructed and stored in the knowledge base.

[0129] The optimization of the multi-way retrieval coarse ranking module is as follows: construct multi-way recall, such as dense vector, sparse vector, full-text keyword retrieval, and MMR retrieval strategy, and use multiple recall strategies to improve the recall effect.

[0130] The reordering module is optimized as follows: The reordering module is used to sort the recalled knowledge, and multi-factor fine-grained ranking is used to improve the reordering effect.

[0131] By optimizing the knowledge base through the above-mentioned points, including index optimization, multi-way retrieval optimization, and re-ranking optimization, the overall knowledge retrieval effect can be improved.

[0132] In the data processing method provided in Embodiment 1 of this disclosure, multiple initial response information is obtained by performing a multi-path knowledge retrieval operation in the target knowledge base based on the target question information. The multi-path knowledge retrieval operation includes at least two methods: a sparse vector-based knowledge retrieval operation and a dense vector-based knowledge retrieval operation. If the multiple initial response information meets preset conditions, the multiple initial response information is sorted to obtain processed initial response information. Based on the processed initial response information, the target response information corresponding to the target question information is determined. This solves the technical problem that retrieving response information for related questions using keyword matching results in low accuracy of the response information. In this solution, when performing knowledge retrieval based on the target question information, compared to keyword matching in the prior art, the sparse vector-based knowledge retrieval operation and the dense vector-based knowledge retrieval operation can effectively improve the knowledge coverage and accuracy of the retrieval, thereby achieving the technical effect of improving the accuracy of answering questions.

[0133] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this disclosure is not limited to the described order of actions, because according to this disclosure, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and components involved are not necessarily essential to this disclosure.

[0134] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solutions of this disclosure, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this disclosure.

[0135] According to an embodiment of this disclosure, a data processing method is also provided, as shown in FIG5, the data processing method comprising:

[0136] Step S501: Obtain the target problem information uploaded by the client;

[0137] Step S502: Based on the target question information, perform a multi-path knowledge retrieval operation in the target knowledge base on the cloud server to obtain multiple initial response information. The multi-path knowledge retrieval operation includes at least: a knowledge retrieval operation based on sparse vectors and a knowledge retrieval operation based on dense vectors. If the multiple initial response information meets preset conditions, sort the multiple initial response information to obtain multiple processed initial response information. Based on the multiple processed initial response information, determine the target response information corresponding to the target question information.

[0138] Step S503: Return the target response information to the client.

[0139] It should be noted that the method for processing the target problem information in the cloud server is the same as the method provided in Embodiment 1 of this disclosure, and will not be repeated here.

[0140] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this disclosure is not limited to the described order of actions, because according to this disclosure, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and components involved are not necessarily essential to this disclosure.

[0141] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solutions of this disclosure, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this disclosure.

[0142] According to an embodiment of this disclosure, a data processing apparatus for implementing the above-described data processing method is also provided, as shown in FIG6. The apparatus includes: a first execution component 601, a first processing component 602, and a first determination component 603.

[0143] The first execution component 601 is configured to perform a multi-way knowledge retrieval operation in the target knowledge base based on the target question information to obtain multiple initial response information. The multi-way knowledge retrieval operation includes at least: a knowledge retrieval operation based on sparse vectors and a knowledge retrieval operation based on dense vectors.

[0144] The first processing component 602 is configured to sort the multiple initial response information when multiple initial response information meets preset conditions, so as to obtain multiple processed initial response information.

[0145] The first determining component 603 is configured to determine the target response information corresponding to the target question information based on the processed multiple initial response information.

[0146] Optionally, in the data processing apparatus provided in Embodiment 3 of this disclosure, the first execution component 601 performs a multi-path knowledge retrieval operation in the target knowledge base based on the target question information to obtain multiple initial response information. The multi-path knowledge retrieval operation includes at least: a sparse vector-based knowledge retrieval operation and a dense vector-based knowledge retrieval operation. The first processing component 602 sorts the multiple initial response information when they meet preset conditions to obtain processed initial response information. The first determining component 603 determines the target response information corresponding to the target question information based on the processed initial response information. This solves the technical problem that retrieving response information for related questions using keyword matching results in low accuracy of the response information. In this solution, when performing knowledge retrieval based on the target question information, compared to keyword matching in the prior art, the sparse vector-based knowledge retrieval operation and the dense vector-based knowledge retrieval operation can effectively improve the knowledge coverage of the retrieval and increase the accuracy of the knowledge retrieval, thereby achieving the technical effect of improving the accuracy of answering questions.

[0147] Optionally, in the data processing apparatus provided in Embodiment 3 of this disclosure, the apparatus further includes: a first acquisition component, configured to acquire multiple knowledge information before performing a multi-way knowledge retrieval operation in the target knowledge base based on the target question information to obtain multiple initial response information; a block segmentation component, configured to perform multi-level block processing on the target knowledge information if the format information corresponding to the target knowledge information is in text format, to obtain processed target knowledge information, wherein the target knowledge information is any one of the multiple knowledge information; and a second determination component, configured to obtain the target knowledge base based on the processed target knowledge information and the knowledge information other than the text format knowledge information among the multiple knowledge information.

[0148] Optionally, in the data processing apparatus provided in Embodiment 3 of this disclosure, the apparatus further includes: a first construction component, configured to, after performing multi-level block processing on the target knowledge information to obtain processed target knowledge information, if the block granularity corresponding to the processed target knowledge information is document granularity, then construct index information based on the document title information corresponding to the processed target knowledge information, wherein the index information is used to match with target question information to perform a knowledge retrieval operation; a second construction component, configured to, if the block granularity corresponding to the processed target knowledge information is paragraph granularity, then construct index information based on the paragraph title information corresponding to the processed target knowledge information; and a third construction component, configured to, if the block granularity corresponding to the processed target knowledge information is block granularity, then construct index information based on the question information that the processed target knowledge information can answer.

[0149] Optionally, in the data processing apparatus provided in Embodiment 3 of this disclosure, the apparatus further includes: a fourth construction component, configured to construct first prompt information based on the document title information when constructing index information based on the document title information corresponding to the processed target knowledge information, wherein the first prompt information is used to instruct the generative big model to summarize the content corresponding to the processed target knowledge information; a second processing component, configured to process the first prompt information through the generative big model to obtain the overview information corresponding to the processed target knowledge information; and a fifth construction component, configured to construct key-value pairs based on the overview information and document title information corresponding to the processed target knowledge information, and store the key-value pairs in the target knowledge base, wherein, for document-level knowledge information, during knowledge retrieval operations, the overview information in the key-value pairs is used to constitute at least multiple initial response information.

[0150] Optionally, in the data processing apparatus provided in Embodiment 3 of this disclosure, the third construction component includes: a first construction sub-component, configured to construct second prompt information based on the processed target knowledge information, wherein the second prompt information is used to indicate that the generative large model outputs question information that can be answered by the processed target knowledge information; a first processing sub-component, configured to process the second prompt information through the generative large model to obtain question information; and a second construction sub-component, configured to construct index information based on the question information.

[0151] Optionally, in the data processing apparatus provided in Embodiment 3 of this disclosure, the apparatus further includes: a judgment component, configured to, after acquiring multiple pieces of knowledge information, determine whether the size of the target knowledge information is greater than a target threshold if the format information corresponding to the target knowledge information is in table format; a conversion component, configured to convert the target knowledge information to unstructured text information if the size of the target knowledge information is less than or equal to the target threshold; and a fifth construction component, configured to construct index information based on the unstructured text information, wherein the index information is used to match with the target question information to perform a knowledge retrieval operation.

[0152] Optionally, in the data processing apparatus provided in Embodiment 3 of this disclosure, the execution component includes: a first operation sub-component, configured to perform a sparse vector knowledge recall operation in the target knowledge base based on the sparse vector corresponding to the target question information, to obtain a first recall result; a second operation sub-component, configured to perform a dense vector knowledge recall operation in the target knowledge base based on the dense vector corresponding to the target question information, to obtain a second recall result; a third operation sub-component, configured to perform a keyword knowledge recall operation in the target knowledge base based on the keyword corresponding to the target question information, to obtain a third recall result; a fourth operation sub-component, configured to perform a knowledge recall operation in the target knowledge base based on the target question information using the maximum boundary correlation algorithm, to obtain a fourth recall result; and a first determination sub-component, configured to obtain multiple initial response information based on the first recall result, the second recall result, the third recall result, and the fourth recall result.

[0153] Optionally, in the data processing apparatus provided in Embodiment 3 of this disclosure, the first operation sub-component includes: a processing component configured to perform vectorization processing on the target question information to obtain a first sparse vector; an acquisition component configured to acquire a second sparse vector corresponding to the index information in the target knowledge base; and an execution component configured to perform a knowledge retrieval operation on the sparse vector in the target knowledge base based on the first sparse vector and the second sparse vector to obtain a first retrieval result.

[0154] Optionally, in the data processing apparatus provided in Embodiment 3 of this disclosure, the execution component includes: a calculation subcomponent, configured to perform similarity calculation on the first sparse vector and the second sparse vector, and determine the target index from the index information in the target knowledge base based on the similarity; an operation subcomponent, configured to perform a knowledge retrieval operation in the target knowledge base based on the target index, to obtain the key-value pair corresponding to the target index and the knowledge information corresponding to the target index; and a determination subcomponent, configured to obtain a first retrieval result based on the key-value pair corresponding to the target index and / or the knowledge information corresponding to the target index.

[0155] Optionally, in the data processing apparatus provided in Embodiment 3 of this disclosure, the apparatus further includes: a second acquisition component, configured to acquire multiple first recall scores corresponding to the initial response information among the multiple initial response information before sorting and processing the multiple initial response information to obtain the processed multiple initial response information, wherein the multiple first recall scores are obtained when performing a multi-way knowledge recall operation in the target knowledge base; a third determination component, configured to determine the matching value corresponding to the initial response information among the multiple initial response information based on the multiple first recall scores, wherein the matching value is used to characterize the degree of matching between the multiple initial response information and the target question information; and a third processing component, configured to determine whether the multiple initial response information meets the preset condition based on the matching value.

[0156] Optionally, in the data processing apparatus provided in Embodiment 3 of this disclosure, the third determining component includes: a second processing subcomponent, configured to normalize multiple first recall scores to obtain normalized first recall scores; a first calculation subcomponent, configured to perform similarity calculation based on the target question information and the directory information corresponding to the initial response information among multiple initial response information to obtain a second recall score; a second calculation subcomponent, configured to perform similarity calculation based on the target question information and the initial response information among multiple initial response information to obtain a third recall score; a third calculation subcomponent, configured to perform similarity calculation based on the target question information and the title information corresponding to the initial response information among multiple initial response information to obtain a fourth recall score; and a fourth calculation subcomponent, configured to calculate a matching value based on the normalized first recall score, second recall score, third recall score, and fourth recall score.

[0157] Optionally, in the data processing apparatus provided in Embodiment 3 of this disclosure, the third processing component includes: a judgment sub-component, configured to judge whether the matching value meets the preset requirements; and a third processing sub-component, configured to, if the matching value meets the preset requirements, then the plurality of initial response information meets the preset conditions.

[0158] Optionally, in the data processing apparatus provided in Embodiment 3 of this disclosure, the apparatus further includes: a generation component, configured to, after determining whether the matching value meets the preset requirements, generate multiple similar question information based on the target question information through a generative large model if the matching value does not meet the preset requirements; and a second execution component, configured to execute again the step of performing a multi-way knowledge recall operation in the target knowledge base based on the multiple similar question information, to obtain multiple initial response information corresponding to the secondary knowledge recall, and to obtain the processed initial response information based on the multiple initial response information corresponding to the secondary knowledge recall.

[0159] It should be noted that the first execution component 601, the first processing component 602, and the first determining component 603 mentioned above correspond to steps S201 to S203 in the previous embodiment. The three components and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the previous embodiment. It should be noted that the above components, as part of the device, can run in the computer terminal 10 provided in the previous embodiment.

[0160] It should be noted that the preferred implementation schemes involved in the above embodiments of this disclosure are the same as the schemes, application scenarios and implementation processes provided in Embodiment 1, but are not limited to the schemes provided in Embodiment 1.

[0161] Embodiments of this disclosure can provide an electronic device, which can be any electronic device in a group of electronic device terminals. Optionally, in this embodiment, the aforementioned electronic device can also be replaced with a terminal device such as a mobile terminal.

[0162] Optionally, in this embodiment, the aforementioned electronic device may be located in at least one of a plurality of network devices in a computer network.

[0163] In this embodiment, the electronic device described above can execute the program code of the steps in the data processing method provided in any of the above embodiments.

[0164] Optionally, FIG7 is a structural block diagram of an electronic device according to an embodiment of the present disclosure. As shown in FIG7, the electronic device 70 may include: one or more (only one is shown in FIG7) processors 702 and memory 704. The electronic device 70 may also include a storage controller for controlling and managing the memory 704; the electronic device 70 may also include a peripheral interface for connecting radio frequency components, audio components, and a display screen, etc.

[0165] The memory can be used to store software programs and components, such as the program instructions / components corresponding to the data processing method and apparatus in this disclosure embodiment. The processor executes various functional applications and data processing by running the software programs and components stored in the memory, thereby implementing the aforementioned data processing method. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the electronic device 70 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0166] The processor can invoke information and application programs stored in the memory through the transmission device to execute the steps in the data processing method provided in any of the above embodiments.

[0167] It will be understood by those skilled in the art that the structure shown in FIG7 is merely illustrative, and the electronic device 70 may also be a smartphone (such as an Android phone, an iOS phone, etc.), a tablet computer, a PDA, a mobile internet device (MID), a PAD, or other terminal device. FIG7 does not limit the structure of the aforementioned electronic device. For example, the electronic device 70 may also include more or fewer components (such as a network interface, a display device, etc.) than shown in FIG7, or may have a different configuration than that shown in FIG7.

[0168] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0169] The embodiments of this disclosure also provide a computer-readable storage medium. Optionally, in this embodiment, the computer program product described above can be used to store the program code executed by the data processing method provided in Embodiment 1 above.

[0170] Embodiments of this disclosure also provide a computer program product. Optionally, in this embodiment, the computer program product can be used to store the program code executed by the data processing method provided in Embodiment 1.

[0171] Optionally, in this embodiment, the computer program product may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals.

[0172] The sequence numbers of the embodiments disclosed above are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0173] In the above embodiments of this disclosure, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0174] In the several embodiments provided in this disclosure, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of components is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple components or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interface; the indirect coupling or communication connection of components or components may be electrical or other forms.

[0175] The components described as separate parts may or may not be physically separate. The components shown as components may or may not be physical components; that is, they may be located in one place or distributed across multiple network components. Some or all of the components can be selected to achieve the purpose of this embodiment according to actual needs.

[0176] Furthermore, the functional components in the various embodiments of this disclosure can be integrated into a single processing component, or each component can exist physically separately, or two or more components can be integrated into a single component. The integrated components described above can be implemented in hardware or as software functional components.

[0177] If the integrated components are implemented as software functional components and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0178] The above description is only a preferred embodiment of this disclosure. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles of this disclosure, and these improvements and modifications should also be considered within the scope of protection of this disclosure. Industrial applicability

[0179] The solution provided in this disclosure can be applied to data processing. It employs a multi-path knowledge retrieval operation based on target question information in a target knowledge base to obtain multiple initial response messages. The multi-path knowledge retrieval operation includes at least two methods: sparse vector-based knowledge retrieval and dense vector-based knowledge retrieval. If the multiple initial response messages meet preset conditions, they are sorted to obtain processed initial response messages. Based on the processed initial response messages, the target response message corresponding to the target question information is determined. By performing multi-path knowledge retrieval operations in the target knowledge base, the coverage of knowledge retrieval is improved, thereby enhancing the accuracy of answering questions. This solves the technical problem that retrieving response information for related questions using keyword matching results in relatively low accuracy of the response information.

Claims

1. A data processing method, comprising: Based on the target question information, a multi-way knowledge retrieval operation is performed in the target knowledge base to obtain multiple initial response information. The multi-way knowledge retrieval operation includes at least: a knowledge retrieval operation based on sparse vectors and a knowledge retrieval operation based on dense vectors. If the multiple initial response messages meet the preset conditions, the multiple initial response messages are sorted to obtain multiple processed initial response messages; Based on the processed initial response information, the target response information corresponding to the target question information is determined.

2. The method according to claim 1, wherein, Before performing a multi-way knowledge retrieval operation in the target knowledge base based on the target question information to obtain multiple initial response information, the method further includes: Acquire multiple pieces of knowledge and information; If the format information corresponding to the target knowledge information is in text format, then the target knowledge information is processed into multi-level blocks to obtain the processed target knowledge information, wherein the target knowledge information is any one of the multiple knowledge information; The target knowledge base is obtained based on the processed target knowledge information and the knowledge information other than text format among the multiple knowledge information.

3. The method according to claim 2, wherein, After performing multi-level block processing on the target knowledge information to obtain the processed target knowledge information, the method further includes: If the granularity of the block corresponding to the processed target knowledge information is the document granularity, then based on the document title information corresponding to the processed target knowledge information, index information is constructed, wherein the index information is used to match with the target question information to perform a knowledge retrieval operation.

4. The method according to claim 3, wherein, After performing multi-level block processing on the target knowledge information to obtain the processed target knowledge information, the method further includes: If the block granularity corresponding to the processed target knowledge information is paragraph granularity, then the index information is constructed based on the paragraph title information corresponding to the processed target knowledge information.

5. The method according to claim 4, wherein, After performing multi-level block processing on the target knowledge information to obtain the processed target knowledge information, the method further includes: If the granularity of the processed target knowledge information is block-level, then the index information is constructed based on the question information that can be answered by the processed target knowledge information.

6. The method according to claim 5, wherein, When constructing index information based on the document title information corresponding to the processed target knowledge information, the method further includes: A first prompt information is constructed based on the document title information, wherein the first prompt information is used to instruct the generative big model to summarize and process the content corresponding to the processed target knowledge information. The first prompt information is processed by the generative large model to obtain the overview information corresponding to the processed target knowledge information; Based on the overview information corresponding to the processed target knowledge information and the document title information, key-value pairs are constructed and stored in the target knowledge base. For document-level knowledge information, during knowledge retrieval operations, the overview information in the key-value pairs is used to constitute at least the plurality of initial response information.

7. The method according to claim 5, wherein, If the granularity of the processed target knowledge information is block-level, then constructing the index information based on the question information that can be answered by the processed target knowledge information includes: Based on the processed target knowledge information, a second prompt information is constructed, wherein the second prompt information is used to instruct the generative large model to output the question information that the processed target knowledge information can answer; The second prompt information is processed by the generative large model to obtain the problem information; Based on the problem information, the index information is constructed.

8. The method according to claim 2, wherein, After acquiring multiple pieces of knowledge information, the method further includes: If the format of the target knowledge information is a table, then determine whether the size of the target knowledge information is greater than the target threshold. If the size of the target knowledge information is less than or equal to the target threshold, then the target knowledge information is formatted to obtain unstructured text information. Based on the unstructured text information, index information is constructed, wherein the index information is used to match with the target question information to perform a knowledge retrieval operation.

9. The method according to claim 1, wherein, Based on the target question information, a multi-path knowledge retrieval operation is performed in the target knowledge base to obtain multiple initial response information, including: Based on the sparse vector corresponding to the target question information, a sparse vector knowledge retrieval operation is performed in the target knowledge base to obtain a first retrieval result. Based on the dense vector corresponding to the target question information, a knowledge retrieval operation is performed on the dense vector in the target knowledge base to obtain a second retrieval result; Based on the keywords corresponding to the target question information, a keyword knowledge retrieval operation is performed in the target knowledge base to obtain a third retrieval result; Based on the target question information, a fourth recall result is obtained by performing a knowledge recall operation in the target knowledge base using the maximum boundary correlation algorithm; Based on the first recall result, the second recall result, the third recall result, and the fourth recall result, the plurality of initial response information is obtained.

10. The method according to claim 8, wherein, Based on the sparse vector corresponding to the target question information, a sparse vector knowledge retrieval operation is performed in the target knowledge base to obtain the first retrieval result, including: The target problem information is vectorized to obtain a first sparse vector; Obtain the second sparse vector corresponding to the index information in the target knowledge base; Based on the first sparse vector and the second sparse vector, a sparse vector knowledge recall operation is performed in the target knowledge base to obtain the first recall result.

11. The method according to claim 10, wherein, Based on the first sparse vector and the second sparse vector, a sparse vector knowledge retrieval operation is performed in the target knowledge base to obtain the first retrieval result, which includes: The similarity between the first sparse vector and the second sparse vector is calculated, and the target index is determined from the index information in the target knowledge base based on the similarity. Based on the target index, a knowledge retrieval operation is performed in the target knowledge base to obtain the key-value pair corresponding to the target index and the knowledge information corresponding to the target index. The first recall result is obtained based on the key-value pairs corresponding to the target index and / or the knowledge information corresponding to the target index.

12. The method according to claim 1, wherein, Before sorting the multiple initial response messages to obtain processed multiple initial response messages, the method further includes: Obtain multiple first recall scores corresponding to the initial response information from the multiple initial response information, wherein the multiple first recall scores are obtained when performing multi-way knowledge recall operations in the target knowledge base; Based on the plurality of first recall scores, a matching value is determined for the initial response information among the plurality of initial response information, wherein the matching value is used to characterize the degree of matching between the plurality of initial response information and the target question information; Based on the matching value, determine whether the multiple initial response messages meet the preset conditions.

13. The method according to claim 12, wherein, Based on the plurality of first recall scores, the matching value corresponding to the initial response information among the plurality of initial response information includes: The multiple first recall scores are normalized to obtain normalized first recall scores; A second recall score is obtained by calculating the similarity between the target question information and the directory information corresponding to the initial response information among the multiple initial response information; A third recall score is obtained by calculating the similarity between the target question information and the initial response information among the multiple initial response information; A fourth recall score is obtained by calculating the similarity between the target question information and the title information corresponding to the initial response information among the multiple initial response information; The matching value is obtained by calculating based on the normalized first recall score, second recall score, third recall score and fourth recall score.

14. The method according to claim 12, wherein, Determining whether the plurality of initial response messages meet the preset conditions based on the matching value includes: Determine whether the matching value meets the preset requirements; If the matching value meets the preset requirements, then the plurality of initial response messages meet the preset conditions.

15. The method according to claim 14, wherein, After determining whether the matching value meets the preset requirements, the method further includes: If the matching value does not meet the preset requirements, then the generative large model generates multiple similar problem information based on the target problem information; The process of performing a multi-path knowledge retrieval operation in the target knowledge base based on the multiple similar question information is executed again to obtain multiple initial response information corresponding to the secondary knowledge retrieval, and the processed initial response information is obtained based on the multiple initial response information corresponding to the secondary knowledge retrieval.

16. A data processing method, comprising: Obtain the target issue information uploaded by the client; Based on the target question information, a multi-path knowledge retrieval operation is performed in the target knowledge base on a cloud server to obtain multiple initial response information. The multi-path knowledge retrieval operation includes at least two methods: a sparse vector-based knowledge retrieval operation and a dense vector-based knowledge retrieval operation. If the multiple initial response information meets preset conditions, the multiple initial response information is sorted to obtain processed multiple initial response information. Based on the processed multiple initial response information, the target response information corresponding to the target question information is determined. The target response information is returned to the client.

17. A data processing apparatus, comprising: The first execution component is configured to perform a multi-way knowledge retrieval operation in the target knowledge base based on the target question information to obtain multiple initial response information. The multi-way knowledge retrieval operation includes at least: a knowledge retrieval operation based on sparse vectors and a knowledge retrieval operation based on dense vectors. The first processing component is configured to sort the multiple initial response information when the multiple initial response information meets the preset conditions, so as to obtain multiple processed initial response information. The first determining component is configured to determine the target response information corresponding to the target question information based on the processed multiple initial response information.

18. A computer-readable storage medium comprising a stored program, wherein, When the program is running, it controls the device containing the storage medium to execute the data processing method according to any one of claims 1 to 16.

19. An electronic device comprising: Memory, which stores executable programs; A processor configured to run the program, wherein the program, when running, performs the data processing method according to any one of claims 1 to 16.

20. A computer program product comprising a stored computer program that, when executed by a processor, implements the data processing method of any one of claims 1 to 16.

Images