Information reasoning method and system and electronic equipment

Abstract

The invention discloses an information reasoning method and system and electronic equipment, and relates to the field of large model technology and natural language processing. The method comprises the following steps: detecting query information to be reasoned; the query information is decomposed to obtain sub-query information of different reasoning stages, and the sub-query information of the different reasoning stages has a logic association relationship; determining sub-reply information matched with the sub-query information of the different reasoning stages; the sub-reply information matched with the sub-query information is verified, a verification result corresponding to the sub-query information is obtained, and the verification result is used for representing the feasible degree of the sub-query information to the reasoning query information; and in response to the verification result being greater than the verification result threshold, reasoning the sub-query information corresponding to the verification result to obtain reply information corresponding to the query information. The technical problem of low efficiency of information reasoning is solved.

Description

Technical Field

[0001] This application relates to the field of large model technology and natural language processing. Specifically, it relates to an information reasoning method, system, and electronic device. Background Art

[0002] In the related art, in the complex application scenarios of logical reasoning and solving mathematical problems, especially in problems involving many abstract concepts or multi-step reasoning, due to the lack of certain cognitive abilities in current information reasoning, it is still very difficult to solve complex logical reasoning problems (Logical Reasoning) or mathematical problems (Math Problem), and the time spent on solving is relatively long. Therefore, there is still the technical problem of low efficiency in information reasoning.

[0003] For the above technical problems, no effective solution has been proposed yet. Summary of the Invention

[0004] Embodiments of this application provide an information reasoning method, system, and electronic device to at least solve the technical problem of low efficiency in information reasoning.

[0005] According to one aspect of the embodiments of this application, an information reasoning method is provided. The method may include the following steps: detecting query information to be reasoned; decomposing the query information to obtain sub-query information at different reasoning stages, where there is a logical association relationship between the sub-query information at different reasoning stages; determining sub-response information respectively matching the sub-query information at different reasoning stages; verifying the sub-response information matching the sub-query information to obtain a verification result corresponding to the sub-query information, where the verification result is used to represent the feasibility degree of the sub-query information for reasoning the query information; and in response to the verification result being greater than the verification result threshold, reasoning the sub-query information corresponding to the verification result to obtain a response information corresponding to the query information.

[0006] According to another aspect of the embodiments of this application, another information reasoning method is further provided. The method may include the following steps: in response to query information to be reasoned received in a dialogue interface, displaying, on the dialogue interface, the sub-query information at different reasoning stages of the query information, where there is a logical association relationship between the sub-query information at different reasoning stages; displaying, on the dialogue interface, the sub-response information respectively matching the sub-query information at different reasoning stages; and displaying, on the dialogue interface, the response information corresponding to the query information, where the response information is obtained by reasoning the sub-query information corresponding to the verification result in the case that the verification result corresponding to the sub-query information is greater than the verification result threshold, the verification result is obtained by verifying the sub-response information matching the sub-query information, and is used to represent the feasibility degree of the sub-query information for reasoning the query information.

[0007] According to another aspect of the embodiments of the present application, there is also provided a method for generating a model. The method may include: obtaining a query information sample; training a first large model by using sub-query information samples corresponding to the query information sample, where the first large model is used to decompose the input query information to obtain sub-query information at different inference stages, and there is a logical association relationship between the sub-query information at different inference stages; obtaining a reply information sample corresponding to the query information sample; obtaining a verification result sample corresponding to the reply information sample, where the verification result sample is used to at least represent the classification result of the reply information sample; training a second large model by using the verification result sample, where the second large model is used to verify the sub-reply information matched with the sub-query information to obtain a verification result corresponding to the sub-query information, and the verification result is used to characterize the feasibility degree of the sub-query information for inferring the query information, and the sub-query information corresponding to the verification result greater than the verification result threshold is used to infer the reply information corresponding to the query information.

[0008] According to another aspect of the embodiments of the present application, there is also provided an information inference system. The system may include: an information generation end, configured to detect the query information to be inferred, decompose the query information to obtain sub-query information at different inference stages, and determine sub-reply information respectively matched with the sub-query information at different inference stages, where there is a logical association relationship between the sub-query information at different inference stages; an information verification end, configured to verify the sub-reply information matched with the sub-query information to obtain a verification result corresponding to the sub-query information, and in response to the verification result being greater than the verification result threshold, infer the reply information corresponding to the query information from the sub-query information corresponding to the verification result, where the verification result is used to characterize the feasibility degree of the sub-query information for inferring the query information.

[0009] According to another aspect of the embodiments of the present application, there is also provided an electronic device. The electronic device may include a memory and a processor: the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the information inference method of any one of the above is implemented.

[0010] According to another aspect of the embodiments of the present application, there is also provided a processor. The processor is used to run a program, and when the program runs, the information inference method of any one of the above is executed.

[0011] According to another aspect of the embodiments of the present application, there is also provided a computer-readable storage medium. The computer-readable storage medium includes a stored program, and when the program runs, it controls the device where the storage medium is located to execute the information inference method of any one of the above.

[0012] In an embodiment of the present application, after detecting the query information, the query information can be decomposed to determine sub-query information with logical association relationships among different inference stages. And each sub-query information can be replied to obtain corresponding sub-reply information. To ensure the feasibility of the determined reply information, each sub-reply information can be verified, and the verification results corresponding to each sub-reply information can be obtained. By determining whether the verification result is greater than the verification result threshold, it is determined whether the query information is feasible. If the verification result is greater than the verification result threshold, it indicates that the query information has a high degree of feasibility, and the sub-query information can be inferred to determine the reply information of the query information. Since it is considered that in the related art, due to the lack of cognitive ability of the large model, the efficiency of reasoning logical problems with logical relationships is low, in the embodiment of the present application, the original query information can be divided into multiple sub-query information with logical association relationships according to the sequence of inference stages. By replying to the sub-query information in different inference stages, the reply information of the query information can be finally determined, thereby greatly improving the accuracy of solving problems such as logical reasoning, and then achieving the technical effect of improving the efficiency of information inference and solving the technical problem of low efficiency of information inference.

[0013] It is easy to note that the above general description and the following detailed description are only for exemplifying and explaining the present application, and do not constitute a limitation to the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The drawings described herein are used to provide a further understanding of the present application, and constitute a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation to the present application. In the drawings:

[0015] Figure 1 is a schematic diagram of an application scenario of an information inference method according to an embodiment of the present application;

[0016] Figure 2 is a flowchart of an information inference method according to an embodiment of the present application;

[0017] Figure 3 is a flowchart of another information inference method according to an embodiment of the present application;

[0018] Figure 4 is a flowchart of a method for generating a model according to an embodiment of the present application;

[0019] Figure 5 is a schematic diagram of an information inference system according to an embodiment of the present application;

[0020] Figure 6 is a schematic diagram of a cognitive tree according to an embodiment of the present application;

[0021] Figure 7 It is a schematic diagram of a complex task reasoning process according to an embodiment of the present application;

[0022] Figure 8 It is a schematic diagram of an information reasoning device according to an embodiment of the present application;

[0023] Figure 9 It is a schematic diagram of another information reasoning device according to an embodiment of the present application;

[0024] Figure 10 It is a schematic diagram of a model generation device according to an embodiment of the present application;

[0025] Figure 11 It is a structural block diagram of a computer terminal according to an embodiment of the present application;

[0026] Figure 12 It is a block diagram of an electronic device for an information reasoning method according to an embodiment of the present application. Detailed implementation manners

[0027] In order to enable those skilled in the art to better understand the 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 accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present application.

[0028] 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 do not have to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present application described here can be implemented in an order different from those illustrated or described here. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0029] The technical solution provided by this application is mainly implemented using large model technology. Here, a large model refers to a deep learning model with a large number of model parameters, usually containing hundreds of millions, tens of billions, hundreds of billions, trillions, or even more than one quadrillion model parameters. A large model can also be called a Foundation Model. Through pre-training of the large model with a large amount of unlabeled corpus, a pre-trained model with over hundreds of millions of parameters is produced. This model can adapt to a wide range of downstream tasks and has good generalization ability. For example, large language models (LLMs), multi-modal pre-training models, etc.

[0030] It should be noted that when a large model is actually applied, the pre-trained model can be fine-tuned with a small number of samples so that the large model can be applied to different tasks. For example, large models can be widely applied in fields such as Natural Language Processing (NLP), computer vision, and speech processing. Specifically, they can be applied to tasks in the field of computer vision such as Visual Question Answering (VQA), Image Caption (IC), and image generation. They can also be widely applied to tasks in the field of natural language processing such as text-based sentiment classification, text summary generation, and machine translation. Therefore, the main application scenarios of large models include but are not limited to digital assistants, intelligent robots, search, online education, office software, e-commerce, intelligent design, etc.

[0031] First, some nouns or terms that appear during the description of the embodiments of this application are applicable to the following explanations:

[0032] Large language model: A language model that can output corresponding response text through natural language input.

[0033] Embodiment 1

[0034] According to the embodiments of this application, an information reasoning 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 the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0035] Considering that the number of model parameters of the large model is huge and the computing resources of mobile terminals are limited, the above-mentioned information reasoning method provided by the embodiments of this application can be applied to the application scenarios as Figure 1 shown, but not limited to this. In such as Figure 1In the application scenario shown, the large model is deployed in server 10, which can be a cloud server. Server 10 can be connected to one or more client devices 20 through a local area network connection, a wide area network connection, an Internet connection, or other types of data networks. Here, client devices 20 can include, but are not limited to: smartphones, tablets, laptops, palmtop computers, personal computers, smart home devices, in-vehicle devices, etc. The client devices together constitute the client relative to the server. An interaction interface for obtaining application generation instructions can be deployed on the graphical user interface of the client device, and this interaction interface can be a dialogue interface. Client device 20 can interact with the user through the graphical user interface to implement the invocation of the large model, and further implement the information inference method provided by the embodiments of the present application.

[0036] In the embodiments of the present application, the system composed of the client device and the server can perform the following steps: When the client needs to query the large model to answer a question, it can input the query information corresponding to the question to be answered on the interaction interface of its client device. The client device can collect the query information and send it to the server through the network. After receiving the query information, the server can perform the following steps: Step S102, detecting the query information to be inferred; Step S104, decomposing the query information to obtain sub-query information at different inference stages, where there is a logical association relationship between the sub-query information at different inference stages; Step S106, determining the sub-answer information that matches the sub-query information at different inference stages respectively; Step S108, verifying the sub-answer information that matches the sub-query information to obtain the verification result corresponding to the sub-query information, where the verification result is used to represent the feasibility degree of the sub-query information for inferring the query information; Step S110, in response to the verification result being greater than the verification result threshold, inferring the sub-query information corresponding to the verification result to obtain the answer information corresponding to the query information. The answer information can be output to the client device.

[0037] In the above process, the sub-query information determined in the server, the sub-answer information corresponding to the sub-query information, and the verification result obtained by verification, etc. can also be sent to the corresponding client device through the network. The sub-query information, sub-answer information, and verification result, etc. in the information inference process can be displayed on the interaction interface of the client device. And it can be adjusted on the interaction interface according to its own needs and accuracy to ensure the accuracy of the final required answer information. It can also be transmitted to the corresponding client device through the network after the server determines the answer information, and after the client device receives the answer information, the answer information can be displayed on the interaction interface.

[0038] Since it is considered that in the related art, due to the lack of cognitive ability of large models, the efficiency of reasoning about logical problems with logical relationships is low. In the embodiments of the present application, the original query information can be divided into multiple sub-query information with logical association relationships according to the sequence of the reasoning stages. By answering the sub-query information in different reasoning stages, the answer information of the query information can be finally determined, thereby greatly improving the accuracy of the large model in solving problems such as logical reasoning, and further achieving the technical effect of improving the efficiency of information reasoning, and solving the technical problem of low efficiency of information reasoning. It should be noted that when the operating resources of the client device can meet the deployment and operation conditions of the large model, the embodiments of the present application can be carried out in the client device.

[0039] The embodiments of the present application propose the following method from the technical implementation side. In the above operating environment, Embodiment 1 of the present application provides an Figure 2 information reasoning method as shown. Figure 2 It is a flowchart of an information reasoning method according to the present application. As Figure 2 shown, the method may include the following steps:

[0040] Step S202, detect the query information to be reasoned.

[0041] In the technical solution provided in step S202 of the present application above, when the client has a problem that needs to query the large model for an answer, the problem to be answered can be formed into query information and input into the interaction interface of the client. When it is detected that query information is input on the interaction interface, the query information collected on the interaction interface can be transmitted to the server through the network. Whether the query information to be reasoned is received can be detected by the server. Among them, the query information can be the original problem (original problem) that the user of the client needs the large model to answer, and can also be called the initial query. The original problem can be the query question Question (Q) input by the user into the interaction interface. The query Q can be a problem that requires the large model to use its own cognitive ability to reason and answer. For example, it can be a relatively complex task such as a math problem (Math Problem) or a logical reasoning problem. It should be noted that the above query information is only for illustrative purposes and is not specifically limited here. As long as it is a problem that requires the large model to use cognitive ability to reason and analyze for an answer, it is within the protection scope of the embodiments of the present application.

[0042] Optionally, when a user in the client has a question to ask the large model, the user can input query information that can describe the queried question on the interaction interface of the client device. For example, natural language that can describe the relatively complex logical reasoning question can be input. The input of the query information can be confirmed by clicking the corresponding control on the interaction interface. At this time, the interaction interface can send the collected query information to the server through the network. Since the large model capable of reasoning about the query information is deployed in the server. Therefore, in the server, the large model can be used to process the query information. It should be noted that the language describing the query information above is only for illustrative purposes and is not specifically limited here.

[0043] Optionally, before reasoning about the query information, a large model with strong cognitive ability, capable of strong logical reasoning and analysis ability, can be pre-trained and pre-deployed in the server to provide a question-and-answer service for the client associated with the server. That is, through the server where the large model is deployed, a service for solving problems such as logical reasoning and data can be provided for users.

[0044] Optionally, after the server receives the query information transmitted by the network, the pre-deployed large model can be enabled to reason about and analyze the query information.

[0045] Step S204: Decompose the query information to obtain sub-query information at different reasoning stages, where there is a logical association relationship between the sub-query information at different reasoning stages.

[0046] In the technical solution provided in step S204 of the present application above, after detecting the query information to be inferred, the query information can be decomposed to obtain sub-query information at different inference stages. Among them, there is a certain logical association relationship between the sub-query information at different inference stages. Different inference stages can be used to represent the solution process when solving the problem of the query information. For example, if two inference stages are required to solve the query information, it can include the first inference stage (Step1) and the second inference stage (Step2). The inference stage can be an inference chain (Chain-of-thought), decomposition set, or generated result obtained by analyzing the query information. The sub-query information can be used to represent the phased analysis result obtained by reasoning in the corresponding inference stage. For example, it can be the decomposition hypothesis, decomposition D, or smaller problem obtained by decomposition. Each inference stage can include at least one sub-query information. For example, Step1 can include three sub-query information 1a, 1b, and 1c, and Step2 can include three sub-query information 2a, 2b, and 2c. The sub-query information can be a sub-problem decomposed from the original problem, and solving this sub-problem helps to solve the entire original problem. The logical association relationship can be used to represent the continuous language sequence between the sub-query information in the inference stage. For example, 1a in Step1 and 2b in Step2.

[0047] It should be noted that the number of the above-mentioned inference stages and the number and specific content of the sub-query information decomposed in different inference stages are only for illustrative purposes and are not specifically limited here. As long as it is a process and method that can divide a complex query information into multiple relatively simple sub-query information according to the inference analysis stage and solve the entire query information by solving each sub-query information, it is within the protection scope of the embodiments of the present application.

[0048] In the related art, due to the lack of cognitive ability and logical reasoning ability of the large language model, it is relatively easy to face relatively simple conversations. However, when faced with answering relatively complex mathematical problems or logical reasoning problems, not only does it take a long time for reasoning, but also the answer to the problem cannot be obtained. Therefore, there is still the technical problem of low information inference efficiency.

[0049] As an alternative example, in the embodiments of the present application, in view of the above problems, for large models, a new inference framework can be designed to enhance the inference ability of large models in complex mathematical problems and the like. Through this inference framework, an original complex problem that is difficult to solve at once can be decomposed and processed into multiple relatively simple small problems that can be solved at once according to the inference stages. That is, complex query information can be decomposed and processed into simple sub-query information at different inference stages. Thus, the entire complex original problem can be finally solved by gradually inferring and solving the small problems. That is, by replying to the sub-query information, the reply information of the final query information is obtained. Through the above method of decomposing and inferring complex problems, the cognitive ability of large models and the inference ability for complex logical inference problems are enhanced, thereby achieving the technical effect of improving the efficiency of information inference.

[0050] Step S206: Determine sub-reply information that respectively matches the sub-query information at different inference stages.

[0051] In the technical solution provided in step S206 of the present application, after decomposing and processing the query information to obtain sub-query information at different inference stages, the sub-reply information corresponding to each sub-query information can be respectively determined. Among them, the sub-reply information can be used to represent the analysis result of the problem corresponding to the sub-query information.

[0052] Optionally, for the sub-query information included in each inference stage obtained by decomposing and processing the query information, reasoning and analysis can be performed one by one to determine the solution process and the final result for solving each sub-query information. That is, through reasoning and analysis, the sub-reply information corresponding to each sub-query information can be determined.

[0053] Since it is considered very difficult for a large model to solve a relatively complex problem at once. Therefore, in the embodiments of the present application, in order to ensure the efficiency of problem-solving and reduce the difficulty of problem-solving, an Intuitive System can be constructed in the inference framework of the large model. The original problem is analyzed and decomposed through the Intuitive System to generate decomposition hypotheses, that is, to generate sub-query information at different inference stages. By solving simple small problems one by one, the difficulty of solving the complex original problem is weakened, thereby achieving the technical effect of improving the efficiency of the large model in solving problems. Among them, the generation result of the Intuitive System is the sub-reply information of the sub-query information.

[0054] Step S208: Verify the sub-reply information that matches the sub-query information to obtain a verification result corresponding to the sub-query information, where the verification result is used to characterize the feasibility degree of the sub-query information for the inference query information.

[0055] In the technical solution provided in step S208 of the present application, after separately determining the sub-response information that matches the sub-query information in different inference stages, the sub-response information that matches the sub-query information can be verified to obtain the verification result corresponding to the sub-query information. Among them, the verification result can be used to represent the feasibility degree of the sub-query information for the inference query information, that is, it can be used to evaluate the feasibility and acceptability of the hypothesis of the sub-response information, and can be obtained by scoring the feasibility degree corresponding to the sub-response information. The verification result can be a decomposed score (Score|Decomposition). For example, the verification result can be a score of the feasibility degree: 30 points, 60 points or 100 points, or it can be Sure, Likely or Impossible. It should be noted that the above scoring form of the verification result is only for illustrative purposes and is not specifically limited here. As long as it is a form and method that can score the sub-response information corresponding to the sub-query information obtained by decomposing the query information, it is within the protection scope of the embodiments of the present application.

[0056] Optionally, in order to ensure the accuracy of the response information of the query information finally determined, it is necessary to evaluate the sub-response information obtained by solving each sub-query information obtained by decomposition to evaluate the acceptability of the sub-response information, and ensure the accuracy of the determined sub-response information through this method, so as to ensure the accuracy of the response information of the final query information.

[0057] In the embodiments of the present application, in order to ensure the accuracy of solving complex logical reasoning problems, etc., a reflective system can be constructed in the inference framework of the large model, and the reflective system is used to evaluate the sub-response information obtained by solving the sub-query information decomposed and processed in the intuitive system. That is, the role of the reflective system is to evaluate the generation result of the intuitive system to determine its acceptability, and the verified result after evaluation can be obtained. Through the verification result, the accuracy of the sub-response information is ensured, so as to ensure the accuracy of the response information, and further achieve the technical effect of improving the accuracy of the large model inference problem.

[0058] Step S210, in response to the verification result being greater than the verification result threshold, perform inference on the sub-query information corresponding to the verification result to obtain the response information corresponding to the query information.

[0059] In the technical solution provided in step S210 of the present application, after verifying the sub-response information matching the sub-query information to obtain the verification result corresponding to the sub-query information, the magnitude relationship between the verification result and the verification result threshold can be determined. If the verification result is greater than the verification result threshold, it can be indicated that the feasibility degree of the sub-query information for the inference query information is relatively large, and the sub-query information corresponding to the verification result can be inferred to obtain the response information of the query information. Among them, the verification result can be used to represent the correctness of the determined sub-query information and sub-response information, that is, to reflect the correctness of the hypothesis obtained by the verification system for the intuition system. The response information can be used to represent the final result corresponding to the query information. The verification result threshold can be a preset result in advance, or a result set by itself according to the actual information inference situation. The verification result threshold can be 99 points or Sure.

[0060] It should be noted that the above setting method and numerical value of the verification result are only for illustrative purposes and are not specifically limited here. As long as the conditions set can evaluate the sub-query information and sub-response information to ensure the accuracy of the response information, they are all within the protection scope of the embodiments of the present application.

[0061] Optionally, after determining the verification of the sub-query information and sub-response information to obtain the verification result, the magnitude relationship between the verification result and the verification result threshold can be determined. If the verification result is less than or equal to the verification result threshold, it can be indicated that the correctness of the obtained sub-query information and sub-response information is relatively low. If the response information of the final query information is inferred based on the sub-response information at this time, the accuracy cannot be guaranteed. If the verification result is greater than the verification result threshold, it can be indicated that the correctness of the obtained sub-query information and sub-response information is relatively high. At this time, the response information of the final query information can be inferred based on the sub-response information to ensure the accuracy of the response information.

[0062] In the embodiments of the present application, the intuitive system and the reflection system in a newly designed inference framework for large models can enhance the logical reasoning ability and cognitive ability of large models. The intuitive system simplifies complex problems, that is, it can divide the reasoning stages of complex problems according to the logical association relationship and reasoning ability, and divide complex problems into multiple simple small problems. And it can solve each simple small problem through the intuitive system to improve the ability of large models to solve complex problems. However, simply simplifying complex problems through the intuitive system cannot guarantee the accuracy of problem solving. Therefore, a reflection system can be designed to make up for it. The reflection system verifies the sub-query information and sub-response information analyzed by the intuitive system to ensure the accuracy of the solution. In other words, for the intuitive system and reflection system constructed for large models, the intuitive system generates hypotheses for the decomposition of the original problem, and the reflection system is used to verify the correctness of the hypotheses to guide the subsequent generation of the intuitive system. Thus, in the embodiments of the present application, through the complementarity of the intuitive system and the reflection system, the accuracy of the large model answering questions is guaranteed, and the efficiency of the large model answering questions is guaranteed, thereby achieving the technical effect of improving the efficiency of information reasoning.

[0063] Through steps S202 to S210 of the present application above, after detecting the existence of query information in the interaction interface of a certain client, the query information can be transmitted from the client to the server, and the query information is decomposed and processed in the server to determine sub-query information with logical association relationships in different reasoning stages. And it can reply to each sub-query information to obtain corresponding sub-response information. To ensure the feasibility of the determined response information, each sub-response information can be verified, and the verification results corresponding to each sub-response information. By judging whether the verification result is greater than the verification result threshold, it is determined whether the query information is feasible. If the verification result is greater than the verification result threshold, it means that the query information has a high degree of feasibility, and the sub-query information can be inferred to determine the response information of the query information. Considering that in the related art, due to the lack of cognitive ability of large models, the efficiency of reasoning logical problems with logical relationships and other problems is low. In the embodiments of the present application, the original query information can be divided into multiple sub-query information with logical association relationships according to the sequence of reasoning stages. By replying to the sub-query information in different reasoning stages, the response information of the query information can be finally determined, thereby achieving the purpose of greatly improving the accuracy of large models in solving problems such as logical reasoning, and further achieving the technical effect of improving the efficiency of information reasoning, and solving the technical problem of low efficiency of information reasoning.

[0064] The above method of this embodiment will be further introduced below.

[0065] As an alternative implementation, in step S204, the query information is decomposed to obtain sub-query information at different inference stages, including: based on the query information, searching for sub-query information at different inference stages in the sample set, where the sample set includes sub-query information corresponding to different query information.

[0066] In this embodiment, in the process of decomposing the query information to obtain sub-query information at different inference stages, based on the query information, sub-query information at different inference stages can be found in the sample set. Among them, the sample set can include sub-query information corresponding to different query information. The sample set can also be referred to as a training set, a decomposition set, in-context learning, or context, etc.

[0067] Optionally, before using the large model for information inference, a sample set can be prepared in advance. A large amount of query information can be summarized in the sample set, and the sub-query information obtained by dividing each query information according to the inference stage can be summarized for the server to call.

[0068] Optionally, the pre-prepared sample set can be deployed in the intuition system. Thus, after the server detects the query information transmitted by the client device, it can traverse the query information in the sample set in the intuition system and find the sub-query information corresponding to the query information from it.

[0069] Optionally, in the case of a logical reasoning problem, the intuition system can be used to determine the sub-query information corresponding to the query information from the sample set. By the above method, a relatively complex problem can be decomposed into smaller problems. By reasoning about these decomposed small problems, the final goal can be achieved, that is, the reply information of the query information can be obtained.

[0070] As an alternative implementation, based on the query information, searching for sub-query information at different inference stages in the sample set includes: determining the identification information of the query information; in the sample set corresponding to the inference stage, the sub-query information whose identification information matches the identification information of the query information is determined as the sub-query information at the inference stage.

[0071] In this embodiment, in the process of searching for sub-query information at different inference stages in the sample set, the identification information of the query information can be determined. According to the identification information of the query information sent by the current client device, it can be determined from the sample set whether there is identification information corresponding to the identification information. If so, the sub-query information matching the determined identification information can be used as the sub-query information of the current query information at the inference stage, where the identification information can be used to represent the current query.

[0072] Optionally, after the server obtains the query information sent by the client, it can decompose and match the query information according to the intuitive system in the large model of the server. The generation ability of the intuitive system is the basis for constructing the cognitive tree. The ability of the intuitive system can be enhanced through the context method. The query Q can be defined as the ultimate goal of the logical reasoning problem or the data problem, that is, the query information.

[0073] For example, the intuitive system can select a Decoder-Only model, such as the Generative Pre-trained Transformer (abbreviated as GPT) 2-XL model or the Language Learning and Multilingualism Alliance (abbreviated as LLaMA)-7B model, as the intuitive system. It should be noted that the models used in the above intuitive system are only for illustrative purposes and are not specifically limited here. As long as the model can decompose and match the query information, it is within the protection scope of the embodiments of the present application.

[0074] Optionally, in the case of a logical reasoning problem, the decomposition D involves further decomposing the query information into smaller problems. By reasoning about the above decomposition, the final response information can be determined.

[0075] Optionally, in the case of a mathematical problem, the decomposition D refers to the sub-problems decomposed from the original problem. Through the above-decomposed sub-problems, the final response information can be determined, that is, solving the sub-problems helps to solve the entire original problem.

[0076] Optionally, the decomposition set can represent the decomposition set of the examples in the training set, that is, the sub-query information contained in the sample set is obtained by decomposing a large amount of query information in advance.

[0077] Optionally, k examples (such as: query: Q; decomposition: query D) can be retrieved from the inference decomposition set, and then the above examples can be used as the context of the model input. The output can be generated as y~f θ (y|x,z 1...k )), where [y]~f θ (y|x,z 1...k ) can be used to represent a continuous language sequence. z can be used to represent the k examples retrieved from the decomposition set Z, Z = {z1,…z L}.

[0078] As an alternative implementation, in the sample set corresponding to the inference stage, the sub-query information whose identification information matches the identification information of the query information is determined as the sub-query information in the inference stage, including: retrieving, in the sample set corresponding to the inference stage, the sub-query information whose similarity between the identification information and the identification information of the query information is greater than the similarity threshold; and determining the retrieved sub-query information as the sub-query information in the inference stage.

[0079] In this embodiment, in the process of determining, in the sample set corresponding to the inference stage, the sub-query information whose identification information matches the identification information of the query information as the sub-query information in the inference stage, in the sample set corresponding to the inference stage, the sub-query information whose similarity between the identification information and the identification information of the query information is greater than the similarity threshold can be retrieved, and this sub-query information can be determined as the sub-query information in the inference stage, where the similarity can be used to characterize the similarity between the query information in the sample set and the query information sent by the client, and the similarity can be the cosine similarity. It should be noted that the above-mentioned similarity being the cosine similarity is only for illustrative purposes and is not specifically limited here.

[0080] Optionally, the similarity threshold can be a value set according to the actual information inference situation, or a pre-set value. For example, the similarity threshold can be pre-set to 99.9%. It should be noted that the above-mentioned setting method and value of the similarity threshold are only for illustrative purposes and are not specifically limited here.

[0081] Optionally, after obtaining the query information sent by the current client, the query information in the sample set can be traversed, and the similarity between each query information and the query information sent by the current client can be determined. If the similarity reaches the similarity threshold, it can indicate that the similarity between the two is relatively high. At this time, the query information in the sample set with a high similarity to the query information sent by the current client, and the sub-query information corresponding to this query information, can be used as the corresponding sub-query information of the query information sent by the current client in the inference stage.

[0082] Optionally, an intuition system is used to obtain the representation of the current query, that is, the identification information of the query information sent by the current client can be obtained by using the intuition system. The cosine similarity with the representations of other queries in the set can be calculated, that is, the identification information of the query information included in the sample set can be traversed, and during the traversal, the cosine similarity between each identification information in the sample set and the identification information of the query information sent by the client can be determined. At least one query information whose cosine similarity is greater than the similarity threshold can be retrieved from the set.

[0083] As an alternative implementation, determining the identification information of the query information includes: analyzing the query information using a first large model to obtain the identification information of the query information.

[0084] In this embodiment, in the process of determining the identification information of the query information, a first large model can be used to analyze the query information to determine the identification information of the query information. Among them, the first large model can be an intuition system, also known as an implicit extraction module, which is a relatively small-scale model.

[0085] Optionally, when the client has a need to request a large model to answer the question to be queried, the user can input query information describing the content to be queried on the interaction interface of the client. The query information collected by the interaction interface can be transmitted to the server through the network. Using the intuition system in the server, the query information of the current client can be analyzed to obtain the identification information of the query information, that is, using the intuition system, the representation of the current query can be obtained.

[0086] As an alternative implementation, in the sample set corresponding to the inference stage, the sub-query information whose identification information matches the identification information of the query information is determined as the sub-query information in the inference stage, including: analyzing the sample set corresponding to the inference stage and the identification information of the query information using a first large model to obtain the sub-query information whose identification information matches the identification information of the query information; using the first large model to determine the matching sub-query information as the sub-query information in the inference stage.

[0087] In this embodiment, in the process of determining the sub-query information whose identification information matches the identification information of the query information in the sample set corresponding to the inference stage, a first large model can be used to analyze the sample set corresponding to the inference stage and the identification information of the query information, so as to determine the sub-query information corresponding to the identification information that can match the identification information of the current query information. And a first large model can be used to determine the matching sub-query information as the sub-query information in the inference stage, where the first large model can be the model corresponding to the intuition system.

[0088] Optionally, after using the intuition system to obtain the identification information corresponding to the query information sent by the current client, the intuition system can also be used to call the sample set stored in the intuition system based on the current identification information. The intuition system traverses the identification information of each query information in the sample set and determines the similarity between each identification information and the identification information of the current query information. It can be judged whether there is a similarity exceeding the similarity threshold. If so, the sub-query information corresponding to the identification information in the sample set can be determined as the sub-query information of the query information of the current client in the inference stage.

[0089] Optionally, supervised fine-turning (SFT) has been proven effective in understanding the user's intention. Therefore, in the embodiments of this application, the intuitive system can decompose the query Q into sub-problems using context examples, that is, the intuitive system can decompose the query information samples of complex problems into corresponding sub-query information samples. Since a generative model can be used as the intuitive system, during the training of the intuitive system, the loss of the sub-query information samples generated based on the query information samples can be calculated to determine the loss degree of the sub-query information, and the intuitive system can be trained using the loss degree and the sample set.

[0090] Optionally, during autoregressive calculation, the generated text that does not include the given context can be used, that is, the loss calculation is only performed on the sub-query information samples. For example, given a sample of length N, it can be represented by X, where X = {x1,…x i ,…x n}. The sequence length of the context example can be defined as M. The following likelihood function is maximized using the standard language modeling objective: Thus, the intuitive system is trained.

[0091] As an alternative implementation, determining the sub-response information that matches the sub-query information at different reasoning stages includes: using the first large model to analyze the sub-query information at different reasoning stages to obtain the sub-response information that matches the sub-query information at different reasoning stages.

[0092] In this embodiment, during the process of determining the sub-response information that matches the sub-query information at different reasoning stages, the first large model can be used to analyze the sub-query information at different reasoning stages to obtain the sub-response information that matches the sub-query information at different reasoning stages.

[0093] Optionally, after using the first model to determine the identification information of the query information that matches the identification information of the current query information from the sample set, the sub-query information corresponding to the identification information can be determined. The first large model can be used to solve each sub-query information separately to obtain the sub-response information corresponding to each sub-query information.

[0094] Optionally, after using the intuitive system to analyze the current query information and determine the sub-query information in the reasoning stage of the query information, the sub-query information at different reasoning stages can be analyzed to solve the sub-response information corresponding to each sub-query information, so that the response information of the query information can be finally solved.

[0095] In the embodiments of the present application, when solving problems that require logical reasoning or relatively complex mathematical problems, based on the cognitive theory of humans, the process of humans generating cognition can be imitated through the intuitive system and the reflective system. That is, the above two systems are used to improve the cognitive ability. The intuitive system is responsible for generating multiple decomposition hypotheses for the original problem and solving each of the multiple decomposition hypotheses one by one. The reflective system is used to verify the hypotheses and solutions generated by the intuitive system, and select more likely hypotheses for subsequent generation until the final result is reached. Through the iterative generation of the above dual system, the purpose of improving the problem-solving accuracy of the large model can be achieved, thereby realizing the technical effect of improving the accuracy of information reasoning of the large model.

[0096] As an alternative implementation, verifying the sub-response information matching the sub-query information to obtain the verification result corresponding to the sub-query information includes: when running in the current inference stage of different inference stages, verifying the sub-response information matching the sub-query information of the current inference stage to obtain the sub-verification result of the current inference stage, where the verification result includes the sub-verification result; or, verifying the sub-response information of different inference stages to obtain the overall verification result of different inference stages, where the verification result includes the overall verification result.

[0097] In this embodiment, in the process of verifying the sub-response information matching the sub-query information to obtain the verification result corresponding to the sub-query information, when running in each inference stage of different inference stages, the sub-response information matching the sub-query information of the current inference stage can be verified to obtain the sub-verification result of the current inference stage, or the sub-response information of different inference stages can also be verified as a whole to obtain the overall verification result of different inference stages, where the verification result can include the sub-verification result and the overall verification result. The sub-verification result can be used to represent the verification of the intermediate process and can be the score of the current state. The overall verification result can be used to represent the verification of the entire inference chain and can be the overall score of the entire inference chain.

[0098] In the embodiments of the present application, the sub-response information generated by the first large model can be evaluated to determine its acceptability. Two methods can be used to verify it to obtain the verification result. For example, the verification of the intermediate process and the verification of the entire inference chain are used to ensure the rationality of the entire generated hypothesis and the inference process, and thus the technical effect of improving the accuracy of information reasoning of the large model is achieved.

[0099] As an alternative implementation, the sub-response information of the current inference stage in different inference stages is verified to obtain the sub-verification result of the current inference stage, including: verifying the sub-response information of the current inference stage using the second largest model to obtain the sub-verification result; verifying the sub-response information of different inference stages to obtain the overall verification result of different inference stages, including: verifying the sub-response information of different inference stages using the second largest model to obtain the overall verification result.

[0100] In this embodiment, the second largest model can be used to verify the sub-response information of the current inference stage to obtain the sub-verification result, or the second largest model can be used to verify the sub-response information of different inference stages to obtain the overall verification result. Among them, the second largest model can be a reflection system, also known as an explicit inference module.

[0101] Optionally, after using the intuition system to decompose the query information to obtain the sub-query information and the sub-response information corresponding to the sub-query information respectively, the sub-response information can be transmitted to the reflection system, and the reflection system can be used to evaluate the acceptability of the sub-response information determined by the intuition system.

[0102] Optionally, the reflection system is different from the intuition system in function. The intuition system relies on fast intuition for generation, while the function of the reflection system is to evaluate the generation result of the intuition system to determine its acceptability. The reflection system can verify the result through two methods: verification of the intermediate process and verification of the overall inference chain.

[0103] Optionally, based on the above analysis, a second largest model with the same model architecture as the largest model can be determined, and the second largest model can be used to verify the sub-response information in different inference stages and obtain the verification result.

[0104] For example, for the verification of the intermediate process, given the current state S (query: Q and decomposition: D), a reflection system with the same model architecture as the intuition system can be used to generate a score v for verifying the current state, where v can be expressed as V(f θ ,s)~f θ (v|s).

[0105] For another example, for the verification of the overall inference chain, given the complete inference chain as S = {s1,…,s i ,…,s n}, a reflection system can be used to generate an overall score o, where o can be expressed as O(f θ ,S)~f θ (o|S).

[0106] In the embodiment of the present application, the difference between the reflection system and the intuition system is that its important task is to evaluate and verify the feasibility of the entire reasoning chain of the current state, rather than generating quick hypotheses like the intuition system. This evaluation process helps to determine that the generated hypotheses and the reasoning process are reasonable, thus achieving the technical effect of improving the accuracy of information reasoning of the large model.

[0107] Optionally, in the process of training the reflection system, the same training method as the intuition system can be adopted. Using positive and negative samples, the generation model generates classification results from the positive and negative samples, that is, generates corresponding verification result samples, and determines the loss degree of the verification result samples.

[0108] Optionally, since the reflection system mainly focuses on the judgment of state s, the following loss function can be used to determine the loss degree of the verification result samples:

[0109] As an optional implementation manner, the query information is decomposed to obtain sub-query information at different reasoning stages, including: in the sample set corresponding to the current reasoning stage, based on the query information, the sub-query information of the current reasoning stage is found; in the case that different reasoning stages include the next reasoning stage of the current reasoning stage, in the sample set corresponding to the next reasoning stage, based on the sub-verification result and the query information of the current reasoning stage, the sub-query information of the next reasoning stage is found.

[0110] In this embodiment, in the sample set corresponding to the current reasoning stage, based on the query information, the sub-query information in the current reasoning stage can be found. In the case that different reasoning stages include the next reasoning stage of the current reasoning stage, in the sample set corresponding to the next reasoning stage, based on the sub-verification result and the query information of the current reasoning stage, the sub-query information of the next reasoning stage can be found.

[0111] In the embodiment of the present application, two main components are constructed, the intuition system and the reflection system. The intuition system can quickly generate multiple answers using context instances, while the reflection system can use comparative learning to score the answers of the intuition system. This score can guide the intuition system to operate in subsequent generation steps.

[0112] Optionally, after the reflection system determines the verification result, the verification result can be sent to the intuition system. The intuition system can continue with the subsequent generation steps based on the evaluation by the reflection system. That is, after the reflection system evaluates the sub-query information of a certain reasoning stage and obtains the sub-verification result, the sub-verification result can be sent to the intuition system. The intuition system can reason about the sub-query information in the next reasoning stage of the current reasoning stage based on the sub-verification result. If all reasoning stages have been reasoned and evaluated, the reflection system can evaluate the entire reasoning chain, that is, determine the overall verification result.

[0113] As an alternative implementation, the method may further include: determining the query information as the root node, determining the sub-query information as the leaf node, and establishing a target tree structure; reasoning about the sub-query information corresponding to the verification result to obtain the response information corresponding to the query information, including: in the target tree structure, reasoning about the sub-query information corresponding to the verification result to obtain the response information corresponding to the query information.

[0114] In this embodiment, the query information can be determined as the root node of the tree, and the sub-query information can be determined as the leaf node of the tree, thereby establishing a target tree structure. In the target tree structure, the sub-query information corresponding to the verification result can be reasoned to obtain the response information corresponding to the query information, where the target tree structure can be a Cognitive Tree (CogTree), also known as an inference tree.

[0115] Optionally, the inference framework can adopt an iterative method to construct the structure of a cognitive tree by referring to the thinking mode of humans. The root node of the cognitive tree can represent the initial query, that is, the query information, and the leaf nodes can contain simple questions with direct answers, that is, the sub-query information.

[0116] Optionally, for lightweight complex task reasoning, a smaller-scale model can be used to construct a dual-system generated inference tree to effectively enhance the answering ability of the large model in complex mathematical problems and logical reasoning problems, etc.

[0117] Optionally, the cognitive tree uses a smaller-scale and more open-source model to solve the problem of high model inference cost, and the inference tree is a process of generating an inference tree using a dual system, enhancing the problem of insufficient inference ability of the large model itself.

[0118] In the embodiments of the present application, the large model can construct two systems respectively, an intuitive system and a reflective system. The intuitive system can be used to generate hypotheses for the decomposition of the original problem, and the reflective system can be used to verify the correctness of the hypotheses and guide the subsequent generation of the intuitive system. Through the iterative generation of an inference tree by the dual systems, the inference ability of the large model is enhanced, so as to enhance the inference ability of the large model in complex mathematical problems and logical reasoning problems, thereby achieving the technical effect of improving the accuracy of information inference of the large model.

[0119] The embodiments of the present application also provide an information inference method on the human-computer interaction side. Figure 3 It is a flowchart of an information inference method according to an embodiment of the present application. As Figure 3 shown, the method may include the following steps:

[0120] Step S302, in response to the query information to be inferred received in the dialogue interface, on the dialogue interface, display the sub-query information of the query information at different inference stages, where there is a logical association relationship between the sub-query information at different inference stages.

[0121] In the technical solution provided in step S302 of the present application above, in response to the query information to be inferred received in the dialogue interface, on the dialogue interface, the sub-query information of the query information at different inference stages can be displayed, where there is a logical association relationship between the sub-query information at different inference stages.

[0122] Optionally, when the client needs to query the large model for a question to be answered, the query information corresponding to the question to be answered can be input on the interaction interface of its client device. And the query information can be displayed on the dialogue interface.

[0123] Optionally, the client device can collect the query information and send it to the server through the network. Since the large model capable of inferring the query information is deployed in the server. Thus, in the server, the large model can be used to process the query information.

[0124] Optionally, the server can decompose and process the query information to obtain the sub-query information at different inference stages. And the inferred sub-query information can be sent to the client device and displayed on the dialogue interface.

[0125] Optionally, if the user believes that the decomposed sub-query information does not meet the requirements, it can be adjusted through corresponding operations.

[0126] Step S304, on the dialogue interface, display the sub-reply information respectively matching the sub-query information at different inference stages.

[0127] In the technical solution provided in step S304 of the present application, sub-reply information that matches the sub-query information in different inference stages can also be displayed on the dialogue interface.

[0128] Optionally, for each sub-query information included in each inference stage obtained by decomposing and processing the query information, inference and analysis can be performed one by one to determine the solution process and the final result for solving each sub-query information. That is, through inference and analysis, the sub-reply information corresponding to each sub-query information can be determined.

[0129] Optionally, the sub-reply information corresponding to the sub-query information solved by the server can be sent to the corresponding client device and displayed on the interaction interface of the client device. If the user believes that the solved sub-reply information does not meet the requirements, it can be adjusted through corresponding operations.

[0130] Step S306, display the reply information corresponding to the query information on the dialogue interface, where the reply information is obtained by inferring the sub-query information corresponding to the verification result when the verification result corresponding to the sub-query information is greater than the verification result threshold. The verification result is obtained by verifying the sub-reply information that matches the sub-query information and is used to represent the feasibility degree of the sub-query information for the inference query information.

[0131] In the technical solution provided in step S306 of the present application, the reply information corresponding to the query information can be displayed on the dialogue interface, where the reply information can be obtained by inferring the sub-query information corresponding to the verification result when the verification result corresponding to the sub-query information is greater than the verification result threshold. The verification result can be obtained by verifying the sub-reply information that matches the sub-query information and can be used to represent the feasibility degree of the sub-query information for the inference query information.

[0132] Optionally, for each sub-query information included in each inference stage obtained by decomposing and processing the query information, inference and analysis can be performed one by one to determine the solution process and the final result for solving each sub-query information. That is, through inference and analysis, the sub-reply information corresponding to each sub-query information can be determined.

[0133] Optionally, after respectively determining the sub-reply information that matches the sub-query information in different inference stages, the sub-reply information that matches the sub-query information can be verified to obtain the verification result corresponding to the sub-query information.

[0134] Optionally, in order to ensure the accuracy of the reply information for the query information finally determined, it is necessary to evaluate the sub-reply information obtained by solving each sub-query information obtained by decomposition to evaluate the acceptability of the sub-reply information, and ensure the accuracy of the determined sub-reply information through this method, so as to ensure the accuracy of the reply information for the final query information.

[0135] Optionally, after verifying the sub-reply information matching the sub-query information and obtaining the verification result corresponding to the sub-query information, the magnitude relationship between the verification result and the verification result threshold can be judged. If the verification result is greater than the verification result threshold, it can be explained that the sub-query information has a greater feasibility for the inference query information, and the sub-query information corresponding to the verification result can be inferred to obtain the reply information for the query information.

[0136] Optionally, if the verification result is less than or equal to the verification result threshold, it can be explained that the correctness of the obtained sub-query information and sub-reply information is relatively low. If the reply information for the final query information is inferred based on the sub-reply information at this time, the accuracy cannot be guaranteed. If the verification result is greater than the verification result threshold, it can be explained that the correctness of the obtained sub-query information and sub-reply information is relatively high. At this time, the reply information for the final query information can be inferred based on the sub-reply information to ensure the accuracy of the reply information.

[0137] Optionally, after determining the reply information, the reply information can be transmitted to the corresponding client device through the network and can be displayed on the dialogue interface of the client device.

[0138] Optionally, if the user believes that the solved reply information does not meet the requirements, it can be adjusted through corresponding operations.

[0139] Through the above steps S302 to S306 of the present application, in response to the query information to be inferred received in the dialogue interface, on the dialogue interface, the sub-query information at different inference stages of the query information is displayed, where there is a logical association relationship between the sub-query information at different inference stages; on the dialogue interface, the sub-reply information respectively matching the sub-query information at different inference stages is displayed; on the dialogue interface, the reply information corresponding to the query information is displayed, where the reply information is obtained by inferring the sub-query information corresponding to the verification result when the verification result corresponding to the sub-query information is greater than the verification result threshold, the verification result is obtained by verifying the sub-reply information matching the sub-query information, and is used to characterize the feasibility of the sub-query information for the inference query information, thus achieving the technical effect of improving the efficiency of information inference and solving the technical problem of low efficiency of information inference.

[0140] The above method of this embodiment will be further introduced below.

[0141] As an alternative embodiment, the query information is multimodal information, and the types of the multimodal information include at least one of the following: text information containing character information, video frame information containing frame image information, and audio information. The types of the reply information include at least one of the following: text information, image information, video information, and voice information.

[0142] In this embodiment, the query information may be multimodal information. The multimodal information may include at least one of the following: text information containing character information, video frame information containing frame image information, and audio information. The types of the reply information may include at least one of the following: text information, image information, video information, and voice information.

[0143] It should be noted that the types of the query information sent by the above user and the reply information replied by the large model are only for illustrative purposes and are not specifically limited here.

[0144] In the embodiment of the present application, after detecting that there is query information in the interaction interface of a certain client, the query information may be transmitted from the client to the server, and the query information is decomposed and processed in the server to determine sub-query information that has a logical association relationship with each other in different inference stages. And each sub-query information may be replied to obtain corresponding sub-reply information. In order to ensure the feasibility degree of the determined reply information, each sub-reply information may be verified, and the verification results corresponding to each sub-reply information. By determining whether the verification result is greater than the verification result threshold, it is determined whether the query information is feasible. If the verification result is greater than the verification result threshold, it means that the feasibility degree of the query information is relatively high, and the sub-query information may be inferred to determine the reply information of the query information. Since it is considered that in the related art, due to the lack of cognitive ability of the large model, the efficiency of reasoning logical problems with logical relationships and other problems is low, the embodiment of the present application may divide the original query information into multiple sub-query information with logical association relationships according to the sequence of the inference stages. By replying to the sub-query information in different inference stages, the reply information of the query information can be finally determined, thereby achieving the purpose of greatly improving the accuracy of the large model in solving problems such as logical reasoning, and further achieving the technical effect of improving the efficiency of information inference, and solving the technical problem of low efficiency of information inference.

[0145] The embodiment of the present application also provides a method for generating a model on the model training side. Figure 4 It is a flowchart of a method for generating a model according to an embodiment of the present application. As Figure 4 shown, the method may include the following steps:

[0146] Step S402, obtaining a query information sample.

[0147] In the technical solution provided in step S402 of the present application, query information samples for training the first large model can be obtained.

[0148] Step S404: Use the sub-query information samples corresponding to the query information samples to train the first large model, where the first large model is used to decompose the input query information to obtain sub-query information at different inference stages, and there is a logical association relationship between the sub-query information at different inference stages.

[0149] Optionally, query information samples can be obtained, and the query information samples can be input into a generation model for analysis to obtain corresponding sub-query information samples, and the loss degree of the sub-query information samples can be determined. Then, using the loss degree and the sample set, the generation model can be trained to obtain the first large model, where the sub-query information samples can be used to represent the text decomposed and generated based on the query information samples. The sample set can be context examples.

[0150] In the technical solution provided in step S402 of the present application, after obtaining the query information samples, the sub-query information samples corresponding to the query information samples can be used to train the first large model, where the first large model can be used to decompose the input query information to obtain sub-query information at different inference stages, and there is a logical association relationship between the sub-query information at different inference stages.

[0151] Step S406: Obtain the response information samples corresponding to the query information samples.

[0152] In the technical solution provided in step S406 of the present application, the response information samples corresponding to the query information samples can also be obtained.

[0153] Step S408: Obtain the verification result samples corresponding to the response information samples, where the verification result samples are used to represent at least the classification result of the response information samples.

[0154] In the technical solution provided in step S408 of the present application, after obtaining the response information samples corresponding to the query information samples, the sample result samples corresponding to the response information samples can be obtained, where the verification result samples can be used to represent at least the classification result of the response information samples.

[0155] Optionally, obtain the query information samples and the corresponding response information samples, input the query information samples and the response information samples into a generation model for analysis to obtain the corresponding verification result samples. Determine the loss degree of the verification result samples, and use the verification result samples to train the generation model to obtain the second large model

[0156] Step S410: Train a second large model using the verification result samples. The second large model is used to verify the sub-response information matched with the sub-query information to obtain the verification result corresponding to the sub-query information. The verification result is used to represent the feasibility degree of the sub-query information for the inference query information. The sub-query information corresponding to the verification result greater than the verification result threshold is used to infer the response information corresponding to the query information.

[0157] In the technical solution provided in step S410 of the present application, after obtaining the sample result samples corresponding to the response information samples, a second large model can be trained using the verification result samples. The second large model can be used to verify the sub-response information matched with the sub-query information to obtain the verification result corresponding to the sub-query information. The verification result can be used to represent the feasibility degree of the sub-query information for the inference query information. The sub-query information corresponding to the verification result greater than the verification result threshold is used to infer the response information corresponding to the query information.

[0158] Optionally, after determining that the sub-query information and the sub-response information are verified to obtain the verification result, the magnitude relationship between the verification result and the verification result threshold can be judged. If the verification result is less than or equal to the verification result threshold, it can be explained that the correctness of the obtained sub-query information and sub-response information is relatively low. If the response information to the final query information is inferred based on the sub-response information at this time, the accuracy cannot be guaranteed. If the verification result is greater than the verification result threshold, it can be explained that the correctness of the obtained sub-query information and sub-response information is relatively high. At this time, the response information to the final query information can be inferred based on the sub-response information to ensure the accuracy of the response information.

[0159] The above method of this embodiment will be further introduced below.

[0160] As an optional implementation manner, training a first large model using the sub-query information samples corresponding to the query information samples includes: determining the loss degree of the sub-query information samples; training a generation model using the loss degree and the sample set to obtain the first large model, where the sample set includes sub-query information corresponding to different query information, and the generation model is a pre-trained generative large model.

[0161] In this embodiment, in the process of training a first large model using the sub-query information samples corresponding to the query information samples, the loss degree of the sub-query information samples can be determined, and the generation model can be trained using the loss degree and the sample set to obtain the first large model, where the sample set can include sub-query information corresponding to different query information. The generation model is a pre-trained generative large model.

[0162] Optionally, the SFT has been proven to be effective in recognizing the intentions of cognitive users. Therefore, in the embodiments of the present application, the intuitive system can decompose the query Q into sub-problems by using context examples, that is, the intuitive system can decompose the query information samples of complex problems into corresponding sub-query information samples. Since a generative model can be used as the intuitive system, during the training of the intuitive system, the loss of the sub-query information samples generated based on the query information samples can be calculated to determine the loss degree of the sub-query information, and the intuitive system can be trained by using the loss degree and the sample set.

[0163] Optionally, during the autoregressive calculation, the generated text that does not include the given context can be used, that is, the loss calculation is only performed on the sub-query information samples. For example, given a sample of length N, it can be represented by X, where X = {x1, … x i , … x n}. The sequence length of the context example can be defined as M. The following likelihood function is maximized using the standard language modeling objective: Thus, the intuitive system is trained.

[0164] As an alternative implementation, training the second large model using the verification result samples includes: determining the loss degree of the verification result samples; training the generative model using the loss degree of the verification result samples to obtain the second large model, where the generative model is a pre-trained large generative model.

[0165] In this embodiment, during the process of training the second large model using the verification result samples, the loss degree of the verification result samples can be determined, and the generative model can be trained using the loss degree of the verification result samples to obtain the second large model, where the generative model is a pre-trained large generative model.

[0166] Optionally, during the process of training the reflection system, the same training method as the intuitive system can be adopted, and positive and negative samples can be used to enable the generative model to generate classification results from the positive and negative samples, that is, to generate corresponding verification result samples, by determining the loss degree of the verification result samples.

[0167] Optionally, since the reflection system mainly focuses on the judgment of the state s, the following loss function can be used to determine the loss degree of the verification result samples:

[0168] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data that have been authorized by the user or fully authorized by all parties. And the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions, and corresponding operation entrances are provided for users to choose to authorize or refuse.

[0169] It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that this application is not limited by the described action sequence, because according to this application, 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 essential to this application.

[0170] Through the description of the above embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the technical solution of this application, in essence, or the part that makes a contribution to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disc), and includes several instructions for causing a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods of various embodiments of this application.

[0171] Embodiment 2

[0172] According to an embodiment of this application, an information inference system is also provided. Figure 5 is a schematic diagram of an information inference system according to an embodiment of this application. As Figure 5 shown, the information inference system may include: an information generation end 502 and an information verification end 504.

[0173] The information generation end 502 is used to detect the query information to be inferred, decompose the query information to obtain sub-query information at different inference stages, and determine sub-response information that matches the sub-query information at different inference stages, where there is a logical association relationship between the sub-query information at different inference stages.

[0174] In the technical solution provided by the above-mentioned information generation end 502 of the present application, the query information to be inferred can be detected by the information generation end 502, and the query information can be decomposed and processed to obtain sub-query information at different inference stages, and the sub-response information corresponding to each sub-query information can be determined respectively. Among them, there is a logical association relationship between the sub-query information at different inference stages.

[0175] Optionally, the information generation end 502 is used to infer and analyze the query information input on the interaction interface that can describe the queried problem.

[0176] Optionally, the information generation end 502 can divide the query information into multiple sub-query information according to the logical association relationship in the inference stage.

[0177] Optionally, for each sub-query information included in each inference stage obtained by decomposing the query information by the information generation end 502, inference and analysis can be performed one by one to determine the solution process and the final result for solving each sub-query information, that is, through inference and analysis, the sub-response information corresponding to each sub-query information can be determined. And the sub-response information can be transmitted to the information verification end 504.

[0178] The information verification end 504 is used to verify the sub-response information matching the sub-query information to obtain the verification result corresponding to the sub-query information. In response to the verification result being greater than the verification result threshold, the sub-query information corresponding to the verification result is inferred to obtain the response information corresponding to the query information, where the verification result is used to represent the feasibility degree of the sub-query information for inferring the query information.

[0179] In the technical solution provided by the above-mentioned information verification end 504 of the present application, the sub-query information can be verified by the information verification end 504 to determine the verification result. When the verification result is greater than the verification result threshold, the sub-query information corresponding to the verification result can be inferred to obtain the response information corresponding to the query information, where the verification result can be used to represent the feasibility degree of the sub-query information for inferring the query information.

[0180] Optionally, in order to ensure the accuracy of the response information of the query information finally determined, it is necessary to evaluate the sub-response information obtained by solving each sub-query information decomposed by the information verification end 504 to evaluate the acceptability of the sub-response information, and to ensure the accuracy of the determined sub-response information through this method, so as to ensure the accuracy of the response information of the final query information.

[0181] Optionally, after determining and verifying the sub-query information and the sub-response information to obtain a verification result, the magnitude relationship between the verification result and the verification result threshold can be judged. If the verification result is less than or equal to the verification result threshold, it can indicate that the correctness of the obtained sub-query information and sub-response information is relatively low. If the response information of the final query information is inferred based on the sub-response information at this time, the accuracy cannot be guaranteed. If the verification result is greater than the verification result threshold, it can indicate that the correctness of the obtained sub-query information and sub-response information is relatively high. At this time, the response information of the final query information can be inferred based on the sub-response information to ensure the accuracy of the response information.

[0182] Through the above information inference system of the present application, the query information to be inferred is detected by the information generation end, and the query information is decomposed to obtain sub-query information at different inference stages, and sub-response information that matches the sub-query information at different inference stages is determined. Among them, there is a logical association relationship between the sub-query information at different inference stages; the information verification end is used to verify the sub-response information that matches the sub-query information to obtain the verification result corresponding to the sub-query information. In response to the verification result being greater than the verification result threshold, the sub-query information corresponding to the verification result is inferred to obtain the response information corresponding to the query information. Among them, the verification result is used to represent the feasibility degree of the sub-query information for inferring the query information. Thus, the technical effect of improving the efficiency of information inference is achieved, and the technical problem of low efficiency of information inference is solved.

[0183] Embodiment 3

[0184] Currently, large language models have achieved remarkable results in multiple tasks such as creative generation, dialogue systems, and brainstorming. However, large models have certain limitations in complex tasks such as logical reasoning and solving mathematical problems, especially those involving abstract concepts or multi-step reasoning questions.

[0185] With the continuous development of deep learning in tasks such as natural language processing and machine translation, there is an increasing interest in how to apply deep learning to natural language processing. As a result, large language models (such as GPT-3.5) have emerged and have made significant breakthroughs in multiple tasks such as text generation, sentiment analysis, and dialogue systems. Large language models are usually pre-trained based on large-scale text data and then optimized on specific tasks through fine-tuning to generate high-quality text outputs. These models possess powerful natural language understanding and generation capabilities, can understand context, and generate coherent language, so they have received much attention in the field of natural language processing. By continuously improving the model structure and training methods, efforts are being made to improve the performance of large language models to better meet the growing application requirements. The development in this field provides new opportunities and challenges for natural language processing and artificial intelligence research.

[0186] However, in the related art, for language models, it is still very difficult to solve complex logical reasoning problems and mathematical problems. Moreover, traditional language models lack cognitive capabilities. When dealing with problems involving long reasoning chains or multi-step solutions, it is important to evaluate the problem and its current answer. However, the deployment and inference costs of large language models are relatively high, especially when using inference enhancement techniques without parameter updates. These techniques require a large amount of context and multi-step answer generation, further increasing the inference cost and time. Therefore, there is still a technical problem of low efficiency in information inference.

[0187] Optionally, the present application provides a complex task inference method for large language models, which solves the technical problem of low efficiency in information inference. Different from the long inference time for complex tasks caused by poor cognitive capabilities in traditional solutions and the resulting low efficiency in information inference, it solves the technical problem of low efficiency in information inference.

[0188] In an embodiment of the present application, after detecting that there is query information in the interaction interface of a certain client, the query information can be transmitted from the client to the server, and the query information is decomposed and processed in the server to determine sub-query information with logical association relationships among different inference stages. And each sub-query information can be replied to obtain corresponding sub-reply information. To ensure the feasibility of the determined reply information, each sub-reply information can be verified, and the verification results corresponding to each sub-reply information can be obtained. By determining whether the verification result is greater than the verification result threshold, it is determined whether the query information is feasible. If the verification result is greater than the verification result threshold, it means that the query information has a high degree of feasibility, and the sub-query information can be inferred to determine the reply information of the query information.

[0189] Since the embodiment of the present application takes into account that in the related art, due to the lack of cognitive capabilities of large models, the efficiency of inferring logical problems with logical relationships and other problems is low, the original query information can be divided into multiple sub-query information with logical association relationships according to the sequence of inference stages. By replying to the sub-query information at different inference stages, the reply information of the query information can be finally determined, thereby achieving the purpose of greatly improving the accuracy of the large model in solving problems such as logical reasoning, and further realizing the technical effect of improving the efficiency of information inference, and solving the technical problem of low efficiency in information inference.

[0190] The above method of this embodiment will be further introduced below.

[0191] In this embodiment, for complex task reasoning in lightweight large models, a relatively small-scale model (7B) is used to construct a dual-system generation inference tree, effectively enhancing the model's answering ability for complex mathematical problems and logical reasoning problems. A method for solving complex mathematical problems by a large model is proposed. This method is based on the human cognitive theory and imitates the process of human cognition through two systems: the intuitive system and the reflective system. The intuitive system is responsible for generating multiple decomposition hypotheses for the original problem, and the reflective system verifies the hypotheses generated by the intuitive system and selects more likely hypotheses for subsequent generation until the final result is reached. Through the iterative generation of the above dual system, the problem-solving accuracy of the large model can be improved.

[0192] In related technologies, when using models such as ChatGPT or GPT-4 to solve logical reasoning problems or mathematical problems, context instances are required. However, their inference costs (time and space) are very high. For example, the number of model parameters is 175B, and the model is not open source. Smaller models, such as LLaMA-7B, etc., cannot achieve the expected accuracy for complex tasks.

[0193] Optionally, for lightweight complex task reasoning, a relatively small-scale model can be used to construct a dual-system generation inference tree to greatly enhance the large model's answering ability for complex mathematical problems, logical reasoning problems, etc.

[0194] Optionally, a cognitive tree uses a smaller and more open-source model to solve the problem of high model inference cost, and the inference tree is a process of adopting a dual-system generation inference tree, enhancing the problem of insufficient inference ability of the large model itself.

[0195] Optionally, the inference framework can adopt an iterative method, draw on the human thinking mode to construct the structure of a cognitive tree. The root node of the cognitive tree can represent the initial query, that is, the query information, and the leaf node can contain simple questions for direct answers, that is, sub-query information.

[0196] In this embodiment, Figure 6 is a schematic diagram of a cognitive tree according to an embodiment of the present application. As Figure 6 shown, in the face of complex task reasoning, such as logical reasoning problems and mathematical problems, the client can, based on the above problems, issue query information that requires a large model to answer. The query information can be transmitted to the large model of the dual system. Drawing on the human thinking mode, a cognitive tree structure as Figure 6 shown can be proposed. The large model can be decomposed to construct two systems: the intuitive system 601 and the reflective system 602. The intuitive system 601 is used to generate hypotheses for the decomposition of the original problem, that is, the intuitive system can decompose the query information into sub-query information, as Figure 6As shown, the query information can be decomposed into sub-query information A and sub-query information B. Sub-query information A cannot be further divided, and sub-query information B can be further divided into sub-query information B1 and sub-query information B2. A reflection system can be used to verify the correctness of the hypothesis and guide the subsequent generation of the intuition system. That is, the reflection system 602 can be used to evaluate the sub-query information decomposed by the intuition system. After the evaluation meets the requirements, subsequent operations can be performed through the intuition system. That is, subsequently, the intuition system is used to answer each sub-query information to obtain corresponding sub-answer information, and finally the answer information of the query information is obtained to answer relatively complex logical reasoning problems and mathematical problems for the client.

[0197] In the embodiment of the present application, an inference tree is generated through a dual-system iterative method to enhance the inference ability of the large model. The creativity of this method is to design a new inference framework for large language models to enhance the inference ability of large models in complex mathematical problems and logical reasoning problems.

[0198] Optionally, the generation ability of the intuition system is the basis for constructing the cognitive tree. Therefore, a model that only includes Decoder-Only is selected as the intuition system. The ability of the intuition system is enhanced through the context method. Define the query Q as the ultimate goal of the logical reasoning problem or the mathematical problem.

[0199] Optionally, in the case of logical reasoning problems, the decomposition D involves further decomposing the query information into smaller problems, and the ultimate goal can be achieved through reasoning about the above decomposition.

[0200] Optionally, in the case of mathematical problems, the decomposition D refers to the sub-problems derived from the original problem, and solving this sub-problem helps to solve the entire original problem.

[0201] Optionally, the decomposition set can represent the decomposition set of examples in the training set. k examples (such as: query: Q; decomposition: query D) can be retrieved from the inference decomposition set, and then the above examples can be used as the context of the model input. The output can be generated as y~f θ (y|x,z 1...k ) where, [y]~f θ (y|x,z 1...k ) can be used to represent a continuous language sequence. z can be used to represent the k examples retrieved from the decomposition set Z, Z = {z1,…z L}.

[0202] Optionally, after obtaining the query information sent by the current client, the query information in the sample set can be traversed, and the similarity between each query information and the query information sent by the current client can be determined. If the similarity reaches the similarity threshold, it can indicate that the similarity between the two is relatively high. At this time, the query information in the sample set that has a high similarity to the query information sent by the current client, and the sub-query information corresponding to this query information, can be used as the corresponding sub-query information of the query information sent by the current client in the inference stage.

[0203] Optionally, an intuition system is used to obtain the representation of the current query and calculate the cosine similarity with the representations of other queries in the set. That is, the identification information of the query information included in the sample set can be traversed. During the traversal, the cosine similarity between each identification information in the sample set and the identification information of the query information sent by the client can be determined. k queries that meet the requirements can be retrieved from the set.

[0204] Optionally, during the process of training the intuition system, SFT has been proven to be effective in understanding the user's intention. Therefore, in the embodiments of the present application, the intuition system can decompose the query Q into sub-problems using context examples. That is, the intuition system can decompose the query information samples of complex problems into corresponding sub-query information samples. Since a generative model can be used as the intuition system, during the training of the intuition system, the loss of the sub-query information samples generated based on the query information samples can be calculated, the loss degree of the sub-query information can be determined, and the intuition system can be trained using the loss degree and the sample set.

[0205] Optionally, during the autoregressive calculation, the generated text that does not include the given context can be used. That is, only the sub-query information samples are used for loss calculation. For example, given a sample of length N, it can be represented by X, where X = {x1,…x i ,…x n}. The sequence length of the context example can be defined as M. The following likelihood function is maximized using the standard language modeling objective: Thus, the intuition system is trained.

[0206] Optionally, the reflection system has a different function from the intuition system. The intuition system relies on quick intuition for generation, while the function of the reflection system is to evaluate the generation results of the intuition system to determine its acceptability. The reflection system can verify the results through two methods: verification of the intermediate process and verification of the overall inference chain.

[0207] Optionally, for the verification of the intermediate process, given the current state S (query: Q and decomposition: D), a reflection system with the same model architecture as the intuitive system can be used to generate a score v for verifying the current state, where v can be expressed as V(f θ ,s)~f θ (v|s).

[0208] Optionally, for the verification of the overall inference chain, given the completed inference chain as S = {s1,…,s i ,…,s n}, a reflection system can be used to generate an overall score o, which can be expressed as O(f θ ,S)~f θ (o|S).

[0209] In the embodiment of the present application, the difference between the reflection system and the intuitive system is that its important task is to evaluate and verify the feasibility of the entire inference chain of the current state, rather than generating quick hypotheses like the intuitive system. This evaluation process helps to determine that the process of generating hypotheses and inferences is reasonable, thus achieving the technical effect of improving the accuracy of information inference by the large model.

[0210] Optionally, in the process of training the reflection system, the same training method as the intuitive system can be adopted, and positive and negative samples can be used to let the model generate classification results therefrom. Since the reflection system mainly focuses on the judgment of the state s, the loss function can be defined as follows:

[0211] For example, Figure 7 is a schematic diagram of a complex task inference process according to an embodiment of the present application, such as Figure 7As shown, if the query information that the client wants to query from the large model is "Wen earns 12 yuan per hour as a babysitter. Yesterday, she only worked as a babysitter for 50 minutes. How much money did she earn?", at this time, the above query information of the client can be sent to the intuition system in the server. The query information can be decomposed by the intuition system according to the reasoning stage. If the reasoning stage includes reasoning stage 1 and reasoning stage 2. Through the intuition system, it can be decomposed into three sub-query information and corresponding sub-answer information in reasoning stage 1, 1a, 1b, and 1c. Among them, the sub-query information and sub-answer information 1a can be "How much does Wen earn per minute? Wen earns 12 / 60 = 0.2 yuan per minute"; the sub-query information and sub-answer information 1b can be "How much does the babysitter charge per hour? The babysitter charges 12 yuan per hour"; the sub-query information and sub-answer information 1c can be "How many hours is 50 minutes equal to? 50 minutes is equal to 1 hour". Through the intuition system, it can be decomposed into three sub-query information and corresponding sub-answer information in reasoning stage 2, 2a, 2b, and 2c. Among them, the sub-query information and sub-answer information 2a can be "How long did Wen work yesterday? Wen worked for 1 hour yesterday"; the sub-query information and sub-answer information 2b can be "How much money did Wen earn? She earned 0.2 x 50 = 10 yuan for working 50 minutes"; the sub-query information and sub-answer information 2c can be "How many minutes did Wen work as a babysitter yesterday? 50 minutes".

[0212] For another example, as Figure 7 shown, after the intuition system decomposes the sub-query information and determines the sub-answer information, the sub-query information and sub-answer information can be transmitted to the reflection system. The reflection system is used to evaluate the sub-answer information and sub-query information determined by the intuition system. For example, for 1a, the reflection system evaluates it as certain; for 1b, the reflection system evaluates it as possible; for 1c, the reflection system evaluates it as impossible. For 2a, the reflection system evaluates it as impossible; for 2b, the reflection system evaluates it as certain; for 2c, the reflection system evaluates it as possible. Through the reflection system, the current state score and the overall score can be determined. That is, through the reflection system, each sub-query information and sub-answer information in each reasoning stage can be evaluated, and the entire reasoning chain of the intuition system can also be evaluated. Thus, if the evaluation shows that the current sub-answer information and sub-query information are poor, it can be returned to the intuition system for re-decomposition. Finally, the answer information to the query information can be obtained as "Wen earned 10 yuan yesterday.", and this answer information can be transmitted through the network to the interaction interface of the corresponding client device for display.

[0213] It should be noted that the preferred implementation schemes involved in the above embodiments of the present application are the same as the schemes, application scenarios, and implementation processes provided in Embodiment 1, but are not limited to the schemes provided in Embodiment 1.

[0214] Embodiment 4

[0215] According to an embodiment of the present application, there is also provided an information inference device for implementing the above Figure 2 information inference method shown.

[0216] Figure 8 FIG. is a schematic diagram of an information inference device according to an embodiment of the present application. As Figure 8 shown, the information inference device 800 may include: a detection unit 802, a processing unit 804, a determination unit 806, a verification unit 808, and an inference unit 810.

[0217] The detection unit 802 is configured to detect query information to be inferred.

[0218] The processing unit 804 is configured to decompose the query information to obtain sub-query information at different inference stages, where there is a logical association relationship between the sub-query information at different inference stages.

[0219] The determination unit 806 is configured to determine sub-response information that respectively matches the sub-query information at different inference stages.

[0220] The verification unit 808 is configured to verify the sub-response information that matches the sub-query information to obtain a verification result corresponding to the sub-query information, where the verification result is used to characterize the feasibility degree of the sub-query information for inferring the query information.

[0221] The inference unit 810 is configured to, in response to the verification result being greater than a verification result threshold, infer the sub-query information corresponding to the verification result to obtain a response information corresponding to the query information.

[0222] It should be noted here that the above detection unit 802, processing unit 804, determination unit 806, verification unit 808, and inference unit 810 correspond to steps S202 to S210 in Embodiment 1. The instances and application scenarios implemented by the five units and the corresponding steps are the same, but are not limited to the content disclosed in the above Embodiment 1. It should be noted that the above units may be hardware components or software components stored in a memory and processed by one or more processors, and the above units may also be part of the device and can run in the computer terminal provided in Embodiment 5.

[0223] According to an embodiment of the present application, there is also provided an information inference device for implementing the above Figure 3 information inference method shown.

[0224] Figure 9 FIG. is a schematic diagram of another information inference device according to an embodiment of the present application. As Figure 9As shown in the figure, the information inference device 900 may include: a first display unit 902, a second display unit 904, and a third display unit 906.

[0225] The first display unit 902 is configured to display, on the dialogue interface, sub-query information at different inference stages of the query information in response to the query information to be inferred received in the dialogue interface, wherein there is a logical association relationship between the sub-query information at different inference stages.

[0226] The second display unit 904 is configured to display, on the dialogue interface, sub-reply information respectively matching the sub-query information at different inference stages.

[0227] The third display unit 906 is configured to display, on the dialogue interface, the reply information corresponding to the query information, wherein the reply information is obtained by inferring the sub-query information corresponding to the verification result when the verification result corresponding to the sub-query information is greater than the verification result threshold, the verification result is obtained by verifying the sub-reply information matching the sub-query information, and is used to characterize the feasibility degree of the sub-query information for inferring the query information.

[0228] It should be noted here that the above-mentioned first display unit 902, second display unit 904, and third display unit 906 correspond to steps S302 to S306 in Embodiment 1. The instances and application scenarios implemented by the three units and the corresponding steps are the same, but are not limited to the content disclosed in the above-mentioned Embodiment 1. It should be noted that the above-mentioned units may be hardware components or software components stored in a memory and processed by one or more processors, and the above-mentioned units may also be part of the device and can run in the computer terminal provided in Embodiment 5.

[0229] According to an embodiment of the present application, there is also provided a model generation device for implementing the Figure 4 model generation method shown above.

[0230] Figure 10 is a schematic diagram of a model generation device according to an embodiment of the present application. As Figure 10 shown in the figure, the model generation device 1000 may include: a first acquisition unit 1002, a first training unit 1004, a second acquisition unit 1006, a third acquisition unit 1008, and a second training unit 1010.

[0231] The first acquisition unit 1002 is configured to acquire query information samples.

[0232] The first training unit 1004 is configured to train a first large model by using sub-query information samples corresponding to query information samples, where the first large model is configured to decompose input query information to obtain sub-query information at different inference stages, and there is a logical association relationship between the sub-query information at different inference stages.

[0233] The second acquisition unit 1006 is configured to acquire reply information samples corresponding to query information samples.

[0234] The third acquisition unit 1008 is configured to acquire verification result samples corresponding to reply information samples, where the verification result samples are used to at least represent the classification results of the reply information samples.

[0235] The second training unit 1010 is configured to train a second large model by using the verification result samples, where the second large model is configured to verify sub-reply information matched with sub-query information to obtain verification results corresponding to the sub-query information, and the verification results are used to characterize the feasibility degree of the sub-query information for inferring query information. The sub-query information corresponding to the verification results greater than the verification result threshold is used to infer the reply information corresponding to the query information.

[0236] It should be noted here that the above first acquisition unit 1002, first training unit 1004, second acquisition unit 1006, third acquisition unit 1008, and second training unit 1010 correspond to steps S402 to S410 in Embodiment 1. The instances and application scenarios implemented by the five units and the corresponding steps are the same, but are not limited to the content disclosed in the above Embodiment 1. It should be noted that the above units may be hardware components or software components stored in a memory and processed by one or more processors, and the above units may also be part of a device and can run in the computer terminal provided in Embodiment 5.

[0237] In the above device, when the client has a need to request the large model to reply to the content to be queried, query information describing the content to be queried can be input on the interaction interface of the client. After detecting the existence of query information in the interaction interface of a certain client, the query information can be transmitted from the client to the server, where the query information is decomposed and processed to determine sub-query information that has a logical association relationship with each other in different inference stages. And each sub-query information can be replied to obtain corresponding sub-reply information. To ensure the feasibility of the determined reply information, each sub-reply information can be verified, and the verification results corresponding to each sub-reply information can be obtained. By judging whether the verification result is greater than the verification result threshold, it is determined whether the query information is feasible. If the verification result is greater than the verification result threshold, it means that the query information has a relatively high feasibility, and the sub-query information can be inferred to determine the reply information of the query information. Since it is considered that in the related art, due to the lack of cognitive ability of the large model, the efficiency of reasoning logical problems with logical relationships and other problems is low, the embodiments of the present application can divide the original query information into multiple sub-query information with logical association relationships according to the sequence of inference stages. By replying to the sub-query information in different inference stages, the reply information of the query information can be finally determined, thereby achieving the purpose of greatly improving the accuracy of the large model in solving problems such as logical reasoning, and further realizing the technical effect of improving the efficiency of information inference, and solving the technical problem of low efficiency of information inference.

[0238] Embodiment 5

[0239] The embodiments of the present application can provide a computer terminal, and the computer terminal can be any computer terminal device in a computer terminal group. Optionally, in this embodiment, the above computer terminal can also be replaced with a terminal device such as a mobile terminal.

[0240] Optionally, in this embodiment, the above computer terminal can be located in at least one of multiple network devices in a computer network.

[0241] In this embodiment, the above computer terminal can execute the program code of the following steps in the information inference method: detecting the query information to be inferred; decomposing and processing the query information to obtain sub-query information in different inference stages, where the sub-query information in different inference stages has a logical association relationship; determining sub-reply information that matches the sub-query information in different inference stages respectively; verifying the sub-reply information that matches the sub-query information to obtain the verification result corresponding to the sub-query information, where the verification result is used to represent the feasibility of the sub-query information for inferring the query information; and in response to the verification result being greater than the verification result threshold, inferring the sub-query information corresponding to the verification result to obtain the reply information corresponding to the query information.

[0242] Optionally, Figure 11 is a structural block diagram of a computer terminal according to an embodiment of the present application. As Figure 11 shown, the computer terminal A may include: one or more (only one is shown in the figure) processors 1102, a memory 1104, a storage controller, and a peripheral interface, where the peripheral interface is connected to a radio frequency module, an audio module, and a display.

[0243] Among them, the memory can be used to store software programs and modules, such as program instructions / modules corresponding to the information inference method and device in the embodiments 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, implements the above-mentioned information inference method. 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 memories. In some instances, the memory may further include a memory remotely provided with respect to the processor, and these remote memories can be connected to the computer terminal A through a network. Examples of the above network include but are not limited to the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof.

[0244] The processor can call the information and application programs stored in the memory through a transmission device to execute the following steps: based on the query information, search for sub-query information at different inference stages in the sample set, where the sample set includes sub-query information corresponding to different query information.

[0245] Optionally, the above processor may further execute the program code of the following steps: determine the identification information of the query information; in the sample set corresponding to the inference stage, determine the sub-query information whose identification information matches the identification information of the query information as the sub-query information of the inference stage.

[0246] Optionally, the above processor may further execute the program code of the following steps: retrieve, in the sample set corresponding to the inference stage, sub-query information whose similarity between the identification information and the identification information of the query information is greater than a similarity threshold; determine the retrieved sub-query information as the sub-query information of the inference stage.

[0247] Optionally, the above processor may further execute the program code of the following steps: analyze the query information using a first large model to obtain the identification information of the query information.

[0248] Optionally, the above processor may further execute the program code of the following steps: analyze the sample set corresponding to the inference stage and the identification information of the query information using a first large model to obtain sub-query information whose identification information matches the identification information of the query information; use the first large model to determine the matching sub-query information as the sub-query information of the inference stage.

[0249] Optionally, the above-mentioned processor may also execute program code for the following steps: obtain a query information sample; input the query information sample into a generation model for analysis to obtain a corresponding sub-query information sample; determine the loss degree of the sub-query information sample; use the loss degree and the sample set to train the generation model to obtain a first large model.

[0250] Optionally, the above-mentioned processor may also execute program code for the following steps: use the first large model to analyze the sub-query information at different inference stages to obtain sub-response information that matches the sub-query information at different inference stages.

[0251] Optionally, the above-mentioned processor may also execute program code for the following steps: when running to the current inference stage among different inference stages, verify the sub-response information that matches the sub-query information at the current inference stage to obtain the sub-verification result at the current inference stage, where the verification result includes the sub-verification result; or, verify the sub-response information at different inference stages to obtain the overall verification result at different inference stages, where the verification result includes the overall verification result.

[0252] Optionally, the above-mentioned processor may also execute program code for the following steps: use a second large model to verify the sub-response information at the current inference stage to obtain a sub-verification result; use the second large model to verify the sub-response information at different inference stages to obtain an overall verification result.

[0253] Optionally, the above-mentioned processor may also execute program code for the following steps: obtain a query information sample and a corresponding response information sample; input the query information sample and the response information sample into a generation model for analysis to obtain a corresponding verification result sample; determine the loss degree of the verification result sample; use the verification result sample to train the generation model to obtain a second large model.

[0254] Optionally, the above-mentioned processor may also execute program code for the following steps: in the sample set corresponding to the current inference stage, search for the sub-query information at the current inference stage based on the query information; when different inference stages include the next inference stage of the current inference stage, in the sample set corresponding to the next inference stage, search for the sub-query information at the next inference stage based on the sub-verification result and the query information at the current inference stage.

[0255] Optionally, the above-mentioned processor may also execute program code for the following steps: determine the query information as the root node and the sub-query information as the leaf node to establish a target tree structure; perform inference on the sub-query information corresponding to the verification result to obtain the response information corresponding to the query information, including: in the target tree structure, perform inference on the sub-query information corresponding to the verification result to obtain the response information corresponding to the query information.

[0256] As another alternative example, the processor can call the information and application programs stored in the memory through the transmission device to execute the following steps: in response to the query information to be inferred received in the dialogue interface, on the dialogue interface, display the sub-query information of the query information at different inference stages, where there is a logical association relationship between the sub-query information at different inference stages; on the dialogue interface, display the sub-reply information respectively matching the sub-query information at different inference stages; on the dialogue interface, display the reply information corresponding to the query information, where the reply information is obtained by inferring the sub-query information corresponding to the verification result when the verification result corresponding to the sub-query information is greater than the verification result threshold, the verification result is obtained by verifying the sub-reply information matching the sub-query information, and is used to represent the feasibility degree of the sub-query information for inferring the query information.

[0257] Optionally, the above-mentioned processor can also execute the program code of the following steps: the query information is multi-modal information, and the types of the multi-modal information include at least one of the following: text information containing character information, video frame information containing frame image information, audio information, and the types of the reply information include at least one of the following: text information, image information, video information, and voice information.

[0258] As another alternative example, the processor can call the information and application programs stored in the memory through the transmission device to execute the following steps: obtain a query information sample; use the sub-query information sample corresponding to the query information sample to train and obtain a first large model, where the first large model is used to decompose the input query information to obtain the sub-query information at different inference stages, and there is a logical association relationship between the sub-query information at different inference stages; obtain the reply information sample corresponding to the query information sample; obtain the verification result sample corresponding to the reply information sample, where the verification result sample is used to at least represent the classification result of the reply information sample; use the verification result sample to train and obtain a second large model, where the second large model is used to verify the sub-reply information matching the sub-query information to obtain the verification result corresponding to the sub-query information, and the verification result is used to represent the feasibility degree of the sub-query information for inferring the query information, and the sub-query information corresponding to the verification result greater than the verification result threshold is used to infer the reply information corresponding to the query information.

[0259] Optionally, the above-mentioned processor can also execute the program code of the following steps: determine the loss degree of the sub-query information sample; use the loss degree and the sample set to train and generate a model to obtain a first large model, where the sample set includes the sub-query information corresponding to different query information, and the generation model is a pre-trained generative large model.

[0260] Optionally, the above-mentioned processor may also execute the program code of the following steps: determining the loss degree of the verification result sample; training a generation model by using the loss degree of the verification result sample to obtain a second large model, where the generation model is a pre-trained generative large model.

[0261] By adopting the embodiments of the present application, when a client has a need to request a large model to reply to the content to be queried, query information describing the content to be queried may be input on the interaction interface of the client. After detecting that there is query information in the interaction interface of a certain client, the query information may be transmitted from the client to the server, and the query information is decomposed and processed in the server to determine sub-query information that has a logical association relationship with each other in different inference stages. And the respective sub-query information may be replied to obtain corresponding sub-reply information. In order to ensure the feasibility degree of the determined reply information, each sub-reply information may be verified, and the verification results corresponding to each sub-reply information. By determining whether the verification result is greater than the verification result threshold, it is determined whether the query information is feasible. If the verification result is greater than the verification result threshold, it indicates that the query information has a relatively high feasibility degree, and the sub-query information may be inferred to determine the reply information of the query information. Since it is considered that in the related art, due to the lack of cognitive ability of the large model, the efficiency of inferring logical problems with logical relationships and other problems is low, the embodiments of the present application may divide the original query information into multiple sub-query information with logical association relationships according to the sequence of the inference stages. By replying to the sub-query information in different inference stages, the reply information of the query information may be finally determined, thereby achieving the purpose of greatly improving the accuracy of the large model in solving problems such as logical reasoning, and further realizing the technical effect of improving the efficiency of information inference, and solving the technical problem of low efficiency of information inference.

[0262] Those of ordinary skill in the art can understand that Figure 11 The structure shown is only for illustration, and the computer terminal may also be a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a handheld computer, and a mobile Internet device (Mobile Internet Devices, MID), a PAD and other terminal devices. Figure 11 It does not limit the structure of the above-mentioned electronic device. For example, computer terminal A may also include more or fewer components (such as a network interface, a display device, etc.) than those shown Figure 11 in the figure, or have a different configuration from that shown Figure 11 in the figure.

[0263] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by a program instructing the relevant hardware of the terminal device. The program can be stored in a computer-readable storage medium, and the storage medium can include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, an optical disk, etc.

[0264] Embodiment 6

[0265] An embodiment of the present application further provides a computer-readable storage medium. Optionally, in this embodiment, the above computer-readable storage medium can be used to store the program code executed by the information inference method provided in the first embodiment above.

[0266] Optionally, in this embodiment, the above storage medium can be located in any one of the computer terminals in the computer terminal group in the computer network, or in any one of the mobile terminals in the mobile terminal group.

[0267] Optionally, in this embodiment, the storage medium is set to store program code for performing the following steps: detecting query information to be inferred; decomposing the query information to obtain sub-query information at different inference stages, where there is a logical association relationship between the sub-query information at different inference stages; determining sub-response information respectively matching the sub-query information at different inference stages; verifying the sub-response information matching the sub-query information to obtain a verification result corresponding to the sub-query information, where the verification result is used to characterize the feasibility degree of the sub-query information for inferring the query information; and in response to the verification result being greater than the verification result threshold, inferring the sub-query information corresponding to the verification result to obtain a response information corresponding to the query information.

[0268] Optionally, in this embodiment, the storage medium is set to store program code for performing the following steps: in response to the query information to be inferred received in the dialogue interface, on the dialogue interface, display the sub-query information of the query information at different inference stages, where there is a logical association relationship between the sub-query information at different inference stages; on the dialogue interface, display the sub-response information respectively matching the sub-query information at different inference stages; and on the dialogue interface, display the response information corresponding to the query information, where the response information is obtained by inferring the sub-query information corresponding to the verification result in the case that the verification result corresponding to the sub-query information is greater than the verification result threshold, the verification result is obtained by verifying the sub-response information matching the sub-query information, and is used to characterize the feasibility degree of the sub-query information for inferring the query information.

[0269] Optionally, in this embodiment, the storage medium is configured to store program code for performing the following steps: obtaining a query information sample; training a first large model by using sub-query information samples corresponding to the query information sample, where the first large model is used to decompose the input query information to obtain sub-query information at different inference stages, and there is a logical association relationship between the sub-query information at different inference stages; obtaining a reply information sample corresponding to the query information sample; obtaining a verification result sample corresponding to the reply information sample, where the verification result sample is used to at least represent the classification result of the reply information sample; training a second large model by using the verification result sample, where the second large model is used to verify the sub-reply information matched with the sub-query information to obtain a verification result corresponding to the sub-query information, and the verification result is used to characterize the feasibility degree of the sub-query information for inferring the query information, and the sub-query information corresponding to the verification result greater than the verification result threshold is used to infer the reply information corresponding to the query information.

[0270] In an embodiment of the present application, when a client has a need to request a large model to reply to the content to be queried, query information describing the content to be queried can be input on the interaction interface of the client. After detecting that there is query information in the interaction interface of a certain client, the query information can be transmitted from the client to the server, and the query information is decomposed and processed in the server to determine sub-query information with a logical association relationship between different inference stages. And sub-reply information can be obtained by replying to each sub-query information. To ensure the feasibility degree of the determined reply information, each sub-reply information can be verified, and the verification result corresponding to each sub-reply information. By judging whether the verification result is greater than the verification result threshold, it is determined whether the query information is feasible. If the verification result is greater than the verification result threshold, it means that the query information has a high feasibility degree, and the sub-query information can be inferred to determine the reply information corresponding to the query information. Since it is considered that in the related art, due to the lack of cognitive ability of the large model, the efficiency of inferring logical problems with logical relationships and other problems is low, in the embodiment of the present application, the original query information can be divided into multiple sub-query information with a logical association relationship according to the sequence of the inference stages. By replying to the sub-query information at different inference stages, the reply information corresponding to the query information can be finally determined, thereby achieving the purpose of greatly improving the accuracy of the large model in solving logical reasoning and other problems, and further realizing the technical effect of improving the efficiency of information inference, and solving the technical problem of low efficiency of information inference.

[0271] Embodiment 7

[0272] An embodiment of the present application can provide an electronic device, and the electronic device can include a memory and a processor. Figure 12It is a block diagram of an electronic device for an information inference method according to an embodiment of the present application. The electronic device is intended to represent various forms of digital computers, such as, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as, personal digital processors, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present application described and / or claimed herein.

[0273] As Figure 12 shown, the device 1200 includes a computing unit 1201, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1202 or a computer program loaded from a storage unit 1208 into a random access memory (RAM) 1203. In the RAM 1203, various programs and data required for the operation of the device 1200 can also be stored. The computing unit 1201, the ROM 1202, and the RAM 1203 are connected to each other via a bus 1204. An input / output (I / O) interface 1205 is also connected to the bus 1204.

[0274] A plurality of components in the device 1200 are connected to the I / O interface 1205, including: an input unit 1206, such as a keyboard, a mouse, etc.; an output unit 1604, such as various types of displays, speakers, etc.; a storage unit 1208, such as a magnetic disk, an optical disk, etc.; and a communication unit 1209, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1209 allows the device 1200 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0275] The computing unit 1201 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1201 include, but are not limited to, a Central Processing Unit (CPU), a Graphic Processing Unit (GPU), various dedicated Artificial Intelligence (AI) computing chips, various computing units running machine learning model algorithms, a Digital Signal Processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1201 executes the various methods and processes described above, such as the information inference method. For example, in some embodiments, the information inference method can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 1208. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device 1200 via the ROM 1202 and / or the communication unit 1209. When the computer program is loaded into the RAM 1203 and executed by the computing unit 1201, one or more steps of the information inference method described above can be executed. Alternatively, in other embodiments, the computing unit 1201 can be configured to execute the information inference method by any other suitable means (e.g., by means of firmware).

[0276] The various embodiments of the systems and techniques described above in this specification can be implemented in digital electronic circuitry, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard parts (ASSPs), system on chip (SOC) systems, complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include: being implemented in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be a special-purpose or general-purpose programmable processor that receives data and instructions from, and transmits data and instructions to, a storage system, at least one input device, and at least one output device.

[0277] The program code for implementing the methods of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the program codes, when executed by the processor or controller, cause the functions / operations specified in the flowchart and / or block diagram to be implemented. The program code can be executed entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine, or entirely on the remote machine or server.

[0278] In the context of this application, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0279] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a cathode ray tube or a liquid crystal display, a monitor); and a keyboard and a pointing device (e.g., a mouse or a trackball) through which the user can provide input to the computer. Other kinds of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0280] The systems and techniques described herein can be implemented in a computing system including backend components (e.g., as a data server), or a computing system including middleware components (e.g., an application server), or a computing system including frontend components (e.g., a user computer having a graphical user interface or a web browser through which the user can interact with an implementation of the systems and techniques described herein), or a computing system including any combination of such backend components, middleware components, or frontend components. The components of the system can be interconnected to each other by digital data communication in any form or medium (e.g., a communication network). Examples of communication networks include: local area networks, wide area networks, and the Internet.

[0281] A computer system can include a client device and a server. The client device and the server are generally remote from each other and typically interact through a communication network. The relationship between the client device and the server is created by computer programs running on the respective computers and having a client device-server relationship with each other. The server can be a cloud server, or a server of a distributed system, or a server incorporating a blockchain.

[0282] It should be noted that the serial numbers of the above embodiments of the present application are only for description and do not represent the advantages or disadvantages of the embodiments.

[0283] In the above embodiments of the present application, the descriptions of the various embodiments have their own emphases. For parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0284] In several embodiments provided by the present application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are merely illustrative. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods. 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 displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of units or modules can be in electrical or other forms.

[0285] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place, or they can be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0286] In addition, each functional unit in various embodiments of the present application can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0287] 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 application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in various embodiments of the present application. And the aforementioned storage medium includes: various media such as USB flash drives, read-only memories, random access memories, mobile hard disks, magnetic disks, or optical discs that can store program codes.

[0288] The above is only the preferred embodiment of the present application. It should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present application, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. An information inference method, characterized in that, Including: Detecting query information to be inferred; Decomposing and processing the query information to obtain sub-query information at different inference stages, where there is a logical association relationship between the sub-query information at different inference stages; Determining sub-response information respectively matching the sub-query information at different inference stages; Verifying the sub-response information matching the sub-query information to obtain a verification result corresponding to the sub-query information, where the verification result is used to characterize the feasibility degree of the sub-query information for inferring the query information; In response to the verification result being greater than the verification result threshold, inferring the sub-query information corresponding to the verification result to obtain a response information corresponding to the query information.

2. The method according to claim 1, wherein Decomposing and processing the query information to obtain sub-query information at different inference stages, including: Based on the query information, searching for the sub-query information at different inference stages in the sample set, where the sample set includes the sub-query information corresponding to different query information.

3. The method according to claim 2, wherein Based on the query information, searching for the sub-query information at different inference stages in the sample set, including: Determining the identification information of the query information; In the sample set corresponding to the inference stage, determining the sub-query information whose identification information matches the identification information of the query information as the sub-query information of the inference stage.

4. The method according to claim 3, wherein In the sample set corresponding to the inference stage, determining the sub-query information whose identification information matches the identification information of the query information as the sub-query information of the inference stage, including: In the sample set corresponding to the inference stage, retrieving sub-query information whose similarity between the identification information and the identification information of the query information is greater than the similarity threshold; Determining the retrieved sub-query information as the sub-query information of the inference stage.

5. The method according to claim 3, wherein Determining the identification information of the query information, including: Using a first large model to analyze the query information to obtain the identification information of the query information.

6. The method according to claim 5, wherein In the sample set corresponding to the inference stage, determining the sub-query information whose identification information matches the identification information of the query information as the sub-query information of the inference stage, including: Using the first large model to analyze the sample set corresponding to the inference stage and the identification information of the query information to obtain sub-query information whose identification information matches the identification information of the query information; Using the first large model to determine the matching sub-query information as the sub-query information of the inference stage.

7. The method according to claim 5, characterized in that, Determining sub-response information respectively matching the sub-query information at different inference stages, including: Using the first large model to analyze the sub-query information at different inference stages to obtain sub-response information matching the sub-query information at different inference stages.

8. The method according to claim 1, characterized in that Verifying the sub-response information matching the sub-query information to obtain a verification result corresponding to the sub-query information, including: When running to the current inference stage among the different inference stages, verify the sub-response information matching the sub-query information of the current inference stage to obtain the sub-verification result of the current inference stage, where the verification result includes the sub-verification result; or, Verify the sub-response information of the different inference stages to obtain the overall verification result of the different inference stages, where the verification result includes the overall verification result.

9. The method according to claim 8, characterized in that, Verifying the sub-response information of the current inference stage among the different inference stages to obtain the sub-verification result of the current inference stage includes: Using the second largest model to verify the sub-response information of the current inference stage to obtain the sub-verification result; Verifying the sub-response information of the different inference stages to obtain the overall verification result of the different inference stages includes: using the second largest model to verify the sub-response information of the different inference stages to obtain the overall verification result.

10. The method according to claim 8, wherein Decompose the query information to obtain sub-query information of different inference stages, including: In the sample set corresponding to the current inference stage, search for the sub-query information of the current inference stage based on the query information; When the next inference stage of the current inference stage is included in the different inference stages, in the sample set corresponding to the next inference stage, search for the sub-query information of the next inference stage based on the sub-verification result of the current inference stage and the query information.

11. The method according to any one of claims 1 to 10, characterized in that, The method further includes: Determine the query information as the root node and the sub-query information as the leaf node to establish a target tree structure; Infer the response information corresponding to the query information from the sub-query information corresponding to the verification result, including: in the target tree structure, infer the response information corresponding to the query information from the sub-query information corresponding to the verification result.

12. An information inference method, characterized in that, Includes: In response to the query information to be inferred received on the dialogue interface, display the sub-query information of the query information at different inference stages on the dialogue interface, where there is a logical association relationship between the sub-query information at different inference stages; On the dialogue interface, display the sub-response information respectively matching the sub-query information of the different inference stages; Display the response information corresponding to the query information on the dialogue interface, where the response information is obtained by inferring the sub-query information corresponding to the verification result when the verification result corresponding to the sub-query information is greater than the verification result threshold, the verification result is obtained by verifying the sub-response information matching the sub-query information, and is used to characterize the feasibility of the sub-query information for inferring the query information.

13. The method according to claim 12, wherein The query information is multimodal information, and the types of the multimodal information include at least one of the following: text information containing character information, video frame information containing frame image information, audio information, and the types of the response information include at least one of the following: text information, image information, video information, and voice information.

14. A method for generating a model, characterized in that, Includes: Obtain a query information sample; Training a first large model using the sub-query information samples corresponding to the query information samples, where the first large model is used to decompose the input query information to obtain sub-query information at different inference stages, and there is a logical association relationship between the sub-query information at different inference stages; Obtaining the reply information samples corresponding to the query information samples; Obtaining the verification result samples corresponding to the reply information samples, where the verification result samples are used to at least represent the classification results of the reply information samples; Training a second large model using the verification result samples, where the second large model is used to verify the sub-reply information matched by the sub-query information to obtain the verification result corresponding to the sub-query information, and the verification result is used to characterize the feasibility of the sub-query information for inferring the query information, and the sub-query information corresponding to the verification result greater than the verification result threshold is used to infer the reply information corresponding to the query information.

15. The method according to claim 14, wherein Training a first large model using the sub-query information samples corresponding to the query information samples includes: Determining the loss degree of the sub-query information samples; Training a generation model using the loss degree and a sample set to obtain the first large model, where the sample set includes sub-query information corresponding to different query information, and the generation model is a pre-trained generative large model.

16. The method according to claim 14, wherein Training a second Large model includes: Determining the loss degree of the verification result samples; Training a generation model using the loss degree of the verification result samples to obtain the second large model, where the generation model is a pre-trained generative large model.

17. An information inference system, characterized in that, Includes: An information generation end, configured to detect the query information to be inferred, decompose the query information to obtain sub-query information at different inference stages, and determine sub-reply information respectively matched with the sub-query information at different inference stages, where there is a logical association relationship between the sub-query information at different inference stages; An information verification end, configured to verify the sub-reply information matched by the sub-query information to obtain the verification result corresponding to the sub-query information, and in response to the verification result being greater than the verification result threshold, infer the sub-query information corresponding to the verification result to obtain the reply information corresponding to the query information, where the verification result is used to characterize the feasibility of the sub-query information for inferring the query information.

18. An electronic device, characterized in that, Includes: A memory and a processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, and when the computer-executable instructions are executed by the processor, the steps of the method according to any one of claims 1 to 16 are implemented.

Images