CN-122025173-A - Information generation method, training method, device and intelligent agent based on large model

Abstract

The disclosure provides an information generation method, a training device and an intelligent agent based on a large model, relates to the technical field of artificial intelligence, and particularly relates to the technical fields of AI medical treatment, intelligent agent, large model and the like. The information generation method based on the large model is specifically realized by analyzing inquiry information of a target object by using the large model to obtain a target reasoning mode, wherein the target reasoning mode indicates the complexity of a reasoning task required to be executed for obtaining an inquiry result based on the inquiry information, and the inquiry information is reasoning by using the large model according to the target reasoning mode to generate target information for describing the inquiry result.

Inventors

• Li Lvxue

Assignees

• 北京百度网讯科技有限公司

Dates

Publication Date

20260512

Application Date

20260129

Claims (20)

1. 1. An information generation method based on a large model comprises the following steps: analyzing inquiry information of a target object by utilizing a large model to obtain a target reasoning mode, wherein the target reasoning mode indicates the complexity of a reasoning task required to be executed for obtaining an inquiry result based on the inquiry information, and And according to the target reasoning mode, reasoning the inquiry information by using the large model, and generating target information for describing the inquiry result.
2. 2. The method of claim 1, wherein the inquiry information includes status description information and examination result information for determining a disease type; the analyzing the inquiry information of the target object by using the large model to obtain a target reasoning mode comprises the following steps: Analyzing the state description information of the target object and the examination result information by using a large model to obtain a first association degree, wherein the first association degree indicates whether the state of the target object is a characteristic state for determining the disease type in the examination result information, and And determining the target reasoning mode according to the first association degree.
3. 3. The method of claim 2, wherein the determining the target inference mode from the first degree of association comprises: in response to determining that the first degree of association indicates that the state of the target object is a characteristic state, determining that the target inference mode is a first inference mode, and And determining the target inference mode as a second inference mode in response to determining that the first degree of association indicates that the state of the target object is a non-characteristic state, wherein the second inference mode has a complexity greater than that of the first inference mode.
4. 4. A method according to claim 2 or 3, wherein the analyzing the inquiry information of the target object by using the large model to obtain the target inference mode further comprises: Acquiring historical case information associated with the inquiry information; Analyzing the historical case information, the state description information and the examination result information by using the large model to obtain a second association degree, wherein the second association degree indicates the probability that the state of the target object is determined to be the disease type in the examination result information in the historical case, and And determining the target reasoning mode according to the first association degree and the second association degree.
5. 5. The method of claim 4, wherein the determining the target inference mode from the first degree of association and the second degree of association comprises: Determining that the target inference mode is a third inference mode in response to determining that the first degree of association indicates that the state of the target object is a non-feature state and the second degree of association is greater than a predetermined association threshold, wherein the third inference mode has a complexity greater than the first inference mode and less than the second inference mode, and And in response to determining that the first degree of association indicates that the state of the target object is a non-characteristic state and the second degree of association is less than or equal to the predetermined association threshold, determining that the target inference mode is the second inference mode.
6. 6. The method of any of claims 1-5, wherein said reasoning about the interview information using the large model in accordance with the target reasoning model to generate target information describing the interview result, comprising: determining an inference path matched with the target inference mode, wherein the inference path indicates a logical association relationship between a plurality of inference tasks to be executed to obtain the inquiry result based on the inquiry information, and And according to the logical association relation, utilizing the large model to infer the inquiry information by executing a plurality of inference tasks, and generating the target information.
7. 7. The method of claim 6, wherein said generating said target information by reasoning about said questioning information by performing a plurality of said reasoning tasks using said large model in accordance with said logical association comprises: According to the logical association relation, utilizing the large model to infer the inquiry information by executing a plurality of inference tasks to generate an intermediate result, wherein the intermediate result represents an initial disease type; Performing correlation analysis on the initial disease type and the inquiry information by using the large model to obtain the matching degree between the initial disease type and the inquiry information, and And generating the target information according to the matching degree.
8. 8. The method of claim 7, wherein the inquiry information includes status description information, the degree of matching includes a first degree of matching; Performing association analysis on the initial disease type and the inquiry information by using the large model to obtain the matching degree between the initial disease type and the inquiry information, wherein the method comprises the following steps: acquiring expected state information associated with the initial disease type; and carrying out association analysis on the expected state information and the state description information by using the large model to obtain a first matching degree between the expected state information and the state description information.
9. 9. The method of claim 8, wherein the inquiry information further comprises historical visit information, the degree of matching further comprising a second degree of matching; performing association analysis on the initial disease type and the inquiry information by using the large model to obtain the matching degree between the initial disease type and the inquiry information, and further comprising: And performing conflict analysis on expected state information associated with the initial disease type and the historical visit information by using the large model to obtain a second matching degree between the expected state information and the historical state in the historical visit information.
10. 10. The method according to any one of claims 7-9, wherein the generating the target information according to the degree of matching comprises: Determining the intermediate result as the target information in response to determining that the degree of matching is greater than a predetermined matching threshold; In response to determining that the degree of matching is less than or equal to the predetermined matching threshold, revising the inference path based on the degree of matching using the large model, and And utilizing the large model to infer the inquiry information according to the revised inference path, and generating the target information.
11. 11. The method of claim 10, wherein the generating the target information according to the degree of matching further comprises: Determining the credibility of the expected state information according to the source of the expected state information; Correcting the matching degree according to the credibility, and And generating the target information according to the corrected matching degree.
12. 12. The method of any of claims 6-11, wherein a plurality of the inference tasks includes N tasks associated with each other, N being an integer greater than 1; The step of utilizing the large model to infer the inquiry information by executing a plurality of reasoning tasks according to the logical association relation to generate the target information comprises the following steps: The method comprises the steps of utilizing the large model to carry out an N-1 reasoning task to infer an N-1 intermediate result to generate an N-1 intermediate result, wherein the N-1 intermediate result is obtained by utilizing the large model through carrying out the N-1 reasoning task, and N is an integer which is more than 1 and less than or equal to N-1; in response to determining that the correlation between the nth intermediate result and the nth-1 intermediate result is less than or equal to a predetermined correlation threshold, returning to perform the nth inference task, and In response to determining that a correlation between the nth intermediate result and the nth-1 intermediate result is greater than a predetermined correlation threshold, inferring the nth intermediate result by performing an nth+1 inference task using the large model, generating an nth+1 intermediate result.
13. 13. A model training method, comprising: Analyzing sample inquiry information of a sample object by using an initial large model to obtain a sample reasoning mode, wherein the sample reasoning mode indicates the complexity of a sample reasoning task required to be executed for obtaining a sample inquiry result based on the sample inquiry information; reasoning the sample inquiry information by using the initial large model according to the sample reasoning mode to generate sample information for describing the sample inquiry result, and Training the initial large model by using the sample information and the sample label based on an objective function to obtain a trained large model; Wherein the sample tag is generated based on the sample interrogation information by invoking at least two pre-trained language models.
14. 14. The method of claim 13, further comprising: generating a first sample reasoning result based on the sample inquiry information by using a first pre-training language model; generating a second sample reasoning result and a sample reasoning process based on the sample inquiry information by using a second pre-training language model; in response to determining that the first sample inference result is the same as the second sample inference result, refining the sample inference process with the second pre-training language model to generate an inference digest, and determining the inference digest as a sample tag of a first inference pattern, and And in response to determining that the first sample inference result is different from the second sample inference result, determining the sample inference process as a sample tag for a second inference mode, wherein the first inference mode has a complexity that is less than a complexity of the second inference mode.
15. 15. The method of claim 14, further comprising: performing a correlation analysis on the sample inquiry information and the reasoning abstract by using the second pre-training language model to obtain a third correlation degree between the sample inquiry information and the reasoning abstract, and And in response to determining that the third degree of association is greater than a predetermined association threshold, determining that the inferential summary is the sample tag.
16. 16. The method of any of claims 13-15, wherein the objective function comprises a loss function and a reward function; The training of the initial large model based on the objective function by using the sample information and the sample label to obtain a trained large model comprises the following steps: Training the initial large model based on the loss function by using the sample information and the sample label to obtain an intermediate large model, and And based on the reward function, performing reinforcement learning training on the middle large model by utilizing the sample information and the sample label to obtain the trained large model.
17. 17. An information generating apparatus based on a large model, comprising: The first analysis module is used for analyzing the inquiry information of the target object by utilizing the large model to obtain a target reasoning mode, wherein the target reasoning mode indicates the complexity of a reasoning task required to be executed for obtaining the inquiry result based on the inquiry information, and The first reasoning module is used for reasoning the inquiry information by utilizing the large model according to the target reasoning mode, and generating target information for describing the inquiry result.
18. 18. A model training apparatus comprising: The system comprises a first analysis module, a second analysis module and a first analysis module, wherein the first analysis module is used for analyzing sample inquiry information of a sample object by utilizing an initial large model to obtain a sample reasoning mode, and the sample reasoning mode indicates the complexity of a sample reasoning task required to be executed for obtaining a sample inquiry result based on the sample inquiry information; A second reasoning module for reasoning the sample inquiry information by using the initial big model according to the sample reasoning mode to generate sample information for describing the sample inquiry result, and The training module is used for training the initial large model by utilizing the sample information and the sample label based on an objective function to obtain a trained large model; Wherein the sample tag is generated based on the sample interrogation information by invoking at least two pre-trained language models.
19. 19. An agent, comprising: the input module is used for receiving inquiry information of the target object; a processing module for determining a target task based on the inquiry information of the target object received by the input module, determining a target big model based on the target task, obtaining target information describing the inquiry result by calling the target big model to execute the method of any one of claims 1-16, and And the output module is used for outputting the target information obtained by the processing module.
20. 20. An electronic device, comprising: at least one processor, and A memory communicatively coupled to the at least one processor, wherein, The memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-16.

Description

Information generation method, training method, device and intelligent agent based on large model Technical Field The disclosure relates to the technical field of artificial intelligence, in particular to the technical fields of AI medical treatment, an agent, a large model and the like, and specifically relates to an information generation method, a training device and an agent based on the large model. Background With the deep application of artificial intelligence technology in the medical auxiliary diagnosis scene, the requirement of users on the accuracy of the AI (ARTIFICIAL INTELLIGENCE ) auxiliary diagnosis result is higher and higher. However, the fixed and mechanical reasoning process not only reduces the utilization rate of computing resources, but also increases the instability of the reasoning diagnosis result. Disclosure of Invention The disclosure provides an information generation method, a training device and an intelligent agent based on a large model. According to one aspect of the disclosure, an information generation method based on a large model is provided, and the information generation method comprises the steps of analyzing inquiry information of a target object by using the large model to obtain a target reasoning mode, wherein the target reasoning mode indicates the complexity of a reasoning task required to be executed for obtaining an inquiry result based on the inquiry information, and reasoning the inquiry information by using the large model according to the target reasoning mode to generate target information for describing the inquiry result. According to another aspect of the disclosure, a model training method is provided, which comprises the steps of analyzing sample inquiry information of a sample object by using an initial large model to obtain a sample reasoning mode, wherein the sample reasoning mode indicates the complexity of sample reasoning tasks required to be executed for obtaining sample inquiry results based on the sample inquiry information, reasoning the sample inquiry information by using the initial large model according to the sample reasoning mode to generate sample information for describing the sample inquiry results, and training the initial large model by using sample information and sample labels based on an objective function to obtain a trained large model, and the sample labels are generated by calling at least two pre-training language models based on the sample inquiry information. According to another aspect of the present disclosure, there is provided an information generating apparatus based on a large model, including a first analysis module and a first inference module. The first analysis module is used for analyzing the inquiry information of the target object by utilizing the large model to obtain a target reasoning mode, wherein the target reasoning mode indicates the complexity of a reasoning task required to be executed for obtaining the inquiry result based on the inquiry information. The first reasoning module is used for reasoning the inquiry information by utilizing the big model according to the target reasoning mode, and generating target information for describing the inquiry result. According to another aspect of the present disclosure, there is provided a model training apparatus including a second analysis module, a second reasoning module, and a training module. The second analysis module is used for analyzing the sample inquiry information of the sample object by utilizing the initial large model to obtain a sample reasoning mode, wherein the sample reasoning mode indicates the complexity of a sample reasoning task required to be executed for obtaining a sample inquiry result based on the sample inquiry information. And the second reasoning module is used for reasoning the sample inquiry information by using the initial large model according to the sample reasoning mode and generating sample information for describing the sample inquiry result. The training module is used for training the initial large model by utilizing sample information and sample labels based on the objective function to obtain a trained large model, wherein the sample labels are generated by calling at least two pre-training language models based on sample inquiry information. According to another aspect of the present disclosure, an agent is provided that includes an input module, a processing module, and an output module. And the input module is used for receiving the inquiry information of the target object. The processing module is used for determining a target task based on the inquiry information of the target object received by the input module, determining a target big model based on the target task, and executing the information generation method based on the big model by calling the target big model to obtain target information for describing the inquiry result. And the output module is used for outputting the target information obtained b