Data processing method, research and development management device, storage medium and test equipment
Through the deductive reasoning and computing power of the big model and the multi-step reasoning process, the problem of insufficient accuracy of information search in multiple rounds of continuous question-and-answer scenarios is solved, and more efficient data retrieval and declared information judgment is achieved.
Patent Information
- Application Number
- CN202411962935.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-30
- Publication Date
- 2025-04-18
AI Technical Summary
In the existing technology, in multiple rounds of continuous question and answer scenarios, the accuracy of information search is insufficient and it is difficult to meet the requirements. In addition, there are difficulties in integrating large models with knowledge graphs or local databases, and the applicability of direct reasoning methods and zero-sample thinking chains are also difficult to meet the retrieval needs.
Using the deductive reasoning computing power that has been learned by the big model, through the first pre-inference step and the second correlation reasoning step, the input information and reference information are obtained and processed, and the first correlation information and the second correlation information are generated, further improving the accuracy of data feature mining and retrieval.
Through the multi-step reasoning process, the accuracy of information search is significantly improved, the AI computing power is extended, and the data retrieval efficiency is improved, especially the accuracy of decision-making on judgment statement information is improved.
Smart Images
Figure CN119849634A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to a data processing method, a research and development management device, a storage medium and a testing device. Background Art
[0002] Processes such as information retrieval and data mining are an important part of the development and application of artificial intelligence (AI), database and other systems. They can find output information with hints, goals or guiding content such as interest, logic, habit, reasonable reasoning, and appropriate deduction based on input information (search requirements, professional knowledge, etc.).
[0003] In scenarios such as product research and development (development tool application), its basic downstream tasks include summarizing the requirements of the detection system, generating test cases from the signals and configurations in the detection documents, or using the detection information to guide the troubleshooting of related issues; among them, judging the authenticity of the statement content is also a basic task; related technologies are usually combined with local databases (documents or knowledge graphs, etc.) for judgment, but the accuracy of its information search in specific scenarios is insufficient and it is difficult to meet the requirements.
[0004] For example, in a multi-round continuous question-answering scenario, the related technology uses reinforcement learning to train its process finalizer (see IM-RAG's public information at SIGIR in July 2024); however, this implementation process requires additional training of a network model, and the operating efficiency needs to be improved.
[0005] The authenticity judgment of the statement content can also be achieved by using a large model reasoning process, such as combining a large model with a zero-shot chain of thought (zero-shot CoT) or a few-shot chain of thought (few-shot CoT), or using the chain of knowledge (CoK); but the technical problem still lies in the efficiency of execution; on the one hand, there are difficulties in integrating large models with knowledge graphs or local databases; at the same time, the applicability of direct reasoning methods and zero-shot chains of thought is difficult to meet retrieval needs. Summary of the invention
[0006] An embodiment of the present invention discloses a data processing method, which uses the deductive reasoning computing power learned by a large model to process a target data set; its core includes a first pre-reasoning step and a second associated reasoning step; its first pre-reasoning step obtains first input information and obtains first reference information and / or second input information; and then obtains first associated information by reasoning according to a preset first model.
[0007] Furthermore, the second association reasoning step obtains second reference information and combines it with the association information of the previous step to further mine data features; the second reference information includes the second input information and the first association information, and the second association information is obtained by reasoning based on the first model.
[0008] Specifically, its first input information may include a first-branch question Q1 and / or a first-branch question A1, and its second input information may include a second-branch question Q2; its second-branch question Q2 and the first input information are asynchronously transmitted to the first model, and the first model asynchronously infers the first input information and the second input information to obtain second related information, and the second related information will be explicitly output or used for further data processing, feature extraction and / or analysis processes.
[0009] Furthermore, the data processing method may also be provided with a third prediction output step; the second associated information is adjusted according to its third reference information to obtain third output information; wherein the second associated information and / or the third output information is used to make a statement judgment and / or generate a thought chain.
[0010] Specifically, the third output information may be expressed using a logical value, which includes a logical true or a logical false; and the logical true may be recorded as true, and the logical false may be recorded as false.
[0011] Furthermore, the first model may adopt a deep learning model, which is used to process a database 001 whose data set size is greater than a preset first lower limit value. The database 001 and its affiliated processing units are used to form the computing power of the above-mentioned reasoning process.
[0012] Among them, the first associated information and the second associated information can adopt a data structure corresponding to a thinking chain, and the thinking chain can adopt a deep learning model and / or a large model, and be inferred in combination with the first reference information and / or the second reference information.
[0013] Specifically, the above reasoning adopts a deductive process, which starts from a specific premise and draws a conclusion through logical deduction, and the logical deduction obtains concrete data through conversion of abstract data.
[0014] Among them, the above-mentioned reasoning can use the knowledge graph to obtain specific data; its first input information and the second input information will asynchronously intervene in the reasoning of the above-mentioned first model or the processing process of the knowledge graph.
[0015] In addition, a search algorithm can be used to generate hyperlink documents to assist the above model and generate a preset return quantity N of first associated information and / or second associated information. The search algorithm returns the top N documents with the highest preset scores, where N is denoted as parameter topk and is a positive integer.
[0016] Specifically, the search algorithm may adopt the BM25 search algorithm and / or the Semantic search algorithm, and the value of the parameter topk, i.e., N, is equal to 3.
[0017] The database 001 may be stored off-site, and the database 001 may receive data access operations via a communication link or a communication system.
[0018] Correspondingly, an embodiment of the present invention also discloses a research and development management device, including a first pre-reasoning unit and a second association reasoning unit; wherein the first pre-reasoning unit is used to obtain first input information and first reference information and / or second input information, and then can infer first association information according to a preset first model; and its second association reasoning unit further obtains second reference information and combines the second input information and the first association information, and then infers the second association information according to the first model, thereby making full use of the information contribution of the first association information to the retrieval process.
[0019] Furthermore, the R&D management device may also be provided with a third prediction output unit, and may adopt the above data processing method to obtain the second associated information and / or adopt the third prediction output unit, and then adjust the second associated information according to the third reference information to obtain the third output information.
[0020] In addition, the R&D management device is not limited to the support of the knowledge graph and the local database; its reasoning can use the knowledge graph or database 001 to obtain specific data, or it can be implemented without the above-mentioned specific data; wherein, the first input information and the second input information asynchronously intervene in the reasoning of the first model or the processing process of the knowledge graph; its database 001 can be stored remotely or accept data access operations via a communication link or communication system.
[0021] In addition, an embodiment of the present invention also discloses a computer storage medium and a testing device; the computer storage medium includes a storage medium body for storing a computer program; when the computer program is executed by a microprocessor, it is used to implement any of the above data processing methods; the testing device includes any of the above R&D management devices and / or computer storage media, which can be used to solve the same technical problems and improve the search accuracy with the help of thinking chain information.
[0022] In summary, the first pre-reasoning step / unit of the present invention obtains the first input information, and obtains the first reference information and / or the second input information, and then obtains the first related information according to the preset first (large) model by reasoning; the second related reasoning step / unit can also obtain the second reference information and combine the second input information and the first related information, and obtain the second related information according to the first (large) model by reasoning; the aforementioned reasoning results can also be converted through the third prediction output step / unit to obtain logical true / false value information; its product adopts a multi-step reasoning process to improve the accuracy of information search, extend AI computing power, combine multi-(two)-step input information, improve data retrieval efficiency, and especially improve the accuracy of judgment and declaration information decision-making.
[0023] It should be noted that the terms "first", "second" and similar terms used in this article are only for describing the various components in the technical solution, and do not constitute a limitation of the technical solution, nor can they be understood as an indication or suggestion of the importance of the corresponding elements; elements with "first", "second" and similar terms indicate that the corresponding technical solution contains at least one of the element. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] In order to more clearly illustrate the technical solution of the present invention and facilitate further understanding of the technical effects, technical features and purposes of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings. The accompanying drawings constitute an essential part of the specification and are used together with the embodiments of the present invention to illustrate the technical solution of the present invention, but do not constitute a limitation to the present invention.
[0025] The same reference numerals in the accompanying drawings represent the same components, and the details of each drawing are as follows.
[0026] Figure 1 The following is an example of the processing flow for declaration judgment in an embodiment of the present invention.
[0027] Figure 2 This is an example of the core processing flow of an embodiment of the present invention in a continuous question-answering scenario.
[0028] Figure 3 This is an example of the processing flow of Liar database (dataset) sample information in an embodiment of the present invention.
[0029] Figure 4 The figure compares the processing effects of the embodiment of the present invention and the direct inference method (461 samples and 3681 samples).
[0030] Figure 5 The following are examples of the results obtained by using the direct answer method for Q1 and A1 in the related art.
[0031] Figure 6 This is an example of Q2 in the related art and the result (A21) obtained by the direct answer method.
[0032] Figure 7 As an example, the result (A22) obtained by combining historical data with the direct answer method in the relevant technology is emphasized on the response to Q1, but no response or deduction is made to Q2.
[0033] Figure 8 This is an example of the reference information (prompt) structure of the embodiment of the present invention.
[0034] Fig. 9 This is an example of the data processing result (A23) (effect) and its first associated information according to an embodiment of the present invention.
[0035] Fig.10 The figure is a schematic diagram of a flow chart of an embodiment of the method of the present invention.
[0036] Fig.11 It is a schematic diagram of the structure of the product embodiment of the present invention.
[0037] Fig.12 This is a schematic diagram of the product layout structure of the product embodiment of the present invention Figure 1 .
[0038] Fig.13 This is a schematic diagram of the product layout structure of the product embodiment of the present invention Figure 2 .
[0039] Fig.14 This is a schematic diagram of the product layout structure of the product embodiment of the present invention Figure 3 .
[0040] Fig.15 This is a schematic diagram of a product usage scenario of an embodiment of the product of the present invention.
[0041] The meanings of the various reference signs / numbers are as follows.
[0042] 001-Database; 009-Communication link or communication system; 011-first input information, such as the first declaration information text_in or text_in1 (can be composed of Q1 and A1); 012-second input information, such as second declaration information text_in2 (which may be composed of Q2); 021-The first question (demand), generated with reference to Q1 and A1, is Question1, which can be written as {Q1, A1}; 022-The second question (demand), generated with reference to Q2, i.e. Question2; 030-Example sample data structure, such as Liar database sample structure; 031-Requirement (task) description information; 033-Prerequisite information (including first premise information, second premise information, etc.); 034- Professional knowledge information (professional data or fields corresponding to the database sample data structure); 035-Conclusion description information; 036-Prediction results; 037-(AI) Reasoning results; 038-(search) qualitative results; 040- F1-score comparison data; 041-461 samples comparison data; 043-3681 sample comparison data; 050- Direct search process (related technology 1); 051- Direct demand information; 053-Direct answer; 055- Directly answer the link information; 060- Local processing flow (related technology 1 continued); 061- Local demand information; 063-local search output (link information); 070-Partial historical information combination process (related technology 2); 071-Multi-step demand information; 073-One-sided historical information; 075-Answer to partial historical information; 092-Examples of multi-step problems; 095-Hyperlink documents, such as a preset number N of documents (results) obtained by the BM25 search algorithm or the Semantic search algorithm; 100-first preliminary reasoning step; 110-first reference information, such as prompt or prompt1 (i.e., first task description information); 111-Key words (part) 199-the first related information, i.e. the first chain of thought (reason); 200-second associative reasoning step; 210 - second reference information, such as prompt2 (including the first input information 011 and the first associated information 199); 211- Potential contact information, i.e., requirements that are potentially related to the previous question; 299-second associated information, i.e., second thought chain (result_long) (explicit expression); 300-third prediction output step; 310-third reference information, such as promp3 (used to convert biased information into false true value prediction information); 399- The third output information, i.e. the prediction result, i.e. result (the final result false and other information); 600- R&D management tools; 610-first pre-reasoning unit; 620-second associative reasoning unit; 630-third prediction output unit; 900- New products or systems to be developed, such as vehicles, smart devices, etc.; 901-Test equipment; 903-Computer storage media; 991- Equipment operators; 999-Other members of the project team. DETAILED DESCRIPTION
[0043] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described below are only for explaining the technical solutions of the present invention, rather than for limiting the present invention. In addition, the parts described in the embodiments or the accompanying drawings are only examples of the relevant parts of the present invention, rather than the whole of the present invention.
[0044] like Figure 1 , Figure 2 , Figure 3 , Fig.10 The data processing method shown includes a first pre-reasoning step 100 and a second associated reasoning step 200; the first pre-reasoning step 100 obtains the first input information 011, and obtains the first reference information 110 and / or the second input information 012, and obtains the first associated information 199 by reasoning according to the preset first model; the second associated reasoning step 200 obtains the second reference information 210, the second reference information 210 includes the second input information 012 and the first associated information 199, and obtains the second associated information 299 by reasoning according to the first model.
[0045] Among them, Figure 1 , Figure 2 As shown, the first input information 011 includes a first-branch question Q1 and / or a first-branch question A1, and the second input information 012 includes a second-branch question Q2. The second-branch question Q2 is asynchronously transmitted to the first model with the first input information 011, and the first model asynchronously infers the first input information 011 and the second input information 012.
[0046] Furthermore, if Fig.10The data processing method shown is also provided with a third prediction output step 300; the second associated information 299 is adjusted according to the third reference information 310 to obtain the third output information 399; the second associated information 299 and / or the third output information 399 are used to make a statement judgment and / or generate a thinking chain; usually, the third output information 399 can be expressed in logical values, and the logical values include logical true or logical false; logical true is recorded as true, and logical false is recorded as false.
[0047] Specifically, its first model adopts a deep learning model, which is used to process a database 001 whose data set size is larger than a preset first lower limit value, that is, to implement a big data deductive reasoning process.
[0048] Among them, the first related information 199 and the second related information 299 adopt a data structure corresponding to a thinking chain, which adopts a deep learning model and / or a large model structure, and is inferred in combination with the first reference information 110 and / or the second reference information 210; the reasoning is achieved through the deductive reasoning computing power learned from big data; the deductive process starts from a specific premise and derives the conclusion through logical deduction, and the logical deduction is converted from abstract data to obtain specific data.
[0049] Among them, the reasoning process can use the knowledge graph to obtain specific data; its first input information 011 and the second input information 012 asynchronously intervene in the reasoning of the first model or the processing process of the knowledge graph.
[0050] Specifically, Fig. 9 As shown, a search algorithm can be used to generate a hyperlink document 095 to assist the model processing process, and generate a preset return quantity of N first associated information 199 and / or second associated information 299. The search algorithm returns the top N documents with the highest preset scores, where N is denoted as parameter topk and N is a positive integer; wherein the search algorithm uses the BM25 algorithm and / or the Semantic algorithm, and the parameter topk value is equal to 3.
[0051] In addition, the database 001 may be stored off-site, and the database 001 may receive data access operations via a communication link or a communication system.
[0052] Accordingly, if Fig.11The R&D management device 600 is also used to solve the same technical problem, and its core includes a first pre-reasoning unit 610 and a second associated reasoning unit 620; the first pre-reasoning unit 610 obtains the first input information 011 and the first reference information 110 and / or the second input information 012, and then obtains the first associated information 199 according to the preset first model reasoning; its second associated reasoning unit 620 obtains the second reference information 210, and its second reference information 210 includes the second input information 012 and the first associated information 199, and then obtains the second associated information 299 according to the above-mentioned first model reasoning process, thereby integrating the effective information of the two-step reasoning process and improving the retrieval accuracy.
[0053] Furthermore, if Fig.11 The R&D management device 600 may also be provided with a third prediction output unit 630; and use any of the above data processing methods to obtain the second associated information 299 and / or use the third prediction output unit 630 to adjust the second associated information 299 according to the third reference information 310 to obtain the third output information 399.
[0054] Among them, in Figures 12 to 15 In the system shown, the reasoning process can use the knowledge graph or database 001 to obtain specific data; Fig.11 The first input information 011 and the second input information 012 shown asynchronously intervene in the reasoning of the first model or the processing process of the knowledge graph; its database 001 can be stored remotely, and the communication link or communication system 009 accepts data access operations.
[0055] In addition, if Figures 12 to 14 The computer storage medium 903 shown includes a storage medium body for storing computer programs; when the computer program is executed by a microprocessor, it is used to implement the above-mentioned data processing method; its test equipment 901 may also include any of the above-mentioned R&D management devices 600 and / or computer storage media 903 to solve the same technical problem.
[0056] In the actual statement judgment task, it is necessary to search the database 001 (including the AI reasoning process) to obtain the required reference information; for example, Figure 1 The processing flow shown; the sample data structure 110 contains some statement information: label is its real label; text is the statement and the background of the person making the statement, location, political party and occasion, etc.
[0057] Specifically, through the processing of the product and method of the embodiment of the present invention, the following can be obtained: Figure 3 The information shown; among them, the big model can infer the conclusion that "the statement and reality are inconsistent", that is, biased, based on its professional knowledge information and identified premises (such as about an organization and an individual).
[0058] Furthermore, the expression "biased" can be converted into a prediction result of "false" through a third qualitative output step (obtaining the third output information 399, that is, obtaining the third prompt).
[0059] Specifically, through testing of the data set (ie, database 001), it was found that compared with related technologies (such as direct inference methods), the F1 scores (F1-score) of the embodiments of the present invention on the test set (461 samples) and the training set (3681 samples) increased by 0.020 and 0.025 respectively.
[0060] Furthermore, if Figure 5-9 As shown, the technical effects achieved by the embodiments of the present invention and related technologies when performing multiple rounds of question and answer are compared, and the effectiveness of their data processing (retrieval) is explained.
[0061] Among them, Figure 5 As shown, the results obtained by the direct answer method for Q1 and A1 in the related technology are given; Figure 6 Q2 in the related art and the result A21 obtained by the direct answer method are given; Figure 7 The result A22 obtained by combining historical data with the direct answer method in the related technology is given; by comparison, it is found that Figure 7 The method focuses on responding to Q1 but not Q2, and the retrieval results cannot meet the needs of actual use.
[0062] Further, Figure 8 The reference information (prompt) data structure of the embodiment of the present invention is given. Fig. 9 The data processing result A23 of the embodiment of the present invention is given; at the same time, its first associated information is also illustrated.
[0063] Specifically, Figure 4 Direct answer to Q2; Figure 5 {Q1, A1} is used as historical information and answered in combination with Q2; the embodiment of the present invention answers in combination with the first associated information.
[0064] like Figure 5 As shown, that is, the Q1 and A1 processing process, the hyperlink part is the searched direct answer document 055; as a reference document, the auxiliary large model generates the above text part, that is, the direct answer 053.
[0065] Specifically, the experimental parameter topk used is "3", that is, the three documents with the highest scores in the results obtained by the BM25 search algorithm are returned; at the same time, the three documents with the highest scores in the results obtained by the Semantic search algorithm are returned.
[0066] Among them, the key used by the search algorithm comes from Q itself; the BM25 algorithm uses the keywords in Q, and the Semantic algorithm uses the semantic encoding of the entire Q.
[0067] Figure 6 As shown, Q2 combines Figure 5 The method directly obtains the A21 result; its search results include documents related to big data, as shown in the local search output 063 in the figure.
[0068] like Figure 7 As shown, the result A22 obtained by the local retrieval method; the search result is more focused on factor 1, that is, the first question Q1, rather than a response to Q2.
[0069] like Figure 8 As shown, the reference information (prompt) structure of the embodiment of the present invention is composed of Fig. 9 It can be seen that the result A23 retrieved by the embodiment of the present invention makes full use of the reference information 093 (prompt), i.e., the thought chain information; the retrieved document takes into account the element 1, communication and system design, and provides a more practical comprehensive output 095 for the demand information (i.e., multiple rounds of questions and answers).
[0070] It is found that the efficiency of data processing can be improved by using R&D management tools 600 or test equipment 901 in development and design: First, the above products can be asked for details of the requirements of the target system (such as element 1), that is, Q1 " Element 2 Product There are requirements for factor 1 storage design "; Then, I found that the answer consisted of some descriptions about the load" Performance requirements: Element 1 storage The overall load of the storage must meet indicator A. You can consider using data processing techniques such as preset method D to achieve this goal. Mark... ”.
[0071] Furthermore, if you want to know whether the current version of the product contains " About the function of realizing indicator C ”, that is, Q2.
[0072] Generally, when a technician (not the physical device corresponding to the embodiment of the present invention, i.e., a human being) thinks (a two-step thinking process), the implicit information (i.e., the thinking chain) includes: "To achieve the preset indicator A, you first need to know the current value of indicator B and then compare it with indicator A." "We need to know the current value of indicator B, and the system needs to be able to achieve indicator C." like Fig. 9 As shown, the embodiment of the present invention expresses (displays and outputs) the first association information 199, that is, the first thought chain 093, through a deductive reasoning process.
[0073] In other words, the final output, A23, includes functions such as real-time load tracking.
[0074] like Figure 6 As shown, in the related technology, the model did not add related information (i.e., thought chain data) to the reference information (i.e., prompt), resulting in the appearance of documents related to big data in the top-ranked search results; therefore, load monitoring during big data collection appeared in the results, which is unrelated to factor 1.
[0075] like Figure 7 As shown, in the related art, historical data is combined with the direct answer method to obtain the result (A22); the result A22 focuses on the response to Q1, but does not respond to Q2; that is, because the model directly puts Q1 and A1 into the reference information (ie prompt), the generated result contains more element 1 demand information, but lacks Q2 itself.
[0076] Among them, putting {Q1, A1} into reference information (i.e. prompt) is the mainstream method of related technologies for multi-round question and answering; obviously, these methods are not applicable in the above scenarios and are difficult to meet actual needs.
[0077] It should be noted that the above embodiments are only for more clearly illustrating the technical solutions of the present invention. Those skilled in the art will understand that the implementation methods of the present invention are not limited to the above contents, and obvious changes, replacements or substitutions based on the above contents do not exceed the scope of the technical solutions of the present invention; other implementation methods will also fall within the scope of the present invention without departing from the concept of the present invention.
Claims
1. A data processing method, characterized in that The method comprises a first pre-reasoning step (100) and a second associated reasoning step (200); wherein: the first pre-reasoning step (100) obtains first input information (011), and obtains first reference information (110) and / or second input information (012), and obtains first associated information (199) by reasoning according to a preset first model; the second associated reasoning step (200) obtains second reference information (210), the second reference information (210) includes the second input information (012) and the first associated information (199), and obtains second associated information (299) by reasoning according to the first model.
2. The data processing method according to claim 1, wherein: The first input information (011) includes a first one-branch question Q1 and / or a first two-branch question A1, the second input information (012) includes a second branch question Q2, the second branch question Q2 and the first input information (011) are asynchronously transmitted to the first model, and the first model asynchronously infers the first input information (011) and the second input information (012).
3. The data processing method as described in claim 1 or 2 further includes a third prediction output step (300); the third prediction output step (300) adjusts the second association information (299) according to the third reference information (310) to obtain third output information (399); the second association information (299) and / or the third output information (399) are used to make statement judgments and / or generate thought chains.
4. The data processing method according to claim 3, wherein: The third output information (399) is expressed in a logical value, wherein the logical value includes logical true or logical false; the logical true is recorded as true, and the logical false is recorded as false.
5. The data processing method according to any one of claims 1, 2 or 4, wherein: The first model includes a deep learning model, which is used to process a database 001 whose data set size is greater than a preset first lower limit value.
6. The data processing method according to claim 5, wherein: The first association information (199) and the second association information (299) use the data structure corresponding to the thought chain, and the thought chain uses the deep learning model and / or the big model and is inferred in combination with the first reference information (110) and / or the second reference information (210).
7. The data processing method according to any one of claims 1, 2, 4 or 6, wherein: The reasoning includes a deductive process, which starts from a specific premise and derives a conclusion through logic, and the logical deduction converts abstract data into concrete data.
8. The data processing method according to claim 7, wherein: The reasoning uses a knowledge graph to obtain the specific data; the first input information (011) and the second input information (012) asynchronously intervene in the reasoning of the first model or the processing process of the knowledge graph.
9. The data processing method according to any one of claims 1, 2, 4, 6 or 8, wherein: A search algorithm is used to generate a hyperlink document (095) to assist the model, and to generate a preset return quantity N of the first associated information (199) and / or the second associated information (299), wherein the search algorithm returns the top N documents with the highest preset scores, where N is denoted as parameter topk and is a positive integer.
10. The data processing method according to claim 9, wherein: The search algorithm includes the BM25 search algorithm and / or the Semantic search algorithm, and the value of the parameter topk, i.e., N, is equal to 3.
11. The data processing method according to claim 10, wherein: The database 001 is stored remotely, and the database 001 receives data access operations via a communication link or a communication system.
12. A research and development management device (600), comprising a first pre-reasoning unit (610) and a second association reasoning unit (620); wherein: The first pre-reasoning unit (610) obtains first input information (011), and obtains first reference information (110) and / or second input information (012), and obtains first associated information (199) by reasoning according to a preset first model; the second associated reasoning unit (620) obtains second reference information (210), the second reference information (210) includes the second input information (012) and the first associated information (199), and obtains second associated information (299) by reasoning according to the first model.
13. The R&D management device (600) as described in claim 12 also includes a third prediction output unit (630); using the data processing method described in claims 1-11 to obtain the second association information (299) and / or using the third prediction output unit (630) to adjust the second association information (299) according to the third reference information (310) to obtain third output information (399).
14. The R&D management device (600) according to claim 12 or 13, wherein: The reasoning uses a knowledge graph or database 001 to obtain specific data; the first input information (011) and the second input information (012) asynchronously intervene in the reasoning of the first model or the processing process of the knowledge graph; the database 001 is stored remotely, and the database 001 accepts data access operations via a communication link or a communication system.
15. A computer storage medium (903), comprising a storage medium body for storing a computer program; when the computer program is executed by a microprocessor, the computer program implements the data processing method according to any one of claims 1 to 11.
16. A testing device (901), comprising the research and development management apparatus (600) according to any one of claims 12 to 14 and / or the computer storage medium (903) according to claim 15.