CN-122021890-A - Question answering method, device, equipment and storage medium

CN122021890ACN 122021890 ACN122021890 ACN 122021890ACN-122021890-A

Abstract

The application belongs to the technical field of artificial intelligence, and relates to a problem solving method, device, equipment and storage medium, wherein before a unitized processing result corresponding to solving resources is transmitted to a plurality of solving layers contained in a target solving data generating model, the unitized processing result is combined by utilizing an iterative combining mechanism, so that the reasoning workload of the target solving data generating model is reduced, and in the plurality of solving layers, the problem solving knowledge integration of a front layer and a back layer is reduced under the condition that the solution resource integration is not reduced as a whole by introducing a progressive pruning fusion mechanism, and the problem solving fusion efficiency of the target solving data generating model is improved. The method is applied to business consultation scenes in the field of financial science and technology or health care, and can be used for effectively fusing solution information without additional training and timely feeding solution data back to a problem consultation terminal.

Inventors

WANG JIANZONG
ZHANG XULONG
XIE JUNFEI

Assignees

平安科技（深圳）有限公司

Dates

Publication Date: 20260512
Application Date: 20260119

Claims (10)

1. A method of solving a problem, comprising the steps of: Acquiring target questions and answering resources; Performing unitization processing on the target problem and the solution resource respectively to obtain a first processing unit set corresponding to the target problem and a second processing unit set corresponding to the solution resource; Adopting an iterative merging mechanism to merge elements in the second processing unit set until the total number of the remaining elements in the second processing unit set meets a preset target retention rate, and obtaining a latest second processing unit set; transmitting the latest second processing unit set and the first processing unit set to a plurality of solution layers contained in a target solution data generation model; According to a progressive pruning fusion mechanism, fusing the elements in the latest second processing unit set and the solution information of the elements in the first processing unit set in the solution layers; and finally fusing the solution results of the solution fusion information respectively obtained by the solution layers, and outputting the finally fused solution results at the target output layer.
2. The method of claim 1, wherein the step of performing unitization processing on the target problem and the solution resource to obtain a first set of processing units corresponding to the target problem and a second set of processing units corresponding to the solution resource, specifically includes: Performing text decomposition on the target problem to obtain word and/or word combination after the target problem is decomposed as the first processing unit set; And decomposing the solution resource to obtain a resource information combination of the decomposed solution resource as the second processing unit set, wherein the resource information combination comprises a video frame and/or picture area and/or a resource combination in the form of words and/or characters.
3. The method for solving the problem according to claim 1, wherein the step of merging the elements in the second processing unit set by using an iterative merging mechanism until the total number of the remaining elements in the second processing unit set meets a preset target retention rate, includes the steps of: Counting the number of elements in the second processing unit set as an initial number; Sequentially comparing the similarity between the elements in the second processing unit set by adopting an iterative comparison mode to obtain a similarity comparison result; based on the similarity comparison result, merging the elements with the maximum similarity in the second processing unit set pairwise; Counting the total number of the residual elements in the second processing unit set after the merging of the round; Calculating a ratio of the total number of the remaining elements to the initial number; and determining the latest second processing unit set according to the ratio and the preset target retention rate.
4. The method of claim 3, wherein the step of sequentially comparing the similarities of the elements in the second processing unit set by using an iterative comparison method to obtain a similarity comparison result further comprises: identifying whether the solution resource is a video resource; And if the solution resource is a video resource, using a video stream time stamp as a comparison control parameter when the elements in the second processing unit set are subjected to similarity comparison, and performing comparison between continuous frames or the same frames to obtain a similarity comparison result.
5. A method of solving a problem according to claim 3, wherein the step of determining the latest second set of processing units according to the ratio and the preset target retention rate specifically includes: identifying the magnitude relation between the ratio and the preset target retention rate; If the ratio exceeds the preset target retention rate, continuing element combination of the second processing unit set after the combination of the round until the ratio does not exceed the preset target retention rate, and stopping continuing the combination; and if the ratio does not exceed the preset target retention rate, the second processing unit set after the round of merging is the latest second processing unit set.
6. The method of claim 1, wherein the plurality of solution layers are set as an early solution layer, an intermediate solution layer, and a later solution layer according to an inference precedence relationship, and before the step of fusing solution information possessed by the elements in the latest second set of processing units and the elements in the first set of processing units by the plurality of solution layers according to the progressive pruning fusion mechanism is performed, the method further comprises: setting the fusion rate of the elements in the second processing unit set to the elements in the first processing unit set for solution information fusion at 100% in the early solution layer; setting a linear decreasing mode that the fusion rate of elements in the second processing unit set to elements in the first processing unit set for solution information fusion is 100% to 0 in the intermediate solution layer; Setting the fusion rate of the elements in the second processing unit set to the elements in the first processing unit set for solution information fusion to be 0 in the later solution layer; The step of fusing the solution information of the elements in the latest second processing unit set and the elements in the first processing unit set in the multiple solution layers according to a progressive pruning fusion mechanism specifically includes: identifying the category of the answering layer to which the current answering layer belongs; If the answer layer category to which the current answer layer belongs is an early answer layer, fusing all answer information contained in all elements in the second processing unit set into elements contained in the first processing unit set; if the answer layer category to which the current answer layer belongs is an intermediate answer layer, acquiring a fusion rate corresponding to the current answer layer, and fusing answer information contained in all elements in the second processing unit set into elements contained in the first processing unit set according to the fusion rate; if the answer layer category to which the current answer layer belongs is a later-stage answer layer, generating an answer result by only using the elements contained in the first processing unit set.
7. The method of claim 6, wherein the step of obtaining a fusion rate corresponding to a current solution layer, and fusing solution information contained in all elements in the second processing unit set to elements contained in the first processing unit set according to the fusion rate specifically comprises: according to a preset attention weight calculation mode, attention weight values of elements in the second processing unit set in the current answering layer are calculated; sorting the elements in the second processing unit set based on the attention weight value to obtain an element sorting result with decreasing attention weight value; screening out target elements with attention weight values ranked at the front from element sorting results by taking the fusion rate as a screening rate; and fusing the solution information contained in all the target elements into the elements contained in the first processing unit set.
8. A problem solving apparatus, comprising: the data acquisition module is used for acquiring target problems and answering resources; the unitization processing module is used for unitizing the target problem and the answering resources respectively to obtain a first processing unit set corresponding to the target problem and a second processing unit set corresponding to the answering resources; The iterative merging module is used for merging the elements in the second processing unit set by adopting an iterative merging mechanism until the total number of the residual elements in the second processing unit set meets a preset target retention rate, so as to obtain a latest second processing unit set; the data transfer module is used for transferring the latest second processing unit set and the first processing unit set to a plurality of solution layers contained in the target solution data generation model; the pruning fusion module is used for fusing the elements in the latest second processing unit set and the answer information of the elements in the first processing unit set in the plurality of answer layers according to a progressive pruning fusion mechanism; and the solution result output module is used for carrying out final solution result fusion on the solution fusion information respectively obtained by the solution layers and outputting the final fused solution result at the target output layer.
9. A computer device comprising a memory and a processor, the memory having stored therein computer readable instructions which when executed by the processor implement the steps of the problem solving method of any of claims 1 to 7.
10. A computer readable storage medium having stored thereon computer readable instructions which when executed by a processor implement the steps of the problem solving method according to any of claims 1 to 7.

Description

Question answering method, device, equipment and storage medium Technical Field The application relates to the technical field of artificial intelligence, and relates to a problem solving method, device, equipment and storage medium, which are applied to solving a target problem by utilizing a large-scale multi-modal language understanding model. Background In recent years, large multimodal language understanding models have made great breakthroughs that not only understand text, but also "see" image and video content. The core architecture of these models is usually based on a neural network called a transducer, which is composed of a plurality of stacked solution layers, especially in the scenario of financial business questions, such as performing answer prediction in combination with user financial problems and stock fluctuation trend lines, and further, for example, performing disorder answers according to patient problems and graphic diagnosis results in medical disorder analysis, the corresponding reasoning process can be used. In the financial business question-and-answer scenario, whether the knowledge carrier in the answer knowledge base is text or an image, it is decomposed into a minimum knowledge processing unit called token. For knowledge content of a text carrier, a token may be a word or a word, for knowledge content of an image carrier, a token is a small area content of an image, for knowledge content of a video carrier, a token is a video frame content within a video. At present, all input token, namely response knowledge content, needs to be respectively subjected to response result calculation and information fusion processing at each processing layer, and particularly in the field of financial business data processing with huge and complicated data information, the processing mode can lead to response efficiency reduction, and is not beneficial to timely feeding response data back to a problem consultation terminal. Disclosure of Invention The embodiment of the application aims to provide a method, a device, equipment and a storage medium for solving a problem, so as to improve the response efficiency and timely feed the solving data back to a problem consultation terminal. In a first aspect, an embodiment of the present application provides a method for solving a problem, which adopts the following technical scheme: a method of solving a problem, comprising the steps of: Acquiring target questions and answering resources; Performing unitization processing on the target problem and the solution resource respectively to obtain a first processing unit set corresponding to the target problem and a second processing unit set corresponding to the solution resource; Adopting an iterative merging mechanism to merge elements in the second processing unit set until the total number of the remaining elements in the second processing unit set meets a preset target retention rate, and obtaining a latest second processing unit set; transmitting the latest second processing unit set and the first processing unit set to a plurality of solution layers contained in a target solution data generation model; According to a progressive pruning fusion mechanism, fusing the elements in the latest second processing unit set and the solution information of the elements in the first processing unit set in the solution layers; and finally fusing the solution results of the solution fusion information respectively obtained by the solution layers, and outputting the finally fused solution results at the target output layer. In a second aspect, an embodiment of the present application further provides a solution device, which adopts the following technical scheme: a problem solving apparatus comprising: the data acquisition module is used for acquiring target problems and answering resources; the unitization processing module is used for unitizing the target problem and the answering resources respectively to obtain a first processing unit set corresponding to the target problem and a second processing unit set corresponding to the answering resources; The iterative merging module is used for merging the elements in the second processing unit set by adopting an iterative merging mechanism until the total number of the residual elements in the second processing unit set meets a preset target retention rate, so as to obtain a latest second processing unit set; the data transfer module is used for transferring the latest second processing unit set and the first processing unit set to a plurality of solution layers contained in the target solution data generation model; the pruning fusion module is used for fusing the elements in the latest second processing unit set and the answer information of the elements in the first processing unit set in the plurality of answer layers according to a progressive pruning fusion mechanism; and the solution result output module is used for carrying out final solution result fusion o