Information synthesis method and device

Abstract

The invention discloses an information synthesis method and device. The method comprises the steps that a seed problem is decomposed through a preset decomposition model, and meta-information related to the seed problem is generated; constructing a first meta-information relation graph according to a meta-information relation between the meta-information, wherein the meta-information relation is used for indicating an association relation between the meta-information in the same seed question and an association relation between the meta-information in different seed questions; according to the meta-information relation between the meta-information nodes in the first meta-information relation graph, the meta-information nodes are combined to generate meta-information combinations, and different meta-information combinations have different meta-information relations; and generating a synthetic question and a corresponding synthetic answer through a question and answer generation model according to the meta-information combination and the corresponding cue word.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to an information synthesis method and device. Background Art

[0002] Although large language models have demonstrated remarkable capabilities in various language tasks, they still have problems in understanding and solving complex problems (e.g., mathematics and programming). An effective way to address this problem is to use large-scale high-quality synthetic data when large language models understand and solve complex problems. However, developing a low-cost and effective data synthesis method remains a challenge.

[0003] Existing data synthesis methods have three main shortcomings:

[0004] (1) Limited scalability: Existing methods find it difficult to synthesize larger-scale data based on less seed data.

[0005] (2) High cost: Current data synthesis methods rely on the help of commercial models, which leads to a significant increase in synthesis costs.

[0006] (3) Similarity to seed data: Due to the excessive reliance on seed data during the synthesis process, the newly generated data is very similar to the seed data, resulting in insufficient diversity in the generated data.

[0007] The data information synthesis methods in the above-mentioned prior art have technical problems such as poor data scalability, high data synthesis cost, and insufficient diversity of generated data, and no effective solution has been proposed yet. Summary of the invention

[0008] The embodiments of the present application provide an information synthesis method and device to at least solve the technical problems of poor data scalability, high data synthesis cost, and insufficient diversity of generated data in the data information synthesis method existing in the prior art.

[0009] According to one aspect of an embodiment of the present application, there is provided an information synthesis method, comprising: decomposing a seed question through a preset decomposition model to generate meta-information related to the seed question; constructing a first meta-information relationship graph according to the meta-information relationship between the meta-information, wherein the meta-information relationship is used to indicate the association relationship between each meta-information in the same seed question and the association relationship between each meta-information in different seed questions; combining the meta-information nodes according to the meta-information relationship between the meta-information nodes in the first meta-information relationship graph to generate a meta-information combination, wherein different meta-information combinations have different meta-information relationships; and generating a synthesized question and a corresponding synthesized answer according to the meta-information combination and the corresponding prompt words through a question-answer generation model.

[0010] According to another aspect of an embodiment of the present application, an information synthesis device is also provided, including: an information generation module, used to decompose a seed question through a preset decomposition model to generate meta-information related to the seed question; a relationship graph generation module, used to construct a first meta-information relationship graph based on the meta-information relationship between the meta-information, wherein the meta-information relationship is used to indicate the association relationship between each meta-information in the same seed question and the association relationship between each meta-information in different seed questions; a combination generation module, used to combine meta-information nodes according to the meta-information relationship between the meta-information nodes in the first meta-information relationship graph to generate meta-information combinations, wherein different meta-information combinations have different meta-information relationships; and an information synthesis module, used to generate a synthesized question and a corresponding synthesized answer according to the meta-information combination and the corresponding prompt words through a question-answer generation model.

[0011] According to another aspect of an embodiment of the present application, there is also provided an information synthesis device, including: a processor; and a memory, connected to the processor, for providing the processor with instructions for processing the following processing steps: decomposing a seed question through a preset decomposition model to generate meta-information related to the seed question; constructing a first meta-information relationship graph according to the meta-information relationship between the meta-information, wherein the meta-information relationship is used to indicate the association relationship between each meta-information in the same seed question and the association relationship between each meta-information in different seed questions; combining the meta-information nodes according to the meta-information relationship between the meta-information nodes in the first meta-information relationship graph to generate a meta-information combination, wherein different meta-information combinations have different meta-information relationships; and generating a synthesized question and a corresponding synthesized answer according to the meta-information combination and the corresponding prompt words through a question-answer generation model.

[0012] In an embodiment of the present application, the computing device decomposes the seed question, so that the meta information corresponding to the seed question can be obtained in multiple aspects and dimensions as the extended seed data, so that the technical solution realizes the expansion of the scale of the seed data and obtains larger-scale seed data. In addition, the technical solution generates a meta information relationship graph by constructing the relationship between the meta information, and then flexibly combines the various meta information nodes in the meta information relationship graph to generate various types of meta information combinations corresponding to the meta information relationship, and uses the meta information combination to generate corresponding questions and answers, thereby generating more types of seed data, expanding the scale of the seed data, reducing the dependence on the original seed data, and increasing the diversity of the synthesized data. In addition, the data models used in the data synthesis of the technical solution are mostly open source models, which avoids relying on the help of commercial models and reduces the synthesis cost. In addition, the technical problems of poor data scalability, high data synthesis cost, and insufficient diversity of generated data in the data information synthesis method are solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0014] Figure 1 is a hardware structure block diagram of a computing device for implementing the method according to Embodiment 1 of the present application;

[0015] Figure 2 is a flowchart of the information synthesis method according to the first aspect of Example 1 of the present application;

[0016] Figure 3 is a schematic diagram of a first metadata relationship diagram according to Embodiment 1 of the present application;

[0017] Figure 4 is a schematic diagram of a second metadata relationship diagram according to Embodiment 1 of the present application;

[0018] Figure 5 is a schematic diagram of a third metadata relationship diagram according to Example 1 of the present application;

[0019] Figure 6 is a schematic diagram of a fourth element information relationship diagram according to Example 1 of the present application;

[0020] Figure 7 is a sequential flow chart of the information synthesis method according to the first aspect of Example 1 of the present application;

[0021] Figure 8 is a schematic diagram of an information synthesis device according to Embodiment 2 of the present application; and

[0022] Fig. 9 It is a schematic diagram of the information synthesis device described in Example 3 of the present application. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the technical solutions of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only embodiments of a part of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present application.

[0024] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0025] Example 1

[0026] According to this embodiment, a method embodiment of an information synthesis method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0027] The method embodiment provided in this embodiment can be executed in a mobile terminal, a computer terminal, a server or a similar computing device. Figure 1 FIG. 1 shows a hardware structure block diagram of a computing device for implementing the information synthesis method. Figure 1 As shown, the computing device may include one or more processors (the processor may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory for storing data, and a transmission device 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 the I / O interface), a network interface, a power supply and / or a camera. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations shown.

[0028] It should be noted that the one or more processors and / or other data processing circuits described above may generally be referred to herein as "data processing circuitry". The data processing circuitry may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuitry may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the computing device. As described in the embodiments of the present application, the data processing circuitry serves as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0029] The memory can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the information synthesis method in the embodiment of the present application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, the information synthesis method of the above-mentioned application is realized. The memory may include a high-speed random access memory, and may also include a 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 a memory remotely arranged relative to the processor, and these remote memories may be connected to the computing device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0030] The transmission device is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computing device. In one example, the transmission device includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device can be a radio frequency (Radio Frequency, RF) module, which is used to communicate with the Internet wirelessly.

[0031] The display may be, for example, a touch screen liquid crystal display (LCD) that may enable a user to interact with a user interface of the computing device.

[0032] It should be noted that, in some optional embodiments, the above Figure 1 The computing device shown may include hardware elements (including circuits), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware elements and software elements. Figure 1 This is merely one example of a particular embodiment and is intended to illustrate the types of components that may be present in the computing devices described above.

[0033] Under the above operating environment, according to the first aspect of this embodiment, a method for synthesizing information is provided. Figure 1 The computing device implementation shown in . Figure 2 A schematic diagram showing the process of the method is shown in FIG. Figure 3 As shown, the method includes:

[0034] S202: Decomposing the seed problem by using a preset decomposition model to generate meta-information related to the seed problem;

[0035] S204: constructing a first meta information relationship graph according to the meta information relationship between the meta information, wherein the meta information relationship is used to indicate the association relationship between each meta information in the same seed question and the association relationship between each meta information in different seed questions;

[0036] S206: combining the meta information nodes according to the meta information relationships between the meta information nodes in the first meta information relationship graph to generate meta information combinations, wherein different meta information combinations have different meta information relationships; and

[0037] S208: Generate a synthetic question and a corresponding synthetic answer through the question-answer generation model according to the meta-information combination and the corresponding prompt words.

[0038] Specifically, the computing device collects seed data for data synthesis. Taking the question-answering system as an example, the computing device collects seed questions for data synthesis. Then, the computing device inputs the seed questions and corresponding prompt words into a preset decomposition model, decomposes the seed questions through the decomposition model, and outputs meta-information corresponding to the seed questions. The prompt words are pre-set prompts for prompting the decomposition model to extract meta-information from the seed questions. And the decomposition model can be a natural language processing model.

[0039] The prompt corresponding to the decomposition model is: "As a mathematics education expert, please analyze the given math problems and their solutions to extract specific math knowledge points. These knowledge points will help teachers create similar exercises to help students understand and master key learning objectives.

[0040] Please follow the requirements below:

[0041] 1. Extract knowledge points: Identify relevant mathematical knowledge points from a given problem and its solution.

[0042] 2. Ensure relevance: Make sure the extracted knowledge points are directly related to the topic, accurate, and concise, and avoid vague concepts.

[0043] 3. Focus on key concepts: Focus on the specific concepts necessary to solve the problem and explain the solution.

[0044] 4. Provide a clear and concise list: Provide a clear and concise list of knowledge points so that educators can design relevant exercises to help students focus on the key learning outcomes they need to master.

[0045] Mathematical Problem: {Seed Problem}

[0046] Solution: {Corresponding answer}

[0047] Limit the list to no more than ten knowledge points, and ensure that the listed knowledge points are strictly related to solving the given mathematical problem and contribute to its conceptual understanding. Please output in the following format:

[0048] Related mathematical knowledge points: 1. 2. 3.

[0052] 4. (continue listing as needed)

[0053] ”

[0054] For example, the seed problem is a math problem, and the meta-information is the knowledge points involved in the math problem, such as "calculus", "limit" or "derivative". The computing device inputs the math problem (i.e., the seed problem) and the prompt word into a preset decomposition model, decomposes the math problem (i.e., the seed problem) through the decomposition model, and outputs the knowledge points (i.e., the meta-information) corresponding to the math problem (i.e., the seed problem).

[0055] It should be noted that this embodiment only uses knowledge points as meta-information to illustrate the process of data synthesis. In addition, the meta-information can also be a subject (such as "mathematics") or a theme (such as "algebra"), which is not specifically limited here.

[0056] Furthermore, the computing device inputs the decomposed meta-information and the corresponding prompt words into a preset large language model, and filters the meta-information through the large language model, for example, filtering out knowledge points that contain mathematical errors, are incomplete, or are too general, to obtain filtered knowledge points (i.e., meta-information). The computing device then inputs the filtered meta-information into the embedding model, clusters the filtered meta-information through the embedding model, and stores the clustered meta-information in a meta-information library for storing various types of meta-information. For each cluster center, a large language model is used to select the most appropriate meta-information to represent the cluster center. The embedding model may be a bge-en-large model.

[0057] The prompt corresponding to the large language model is: "Given a list of similar mathematical knowledge points, choose one knowledge point that best represents all of them. This knowledge point should be the most commonly known and used in mathematical discussions, and make sure it contains all the listed knowledge points.

[0058] Knowledge point list: {knowledge point list}

[0059] How to choose:

[0060] 1. Common usage: Knowledge points should be widely used in schools and academic scenarios.

[0061] 2. Extensive coverage: All key aspects and details of similar knowledge points listed should be covered.

[0062] 3. Standard Terminology: The terms used should be standard and widely accepted in the mathematical community.

[0063] Please review each topic and select the one that best meets these criteria. Provide a brief explanation of your choice.

[0064] The format is as follows:

[0065] Best knowledge point: {knowledge point of your choice}

[0066] Reason: {Explanation}".

[0067] Furthermore, the computing device selects meta-information that meets the meta-information relationship from the meta-information database according to the preset meta-information relationship between the meta-information, wherein the meta-information relationship includes the association relationship between the meta-information in the same seed question and the association relationship between the meta-information in different seed questions.

[0068] For example, the knowledge points (i.e., meta-information) in seed question 1 include: knowledge point A, knowledge point B and knowledge point C; the knowledge points (i.e., meta-information) in seed question 2 include: knowledge point A and knowledge point D; the knowledge points (i.e., meta-information) in seed question 3 include: knowledge point D and knowledge point E; the knowledge points (i.e., meta-information) in seed question 4 include: knowledge point E and knowledge point F.

[0069] Thus, the computing device associates the knowledge points in the same seed question, such as associating knowledge point A, knowledge point B, and knowledge point C of seed question 1, associating knowledge point A and knowledge point D of seed question 2, associating knowledge point D and knowledge point E of seed question 3, and associating knowledge point E and knowledge point F of seed question 4. The computing device then associates the knowledge points in different seed questions, such as associating knowledge point A of seed question 1 with knowledge point E of seed question 3, associating knowledge point A of seed question 1 with knowledge point F of seed question 4, and so on. Afterwards, the computing device uses the meta-information as the meta-information node for generating a meta-information relationship graph, and then connects the meta-information nodes corresponding to the knowledge points with the association relationship (i.e., meta-information) through connecting edges to construct a graph such as Figure 3The meta information relationship diagram shown (ie, the first meta information relationship diagram).

[0070] Furthermore, the computing device determines the meta-information nodes with the meta-information relationship according to the connection edges in the first meta-information relationship graph. Then, the computing device generates a meta-information combination from the meta-information nodes with the meta-information relationship. For example, if knowledge point A, knowledge point B, and knowledge point C have a meta-information relationship, the computing device combines knowledge point A, knowledge point B, and knowledge point C to generate a meta-information combination. For another example, if knowledge point A and knowledge point E have a meta-information relationship, the computing device combines knowledge point A and knowledge point E to generate a meta-information combination.

[0071] Furthermore, after the computing device generates multiple meta information combinations from the first meta information relationship graph, each meta information combination and the corresponding prompt words for prompting the generation of questions and answers are input into a preset question-answering generation model, and the meta information combination is synthesized and processed by the question-answering generation model to generate a synthesized question and a corresponding synthesized answer. The computing device then stores the synthesized question and the synthesized answer in a database, so as to serve as sample data when training the question-answering model.

[0072] As described in the background, although large language models have demonstrated remarkable capabilities in various language tasks, they still have problems in understanding and solving complex problems (e.g., mathematics and programming). An effective way to solve this problem is for large language models to use large-scale, high-quality synthetic data when understanding and solving complex problems. However, developing a low-cost and effective data synthesis method remains a challenge. Existing data synthesis methods have three main disadvantages: (1) Limited scalability: Existing methods have difficulty synthesizing larger-scale data based on less seed data. (2) High cost: Current data synthesis methods rely on the help of commercial models, resulting in a significant increase in the cost of synthesis. (3) Similarity to seed data: Due to excessive reliance on seed data during the synthesis process, the newly generated data is very similar to the seed data, resulting in insufficient diversity in the generated data.

[0073] In view of the technical problems described above, through the technical solutions of the embodiments of the present application, the computing device decomposes the seed problem, so that the meta information corresponding to the seed problem can be obtained in multiple aspects and dimensions as the extended seed data, so that the technical solution realizes the expansion of the scale of the seed data and obtains a larger scale of seed data. In addition, the technical solution generates a meta information relationship graph by constructing the relationship between the meta information, and then flexibly combines the various meta information nodes in the meta information relationship graph to generate various types of meta information combinations corresponding to the meta information relationship, and uses the meta information combination to generate corresponding questions and answers, thereby generating more types of seed data, expanding the scale of the seed data, reducing the dependence on the original seed data, and increasing the diversity of the synthesized data. In addition, the data models used in the data synthesis of the technical solution are mostly open source models, which avoids relying on the help of the commercial model and reduces the synthesis cost. In addition, the technical problems of poor data scalability, high data synthesis cost, and insufficient diversity of generated data in the data information synthesis method are solved.

[0074] Optionally, the operation of constructing a first meta-information relationship graph based on the meta-information relationship between the meta-information includes: selecting meta-information nodes in the same seed problem to form a single-hop relationship; connecting the meta-information nodes with the single-hop relationship through explicit connection edges to construct a second meta-information relationship graph; and constructing the first meta-information relationship graph based on the second meta-information relationship graph and the meta-information relationship.

[0075] Specifically, the computing device filters the meta-information in the same seed question from the meta-information library, so that the meta-information in the same seed question forms a single-hop relationship between each other. For example, the computing device associates the knowledge point A and the knowledge point B in the seed question 1, so as to form a single-hop relationship; the computing device associates the knowledge point A and the knowledge point C in the seed question 1, so as to form a single-hop relationship; the computing device associates the knowledge point B and the knowledge point C in the seed question 1, so as to form a single-hop relationship. For another example, the computing device associates the knowledge point A and the knowledge point D in the seed question 2, so as to form a single-hop relationship. And so on, so that the computing device forms a single-hop relationship between the knowledge points (i.e., meta-information) in all question seeds.

[0076] Further, refer to Figure 4 As shown, the computing device uses the meta information with a single-hop relationship as the meta information node for generating the second meta information relationship graph, and connects them through explicit connection edges (ie, solid line edges), thereby generating Figure 4 The meta-information relationship graph (ie, the second meta-information relationship graph) is shown. A single-hop relationship is an association relationship formed when two meta-information nodes are connected by an explicit connection edge.

[0077] Furthermore, the computing device constructs the first meta information relationship graph according to the second meta information relationship graph and other meta information relationships.

[0078] Therefore, this technical solution, by clearly defining the single-hop relationship between the meta-information nodes in the same seed problem, accurately captures the direct relationship between these meta-information, and ensures the clarity and accuracy of the meta-information relationship graph. In addition, in this technical solution, the meta-information with a single-hop relationship is the information in the original seed problem, so the correlation between the meta-information is strong, so when data is synthesized based on the meta-information with a single-hop relationship, more core synthesized data can be obtained, making the synthesized data more reasonable.

[0079] Optionally, the operation of constructing the first meta-information relationship graph based on the second meta-information relationship graph and the meta-information relationship includes: screening the second meta-information relationship graph, in which two meta-information nodes connected by a common meta-information node form a double-hop relationship; in the second meta-information relationship graph, connecting the meta-information nodes with the double-hop relationship through the first implicit connection edge to construct a third meta-information relationship graph; and constructing the first meta-information relationship graph based on the third meta-information relationship graph and the meta-information relationship.

[0080] Specifically, after the computing device constructs the second metadata relationship graph, it selects two metadata nodes connected by a common metadata node from the second metadata relationship graph, and associates the two metadata nodes to form a double-hop relationship. Figure 4 In the second meta-information relationship diagram shown in the figure, the computing device screens out from the second meta-information relationship diagram that knowledge point A and knowledge point E are two meta-information nodes connected by a common meta-information node (i.e., knowledge point D), and therefore associates knowledge point A and knowledge point E to form a double-hop relationship. The meta-information that forms the double-hop relationship is the meta-information in different seed questions. Thus, the computing device constructs a double-hop relationship between all meta-information nodes in the second meta-information relationship according to the above-mentioned method for generating a double-hop relationship between meta-information. Thus, the meta-information that forms the double-hop relationship also includes: (1) knowledge point B and knowledge point D; (2) knowledge point C and knowledge point D; (3) knowledge point D and knowledge point E.

[0081] Further, the computing device connects the meta-information nodes with the double-hop relationship of the second meta-information relationship graph through the first implicit connection edge (ie, the long dashed line edge), thereby generating Figure 5 The meta-information relationship diagram shown (i.e., the third meta-information relationship diagram).

[0082] Furthermore, the computing device constructs the first meta-information relationship graph according to the third meta-information relationship graph and other meta-information relationships.

[0083] Therefore, the technical solution forms a double-hop relationship by screening and associating two meta-information nodes connected by a common meta-information node in the second meta-information relationship graph, thereby effectively connecting those indirect but equally important meta-information. And in the technical solution, the double-hop relationship is the closest implicit relationship, which is indirectly connected through a node, so that a certain degree of correlation can be retained and the scale of seed data can be expanded.

[0084] Optionally, the operation of constructing a first meta-information relationship graph based on a third meta-information relationship graph and meta-information relationships includes: determining, from the third meta-information relationship graph, other meta-information nodes connected to the core meta-information node through two common meta-information nodes; constructing a three-hop relationship based on the core meta-information node and the corresponding other meta-information nodes, wherein the core meta-information node is used to indicate the meta-information node whose explicit connection edge is greater than a preset threshold; in the third meta-information relationship graph, connecting the meta-information nodes with a three-hop relationship through a second implicit connection edge to construct a fourth meta-information relationship graph; and constructing the first meta-information relationship graph based on the fourth meta-information relationship graph and the meta-information relationships.

[0085] Specifically, after constructing the third meta-information relationship graph, the computing device selects meta-information nodes whose explicit connection edges are greater than a preset threshold from the third meta-information relationship graph as core meta-information nodes. The core meta-information nodes are usually meta-information nodes with more explicit connection edges connected in the meta-information relationship graph. For example, the preset threshold is 2, so the computing device selects meta-information nodes whose explicit connection edges are greater than 2 from the third meta-information relationship graph as core meta-information nodes. Thus, the computing device selects the core meta-information nodes whose explicit connection edges are greater than 2 from the third meta-information graph as "knowledge point A".

[0086] Afterwards, the computing device determines other meta-information nodes connected to the core meta-information node through two common meta-information nodes from the third meta-information relationship graph, that is, there is an explicit path with 3 connecting edges between the core knowledge point and other knowledge points. For example, the core meta-information node "knowledge point A" and other meta-information nodes "knowledge point F" are connected through two common meta-information nodes (i.e., knowledge point D and knowledge point F), thereby associating the core meta-information node "knowledge point A" and other meta-information nodes "knowledge point F" to form a three-hop relationship. The meta-information that constitutes the three-hop relationship is the meta-information in different seed questions.

[0087] Further, the computing device connects the meta-information nodes with a three-hop relationship in the third meta-information relationship graph through a second implicit connection edge (ie, a short dashed line edge), thereby generating Figure 6 The meta-information relationship diagram shown (i.e., the fourth meta-information relationship diagram).

[0088] Furthermore, the computing device constructs the first meta-information relationship graph according to the fourth meta-information relationship graph and the meta-information relationship.

[0089] Therefore, this technical solution can identify other meta-information nodes connected to the core meta-information node through two common meta-information nodes, thereby expanding the coverage of the meta-information network. And as the distance between knowledge points increases, the correlation will gradually weaken. Therefore, this technical solution will search for core meta-information nodes in the meta-information relationship graph whose explicit connection edges exceed the preset threshold, thereby highlighting the key nodes in the meta-information relationship graph and ensuring the existence of core data for meta-information with three-hop relationships.

[0090] Optionally, the operation of constructing the first meta-information relationship graph based on the fourth meta-information relationship graph and the meta-information relationship includes: screening the meta-information nodes in the fourth meta-information relationship graph, the number of which is equal to three and which are connected in pairs to form a group relationship; and constructing the first meta-information relationship graph based on the fourth meta-information relationship graph and the group relationship.

[0091] Specifically, after the computing device constructs the third metadata relationship graph, it selects metadata nodes that are connected in pairs and have a number of metadata nodes equal to three from the fourth metadata relationship graph, and associates the metadata nodes to form a group relationship. Figure 6 In the fourth meta-information relationship graph shown, the computing device screens out three meta-information nodes of knowledge point A, knowledge point B, and knowledge point C from the fourth meta-information relationship graph, and they are connected in pairs, so knowledge point A, knowledge point B, and knowledge point C are associated to form a group relationship. The meta-information nodes forming the group relationship can be meta-information in different seed questions or in the same seed question.

[0092] Further, the computing device constructs the following according to the fourth meta-information relationship graph and the meta-information combination having the group relationship: Figure 1 The meta information relationship diagram shown (ie, the first meta information relationship diagram).

[0093] Therefore, the meta-information nodes with group relationships in this technical solution have strong correlation with each other. Therefore, when data is synthesized according to the meta-information with group relationships, more core synthesized data can be obtained, making the synthesized data more reasonable.

[0094] Optionally, according to the meta-information relationship between the meta-information nodes in the first meta-information relationship graph, the meta-information nodes are combined to generate the operation of the meta-information combination, including: determining from the first meta-information relationship graph meta-information nodes with a single-hop relationship to combine and generate a single-hop meta-information combination; determining from the first meta-information relationship graph meta-information nodes with a double-hop relationship to combine and generate a double-hop meta-information combination; determining from the first meta-information relationship graph meta-information nodes with a three-hop relationship to combine and generate a three-hop meta-information combination; and determining from the first meta-information relationship graph meta-information nodes with a group relationship to combine and generate a group meta-information combination.

[0095] Specifically, the computing device determines the meta-information nodes with a single-hop relationship, such as knowledge point A and knowledge point B, based on the two meta-information nodes connected by the explicit connection edge in the first meta-information graph. Thus, the computing device combines the meta-information nodes with a single-hop relationship to generate a single-hop meta-information combination. For example, knowledge point A and knowledge point B are combined to generate a single-hop meta-information combination. Thus, the computing device combines all the meta-information and the single-hop in the first meta-information relationship graph according to the above method to generate corresponding multiple single-hop meta-information combinations.

[0096] Further, the computing device determines the meta-information nodes with a double-hop relationship, such as knowledge point A and knowledge point E, based on the meta-information nodes connected by the first implicit connection edge in the first meta-information relationship graph. Thus, the computing device combines the meta-information nodes with a double-hop relationship to generate a double-hop meta-information combination. For example, knowledge point A is combined with knowledge point E to generate a double-hop meta-information combination. Thus, the computing device combines all the meta-information nodes in the first meta-information relationship graph according to the above method to generate a corresponding plurality of double-hop meta-information combinations.

[0097] Further, the computing device determines the meta-information nodes with a three-hop relationship, such as knowledge point A and knowledge point F, based on the meta-information nodes connected by the second implicit connection edge in the first meta-information relationship graph. Thus, the computing device combines the meta-information nodes with a three-hop relationship to generate a three-hop meta-information combination. For example, knowledge point A is combined with knowledge point F to generate a three-hop meta-information combination. Thus, the computing device combines all the meta-information nodes in the first meta-information relationship graph according to the above method to generate corresponding multiple three-hop meta-information combinations.

[0098] Further, the computing device determines the meta-information nodes with a group relationship, such as knowledge point A, knowledge point B, and knowledge point C, based on the three meta-information nodes that are connected in pairs using explicit connection edges in the first meta-information relationship graph. Thus, the computing device combines the meta-information nodes with a group relationship to generate a group meta-information combination. For example, knowledge point A, knowledge point B, and knowledge point C are combined to generate a group meta-information combination. Thus, the computing device combines all the meta-information nodes in the first meta-information relationship graph according to the above method to generate a corresponding plurality of group meta-information combinations.

[0099] Therefore, in this technical solution, single-hop meta-information combination is used to synthesize variant questions of seed data, so more question types can be explored by using single-hop meta-information combination for question synthesis. Double-hop meta-information combination and triple-hop meta-information combination are used to synthesize distributed data, thereby increasing the diversity of synthesized data. Group meta-information combination is used to generate comprehensive questions, thereby enriching the difficulty level of synthesized data. In summary, this technical solution greatly expands the scale of seed data and improves the diversity of seed data by generating various types of meta-information combinations.

[0100] Optionally, the method further includes: determining the number of times that the meta-information nodes connected by each connection edge for connecting the meta-information nodes in the first meta-information relationship graph appear in a seed question; and using the number as the weight of the corresponding connection edge.

[0101] Specifically, after generating the first meta-information relationship graph, the computing device determines the number of times the meta-information nodes connected by each connection edge (ie, including explicit connection edges and implicit connection edges) for connecting the meta-information nodes appear in a seed question.

[0102] For explicit connection edges, for example, knowledge point A and knowledge point B are connected by explicit connection edges, so that the computing device counts the number of times knowledge point A and knowledge point B appear in one seed question from all the collected seed questions. For example, if knowledge point A and knowledge point B appear in 3 seed questions at the same time, the number of times knowledge point A and knowledge point B appear in one seed question is 3. Afterwards, the computing device uses the number of occurrences 3 as the weight of the explicit connection edge connecting knowledge point A and knowledge point B.

[0103] For implicit connection edges, the number of times the meta-information nodes connected by the implicit connection edges appear in a seed question is necessarily 0, so the weights of the implicit connection edges are all 0.

[0104] Therefore, this technical solution intuitively reflects the correlation and co-occurrence frequency between different meta-information nodes by counting the number of times the meta-information nodes connected by each connection edge appear in the same seed question, and using this number as the weight of the connection edge.

[0105] Optionally, the operation of generating a synthetic question and a corresponding synthetic answer according to a meta-information combination and a corresponding prompt word through a question-answering generation model includes: screening a meta-information combination that meets the weight requirement as a candidate meta-information combination according to the weight of the connecting edge; generating a prompt word for indicating the generation of a synthetic question according to the candidate meta-information combination through a prompt word generator; generating a synthetic question according to the candidate meta-information combination and the corresponding prompt word through a question synthesis model; assigning a difficulty level to the synthetic question through a scoring model; and generating a synthetic answer according to the corresponding synthetic question through a large language model corresponding to the difficulty level.

[0106] Specifically, before data synthesis, the computing device determines the weight of each connection edge (i.e., explicit connection edge and implicit connection edge) in the first meta information relationship graph. Then, the computing device selects the meta information combination that meets the weight requirement in the first meta information relationship graph as the candidate meta information combination. For example, if the computing device requires to select the connection edge with a weight of 0, i.e., the implicit connection edge, the computing device selects the meta information combination corresponding to the connection edge with a weight of 0 in the first meta information relationship graph as the candidate meta information combination.

[0107] For another example, the computing device may also require that the weight be an arbitrary natural number, that is, the computing device determines all meta-information combinations in the first meta-information relationship graph, including single-hop meta-information combinations with single-hop relationships, double-hop meta-information combinations with double-hop relationships, three-hop meta-information combinations with three-hop relationships, and group meta-information combinations with group relationships, thereby taking the determined meta-information combinations as candidate meta-information combinations.

[0108] Furthermore, the computing device inputs the candidate meta-information combination into a preset prompt word generator, and the prompt word generator generates prompt words for prompting the generation of synthetic questions. The computing device then inputs the candidate meta-information combination and the corresponding prompt words into the question-answering generation model. The question-answering generation model includes a question synthesis model, a scoring model, and a large language model.

[0109] Thus, the question-answering generation model synthesizes the meta-information in each candidate meta-information combination according to the prompt word through the question synthesis model to generate a corresponding synthetic question. One candidate meta-information combination can generate a synthetic question. The question synthesis model can be a deepseek-math-rl model.

[0110] The prompt corresponding to the problem synthesis model is: "You are a math teacher. Now you need to help your students learn the following math knowledge points. Using these knowledge points as a guide, please construct a new, original math problem that requires understanding and application of all these knowledge points.

[0111] Ensure the following:

[0112] 1. The constructed questions must not contain any mathematical logic errors.

[0113] 2. Questions must combine all knowledge points.

[0114] 3. Questions should be of sufficient difficulty and logically coherent.

[0115] Knowledge point 1: {Knowledge point 1}

[0116] Knowledge point 2: {Knowledge point 2}

[0117] Knowledge point 3: {Knowledge point 3}

[0118] Please provide your answer in the following format:

[0119] New question: {a new question you created}".

[0120] Furthermore, the computing device inputs each synthetic question into a preset scoring model, scores the difficulty of each synthetic question through the scoring model, and assigns a difficulty level to the synthetic question according to the score. For example, a score of 0 to 60 is a low difficulty level, a score of 60 to 80 is a medium difficulty level, and a score of 81 to 100 is a high difficulty level. The scoring model may be a BERT model. The training data used to train the scoring model is math question-and-answer data with difficulty ratings. With the training data as supervision, BERT learns how to rate a math problem as low, medium, or high in difficulty.

[0121] Each difficulty level (low, medium, and high) corresponds to a large language model with a different parameter scale. For example, the large language model corresponding to the low difficulty level has a smaller parameter scale, the large language model corresponding to the high difficulty level has a larger parameter scale, and the large language model corresponding to the medium difficulty level has a parameter scale in between.

[0122] Furthermore, the computing device selects a corresponding large language model according to the difficulty level of each synthetic question, and then inputs the synthetic question into the corresponding large language model, analyzes the synthetic question through the large language model, and generates a corresponding answer (i.e., a synthetic answer).

[0123] The prompt word corresponding to the large language model is:

[0124] {topic}

[0125] Please think about solving this problem step by step and put the final answer in \box{}".

[0126] Therefore, this technical solution can quickly locate the required meta-information combination in a large amount of information by screening the meta-information combination that meets specific conditions as candidate objects based on the weight of the connecting edge, effectively reducing unnecessary computing resource consumption and improving the operating efficiency of the entire system. In addition, this technical solution can generate synthetic problems in multiple dimensions and levels by matching the difficulty level of the synthetic problem with the parameter scale of the large language model.

[0127] In addition, after generating the synthetic questions and synthetic answers, the computing device will also evaluate the synthetic questions and synthetic answers. Specifically, in order to save costs, this technical solution only uses open source evaluation models. In order to achieve the effect of using closed source models, this technical solution uses multiple evaluation models to jointly score to filter data. Experiments show that this strategy can achieve 94% of the effect of closed source models.

[0128] For the evaluation of questions, this technical solution adopts a weighted scoring filtering strategy. Synthetic questions are evaluated based on two criteria: (1) logical completeness (no mathematical errors and accurate correlation with the provided knowledge points); (2) expression completeness (clarity, completeness, and no hints or answers in the question). Each synthetic question is scored by multiple evaluation models, with scores ranging from 0 to 1. The computing device then calculates the weighted score and filters out questions with a weighted score lower than 0.85. For the evaluation of synthetic answers, this technical solution adopts a single-vote veto strategy. The evaluation model requires that the synthetic answer is mathematically correct and fully meets the requirements of the question. Each synthetic answer is scored as 0 or 1 based on the consensus of the model. As long as one evaluation model considers the synthetic answer inappropriate, the synthetic answer will be eliminated.

[0129] The computing device inputs the synthesized question and the corresponding prompt words into the evaluation model, and evaluates the synthesized question through the evaluation model.

[0130] The prompt words corresponding to the evaluation of the question are:

[0131] Given these mathematical facts:

[0132] Knowledge point 1: {Knowledge point 1}

[0133] Knowledge point 2: {Knowledge point 1}

[0134] I have constructed a new math problem as follows: {Newly constructed question}

[0135] Please evaluate whether this new math problem effectively covers all the knowledge points provided and identify whether there are any factual or logical errors in the problem. Please provide a floating point score between 0 and 1, where 1 means that the problem effectively covers both knowledge points and has no factual or logical errors, and 0 means that the problem does not effectively cover any knowledge point or has factual or logical errors.

[0136] Please provide a reasonable score strictly in accordance with the above requirements. The format is as follows:

[0137] Rating: {rating}

[0138] Explanation: {Explanation}".

[0139] In addition, the computing device inputs the synthesized answer and the corresponding prompt word into the evaluation model, and the evaluation model evaluates the synthesized answer. The prompt word corresponding to the evaluation of the answer is:

[0140] You are given a mathematical problem and its solution. Your task is to judge whether the provided solution is correct. Please read the problem and solution carefully and consider the following criteria: "

[0141] 1. Computational accuracy: Check that all numerical calculations are performed correctly.

[0142] 2. Logical consistency: Confirm whether the logical steps are carried out in sequence and correctly.

[0143] 3. Solution completeness: Ensure that all parts of the problem have been addressed and the solution is comprehensive.

[0144] If any of these criteria are not met, you should answer "False".

[0145] Question: {title}

[0146] Solution: {Answer}

[0147] Is the solution provided correct? Please answer in the following format:

[0148] Answer: {Give your judgment, true or false}

[0149] Explanation: {Explanation}".

[0150] Method effect: This technical solution uses the above method to synthesize GSDP-MATH, a mathematical reasoning dataset of 1.91 million questions with detailed solutions, at a very low cost. This dataset covers a variety of data distributions, multiple mathematical fields, and multiple difficulty levels, providing high-quality mathematical training data for large language models.

[0151] This technical solution uses GSDP-MATH to fine-tune on multiple baseline models (Qwen1.5-7B, LLaMA3-8B and Mistral-7B), and verifies the effectiveness of the method on multiple mathematical reasoning indicators (GSM8K, MATH, GAOKAO-Math and SVAMP). Using only GSDP-MATH, our model GSDP-7B (based on Mistral-7B) achieves 78.4% accuracy on GSM8K and 37.7% accuracy on MATH, with an overall performance exceeding the baseline model by 26 points, beating all competitors under the same conditions.

[0152] To further demonstrate the advantages of this method, we compared various data synthesis methods in terms of expansion ratio (the ratio of the final amount of synthesized data to the amount of seed data) and the cost of synthesizing a single piece of data. As shown in Table 1, our method achieves the highest expansion ratio of 255. At the same time, since our method does not use a closed-source model, the cost of synthesizing a single piece of data (unit: 0.01 cents) is 5% or even 1% of other methods.

[0153] Table 1 Comparison of various methods in terms of expansion ratio and synthesis cost.

[0154]

[0155] In summary, reference Figure 7 As shown, the sequential method of the computing device to synthesize data according to the seed question is as follows:

[0156] S701: The computing device decomposes the seed problem through a preset decomposition model to generate meta-information related to the seed problem;

[0157] S702: The computing device selects meta information nodes in the same seed problem to form a single-hop relationship;

[0158] S703: The computing device connects the meta information nodes having a single-hop relationship through explicit connection edges to construct a second meta information relationship graph;

[0159] S704: The computing device screens the second meta information relationship graph, and two meta information nodes connected by a common meta information node form a double-hop relationship;

[0160] S705: The computing device connects the meta information nodes having a double-hop relationship in the second meta information relationship graph through the first implicit connection edge to construct a third meta information relationship graph;

[0161] S706: The computing device determines, from the third meta information relationship graph, other meta information nodes connected to the core meta information node through two common meta information nodes;

[0162] S707: The computing device forms a three-hop relationship according to the core meta information node and the corresponding other meta information nodes;

[0163] S708: The computing device connects the meta information nodes having a three-hop relationship in the third meta information relationship graph through the second implicit connection edge to construct a fourth meta information relationship graph;

[0164] S709: The computing device selects the meta-information nodes in the fourth meta-information relationship graph, in which the number of the meta-information nodes is equal to three and the two are connected to form a group relationship;

[0165] S710: The computing device constructs the first meta information relationship graph according to the fourth meta information relationship graph and the group relationship;

[0166] S711: The computing device determines the number of times that the meta-information nodes connected by each connection edge for connecting the meta-information nodes in the first meta-information relationship graph appear in a seed question, and uses the number as the weight of the corresponding connection edge;

[0167] S712: The computing device determines, from the first meta information relationship graph, meta information nodes having a single-hop relationship and combines them to generate a single-hop meta information combination;

[0168] S713: The computing device determines, from the first meta information relationship graph, meta information nodes having a double-hop relationship and combines them to generate a double-hop meta information combination;

[0169] S714: The computing device determines, from the first meta information relationship graph, meta information nodes having a three-hop relationship and combines them to generate a three-hop meta information combination;

[0170] S715: The computing device determines, from the first meta information relationship graph, meta information nodes having a group relationship and combines them to generate a group meta information combination;

[0171] S716: The computing device selects, according to the weight of the connection edge, a meta information combination that meets the weight requirement as a candidate meta information combination;

[0172] S717: The computing device generates a prompt word for indicating the generation of a synthetic question according to the candidate meta-information combination through a prompt word generator;

[0173] S718: The computing device generates a synthetic question according to the candidate meta-information combination and the corresponding prompt words through the question synthesis model;

[0174] S719: The computing device assigns a difficulty level to the synthetic question by using a scoring model;

[0175] S720: The computing device generates a synthetic answer according to the corresponding synthetic question through a large language model corresponding to the difficulty level.

[0176] Thus, according to this embodiment, the computing device decomposes the seed question, so that the meta information corresponding to the seed question can be obtained in multiple aspects and dimensions as the extended seed data, so that the technical solution realizes the expansion of the scale of the seed data and obtains larger-scale seed data. In addition, the technical solution generates a meta information relationship graph by constructing the relationship between the meta information, and then flexibly combines the various meta information nodes in the meta information relationship graph to generate various types of meta information combinations corresponding to the meta information relationship, and uses the meta information combination to generate corresponding questions and answers, thereby generating more types of seed data, expanding the scale of the seed data, reducing the dependence on the original seed data, and increasing the diversity of the synthesized data. In addition, the data models used in the data synthesis of the technical solution are mostly open source models, which avoids relying on the help of commercial models and reduces the synthesis cost. Thus, the technical problems of poor data scalability, high data synthesis cost, and insufficient diversity of generated data in the data information synthesis method are solved.

[0177] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0178] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0179] Example 2

[0180] Figure 8 The information synthesis device 800 according to this embodiment is shown, and the device 800 corresponds to the method according to the first aspect of embodiment 1. Figure 8As shown, the device 800 includes: an information generation module 810, which is used to decompose the seed question through a preset decomposition model to generate meta-information related to the seed question; a relationship graph generation module 820, which is used to construct a first meta-information relationship graph according to the meta-information relationship between the meta-information, wherein the meta-information relationship is used to indicate the association relationship between each meta-information in the same seed question and the association relationship between each meta-information in different seed questions; a combination generation module 830, which is used to combine the meta-information nodes according to the meta-information relationship between the meta-information nodes in the first meta-information relationship graph to generate a meta-information combination, wherein different meta-information combinations have different meta-information relationships; and an information synthesis module 840, which is used to generate a synthetic question and a corresponding synthetic answer according to the meta-information combination and the corresponding prompt words through a question-answer generation model.

[0181] Optionally, the relationship graph generation module 820 includes: a first construction submodule, used to select meta-information nodes in the same seed problem to form a single-hop relationship; a second construction submodule, used to connect the meta-information nodes with a single-hop relationship through explicit connection edges to construct a second meta-information relationship graph; and a third construction submodule, used to construct a first meta-information relationship graph based on the second meta-information relationship graph and the meta-information relationship.

[0182] Optionally, the third construction submodule includes: a first construction unit, used to screen out, in the second meta-information relationship graph, two meta-information nodes connected by a common meta-information node that form a double-hop relationship; a second construction unit, used to connect the meta-information nodes with a double-hop relationship in the second meta-information relationship graph through a first implicit connection edge to construct a third meta-information relationship graph; and a third construction unit, used to construct the first meta-information relationship graph based on the third meta-information relationship graph and the meta-information relationship.

[0183] Optionally, the operation of constructing a first meta-information relationship graph based on a third meta-information relationship graph and meta-information relationships includes: determining, from the third meta-information relationship graph, other meta-information nodes connected to the core meta-information node through two common meta-information nodes; constructing a three-hop relationship based on the core meta-information node and the corresponding other meta-information nodes, wherein the core meta-information node is used to indicate a meta-information node whose explicit connection edge is greater than a preset threshold; in the third meta-information relationship graph, connecting the meta-information nodes with a three-hop relationship through a second implicit connection edge to construct a fourth meta-information relationship graph; and constructing the first meta-information relationship graph based on the fourth meta-information relationship graph and the meta-information relationships.

[0184] Optionally, the operation of constructing the first meta-information relationship graph based on the fourth meta-information relationship graph and the meta-information relationship includes: screening the meta-information nodes in the fourth meta-information relationship graph, the number of which is equal to three and which are connected in pairs to form a group relationship; and constructing the first meta-information relationship graph based on the fourth meta-information relationship graph and the group relationship.

[0185] Optionally, according to the meta-information relationship between the meta-information nodes in the first meta-information relationship graph, the meta-information nodes are combined to generate the operation of the meta-information combination, including: determining from the first meta-information relationship graph meta-information nodes with a single-hop relationship to combine and generate a single-hop meta-information combination; determining from the first meta-information relationship graph meta-information nodes with a double-hop relationship to combine and generate a double-hop meta-information combination; determining from the first meta-information relationship graph meta-information nodes with a three-hop relationship to combine and generate a three-hop meta-information combination; and determining from the first meta-information relationship graph meta-information nodes with a group relationship to combine and generate a group meta-information combination.

[0186] Optionally, the device 800 also includes: a first determination module, used to determine the number of times that the meta-information nodes connected by each connection edge used to connect the meta-information nodes in the first meta-information relationship graph appear in a seed question; and a second determination module, used to use the number of times as the weight of the corresponding connection edge.

[0187] Optionally, the information synthesis module 840 includes: a first generation submodule, used to screen meta-information combinations that meet the weight requirements as candidate meta-information combinations according to the weights of the connecting edges; a second generation submodule, used to generate prompt words for indicating the generation of synthetic questions according to the candidate meta-information combinations through a prompt word generator; a third generation submodule, used to generate synthetic questions according to the candidate meta-information combinations and corresponding prompt words through a question synthesis model; a fourth generation submodule, used to assign difficulty levels to synthetic questions through a scoring model; and a fifth generation submodule, used to generate synthetic answers according to corresponding synthetic questions through a large language model corresponding to the difficulty level.

[0188] Thus, according to this embodiment, the computing device decomposes the seed question, so that the meta information corresponding to the seed question can be obtained in multiple aspects and dimensions as the extended seed data, so that the technical solution realizes the expansion of the scale of the seed data and obtains larger-scale seed data. In addition, the technical solution generates a meta information relationship graph by constructing the relationship between the meta information, and then flexibly combines the various meta information nodes in the meta information relationship graph to generate various types of meta information combinations corresponding to the meta information relationship, and uses the meta information combination to generate corresponding questions and answers, thereby generating more types of seed data, expanding the scale of the seed data, reducing the dependence on the original seed data, and increasing the diversity of the synthesized data. In addition, the data models used in the data synthesis of the technical solution are mostly open source models, which avoids relying on the help of commercial models and reduces the synthesis cost. Thus, the technical problems of poor data scalability, high data synthesis cost, and insufficient diversity of generated data in the data information synthesis method are solved.

[0189] Example 3

[0190] Fig. 9 The information synthesis device 900 according to this embodiment is shown, and the device 900 corresponds to the method according to the first aspect of embodiment 1. Fig. 9 As shown, the device 900 includes: a processor 910; and a memory 920, which is connected to the processor 910 and is used to provide instructions for the processor 910 to process the following processing steps: decomposing the seed question through a preset decomposition model to generate meta-information related to the seed question; constructing a first meta-information relationship graph according to the meta-information relationship between the meta-information, wherein the meta-information relationship is used to indicate the association relationship between each meta-information in the same seed question and the association relationship between each meta-information in different seed questions; combining the meta-information nodes according to the meta-information relationship between the meta-information nodes in the first meta-information relationship graph to generate a meta-information combination, wherein different meta-information combinations have different meta-information relationships; and generating a synthetic question and a corresponding synthetic answer according to the meta-information combination and the corresponding prompt words through a question-answer generation model.

[0191] Optionally, the operation of constructing a first meta-information relationship graph based on the meta-information relationship between the meta-information includes: selecting meta-information nodes in the same seed problem to form a single-hop relationship; connecting the meta-information nodes with the single-hop relationship through explicit connection edges to construct a second meta-information relationship graph; and constructing the first meta-information relationship graph based on the second meta-information relationship graph and the meta-information relationship.

[0192] Optionally, the operation of constructing the first meta-information relationship graph based on the second meta-information relationship graph and the meta-information relationship includes: screening the second meta-information relationship graph, in which two meta-information nodes connected by a common meta-information node form a double-hop relationship; in the second meta-information relationship graph, connecting the meta-information nodes with the double-hop relationship through the first implicit connection edge to construct a third meta-information relationship graph; and constructing the first meta-information relationship graph based on the third meta-information relationship graph and the meta-information relationship.

[0193] Optionally, the operation of constructing a first meta-information relationship graph based on a third meta-information relationship graph and meta-information relationships includes: determining, from the third meta-information relationship graph, other meta-information nodes connected to the core meta-information node through two common meta-information nodes; constructing a three-hop relationship based on the core meta-information node and the corresponding other meta-information nodes, wherein the core meta-information node is used to indicate a meta-information node whose explicit connection edge is greater than a preset threshold; in the third meta-information relationship graph, connecting the meta-information nodes with a three-hop relationship through a second implicit connection edge to construct a fourth meta-information relationship graph; and constructing the first meta-information relationship graph based on the fourth meta-information relationship graph and the meta-information relationships.

[0194] Optionally, the operation of constructing the first meta-information relationship graph based on the fourth meta-information relationship graph and the meta-information relationship includes: screening the meta-information nodes in the fourth meta-information relationship graph, the number of which is equal to three and which are connected in pairs to form a group relationship; and constructing the first meta-information relationship graph based on the fourth meta-information relationship graph and the group relationship.

[0195] Optionally, according to the meta-information relationship between the meta-information nodes in the first meta-information relationship graph, the meta-information nodes are combined to generate the operation of the meta-information combination, including: determining from the first meta-information relationship graph meta-information nodes with a single-hop relationship to combine and generate a single-hop meta-information combination; determining from the first meta-information relationship graph meta-information nodes with a double-hop relationship to combine and generate a double-hop meta-information combination; determining from the first meta-information relationship graph meta-information nodes with a three-hop relationship to combine and generate a three-hop meta-information combination; and determining from the first meta-information relationship graph meta-information nodes with a group relationship to combine and generate a group meta-information combination.

[0196] Optionally, the memory 920 is also used to provide instructions for the processor 910 to process the following processing steps: determining the number of times that the meta-information nodes connected by each connection edge used to connect the meta-information nodes in the first meta-information relationship graph appear in a seed question; and using the number as the weight of the corresponding connection edge.

[0197] Optionally, the operation of generating a synthetic question and a corresponding synthetic answer according to a meta-information combination and a corresponding prompt word through a question-answering generation model includes: screening a meta-information combination that meets the weight requirement as a candidate meta-information combination according to the weight of the connecting edge; generating a prompt word for indicating the generation of a synthetic question according to the candidate meta-information combination through a prompt word generator; generating a synthetic question according to the candidate meta-information combination and the corresponding prompt word through a question synthesis model; assigning a difficulty level to the synthetic question through a scoring model; and generating a synthetic answer according to the corresponding synthetic question through a large language model corresponding to the difficulty level.

[0198] Thus, according to this embodiment, the computing device decomposes the seed question, so that the meta information corresponding to the seed question can be obtained in multiple aspects and dimensions as the extended seed data, so that the technical solution realizes the expansion of the scale of the seed data and obtains larger-scale seed data. In addition, the technical solution generates a meta information relationship graph by constructing the relationship between the meta information, and then flexibly combines the various meta information nodes in the meta information relationship graph to generate various types of meta information combinations corresponding to the meta information relationship, and uses the meta information combination to generate corresponding questions and answers, thereby generating more types of seed data, expanding the scale of the seed data, reducing the dependence on the original seed data, and increasing the diversity of the synthesized data. In addition, the data models used in the data synthesis of the technical solution are mostly open source models, which avoids relying on the help of commercial models and reduces the synthesis cost. Thus, the technical problems of poor data scalability, high data synthesis cost, and insufficient diversity of generated data in the data information synthesis method are solved.

[0199] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0200] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0201] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0202] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0203] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0204] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, 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, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, RandomAccess Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0205] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An information synthesis method, characterized in that: include: Decomposing the seed problem by using a preset decomposition model to generate meta-information related to the seed problem; Constructing a first meta information relationship graph according to the meta information relationship between the meta information, wherein the meta information relationship is used to indicate the association relationship between each meta information in the same seed question and the association relationship between each meta information in different seed questions; According to the meta information relationship between the meta information nodes in the first meta information relationship graph, the meta information nodes are combined to generate meta information combinations, wherein different meta information combinations have different meta information relationships; as well as A synthetic question and a corresponding synthetic answer are generated by a question-answering generation model according to the meta-information combination and the corresponding prompt words.

2. The method according to claim 1, characterized in that The operation of constructing a first meta information relationship graph according to the meta information relationship between the meta information includes: Select the meta-information nodes in the same seed problem to form a single-hop relationship; Connecting the meta information nodes having a single-hop relationship through explicit connection edges to construct a second meta information relationship graph; and The first meta information relationship graph is constructed according to the second meta information relationship graph and the meta information relationship.

3. The method according to claim 2, characterized in that The operation of constructing the first meta information relationship graph according to the second meta information relationship graph and the meta information relationship includes: In the second meta-information relationship graph, two meta-information nodes connected by a common meta-information node form a double-hop relationship; In the second meta-information relationship graph, the meta-information nodes having a double-hop relationship are connected through a first implicit connection edge to construct a third meta-information relationship graph; and The first meta-information relationship graph is constructed according to the third meta-information relationship graph and the meta-information relationship.

4. The method according to claim 3, characterized in that The operation of constructing the first meta information relationship graph according to the third meta information relationship graph and the meta information relationship includes: Determine, from the third meta information relationship graph, other meta information nodes connected to the core meta information node via two common meta information nodes; A three-hop relationship is formed according to the core meta-information node and corresponding other meta-information nodes, wherein the core meta-information node is used to indicate the meta-information node whose explicit connection edge is greater than a preset threshold; In the third metadata relationship graph, metadata nodes having a three-hop relationship are connected through second implicit connection edges to construct a fourth metadata relationship graph; and The first meta-information relationship graph is constructed according to the fourth meta-information relationship graph and the meta-information relationship.

5. The method according to claim 4, characterized in that The operation of constructing the first meta information relationship graph according to the fourth meta information relationship graph and the meta information relationship includes: Filtering the meta-information nodes connected in pairs, in which the number of the meta-information nodes is equal to three, in the fourth meta-information relationship graph to form a group relationship; and The first meta information relationship graph is constructed according to the fourth meta information relationship graph and the group relationship.

6. The method according to claim 5, characterized in that The operation of combining the meta information nodes according to the meta information relationship between the meta information nodes in the first meta information relationship graph to generate a meta information combination includes: Determine the meta information nodes having the single-hop relationship from the first meta information relationship graph and combine them to generate a single-hop meta information combination; Determine the meta information nodes having the double-hop relationship from the first meta information relationship graph and combine them to generate a double-hop meta information combination; Determine the meta information nodes having the three-hop relationship from the first meta information relationship graph and combine them to generate a three-hop meta information combination; and Meta information nodes having the group relationship are determined from the first meta information relationship graph and combined to generate a group meta information combination.

7. The method according to claim 1, characterized in that Also includes: Determine the number of times the meta-information nodes connected by the connection edges for connecting the meta-information nodes in the first meta-information relationship graph appear in a seed question; as well as The number is used as the weight of the corresponding connecting edge.

8. The method according to claim 7, characterized in that The operation of generating a synthetic question and a corresponding synthetic answer according to the meta-information combination and the corresponding prompt words by the question-answer generation model includes: According to the weight of the connection edge, selecting a meta information combination that meets the weight requirement as a candidate meta information combination; Generate a prompt word for indicating generation of the synthetic question according to the candidate meta-information combination by a prompt word generator; Generate the synthetic question according to the candidate meta-information combination and the corresponding prompt words through the question synthesis model; assigning a difficulty level to the synthetic question by a scoring model; and The synthetic answer is generated according to the corresponding synthetic question through a large language model corresponding to the difficulty level.

9. An information synthesis device, characterized in that: include: An information generation module, used to decompose the seed problem through a preset decomposition model to generate meta-information related to the seed problem; A relationship graph generating module, used for constructing a first meta information relationship graph according to the meta information relationship between the meta information, wherein the meta information relationship is used for indicating the association relationship between each meta information in the same seed question and the association relationship between each meta information in different seed questions; a combination generating module, used for combining the meta information nodes according to the meta information relationship between the meta information nodes in the first meta information relationship graph to generate a meta information combination, wherein different meta information combinations have different meta information relationships; as well as The information synthesis module is used to generate a synthetic question and a corresponding synthetic answer according to the meta-information combination and the corresponding prompt words through a question-answer generation model.

10. An information synthesis device, characterized in that: include: processor; as well as A memory, connected to the processor, configured to provide the processor with instructions for processing the following processing steps: Decomposing the seed problem by using a preset decomposition model to generate meta-information related to the seed problem; Constructing a first meta information relationship graph according to the meta information relationship between the meta information, wherein the meta information relationship is used to indicate the association relationship between each meta information in the same seed question and the association relationship between each meta information in different seed questions; According to the meta information relationship between the meta information nodes in the first meta information relationship graph, the meta information nodes are combined to generate meta information combinations, wherein different meta information combinations have different meta information relationships; as well as A synthetic question and a corresponding synthetic answer are generated by a question-answering generation model according to the meta-information combination and the corresponding prompt words.

Images