Target model interaction method and device, electronic equipment and storage medium

Abstract

The invention discloses a target model interaction method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining a reasoning instruction; performing data reasoning based on the reasoning instruction through the target model to obtain reasoning process data and a reasoning result; wherein the reasoning process data comprises a plurality of branch nodes and a plurality of conclusion nodes which are cascaded in a data reasoning process; and displaying the reasoning process data and the reasoning result. Through the mode, the transparency and the interpretability of the reasoning path of the target model can be improved, and the efficiency of generating the model output result meeting the user expectation by the target model is improved.

Description

Technical Field

[0001] The present invention relates to the field of model interaction, and in particular to an interaction method, device, electronic device and storage medium of a target model. Background Art

[0002] Artificial Intelligence (AI) big models refer to machine learning models with super-large-scale parameters and complex computing structures. They are usually able to process massive amounts of data and complete various complex tasks. Currently, AI big models can be applied to business areas such as creative assistance and automated decision-making, such as writing, composing, painting, and video.

[0003] When the large artificial intelligence model performs reasoning based on user instructions, if there are problems such as inaccurate instructions described by the user or the training direction does not match the user's instructions, it is easy for the large artificial intelligence model to have misunderstanding deviations, resulting in the user not obtaining the expected model output results.

[0004] In this case, users usually adjust the description of the instructions and try to obtain the expected model output results by regeneration. However, users do not understand the decision path of the large artificial intelligence model, and the adjusted instructions may still be inaccurate, resulting in the failure to obtain the expected model output results. Summary of the invention

[0005] The present invention provides a method, device, electronic device and storage medium for interacting with a target model, so as to reduce the problem that the user's instructions are still inaccurate after adjustment, resulting in the model still being unable to output model content that meets expectations.

[0006] The target model interaction method provided by the present invention includes: obtaining reasoning instructions; performing data reasoning based on the reasoning instructions through the target model to obtain reasoning process data and reasoning results; wherein the reasoning process data includes multiple branch nodes and multiple conclusion nodes cascaded in the data reasoning process; and displaying the reasoning process data and reasoning results.

[0007] In some embodiments, data reasoning is performed based on reasoning instructions through a target model to obtain reasoning process data and reasoning results, including: performing data reasoning based on reasoning instructions through a target model to generate multiple branch nodes, and obtaining multiple conclusion nodes based on the multiple branch nodes; marking each branch node, and recording the association relationship between each branch node and the association relationship between each conclusion node and the corresponding branch node to obtain reasoning process data; determining a target conclusion node from multiple conclusion nodes, and generating reasoning results based on the target conclusion node and its corresponding multiple branch nodes.

[0008] In some embodiments, the reasoning process data and the reasoning results are displayed, including: displaying the reasoning process data and the reasoning results, and distinguishing and displaying a target conclusion node and multiple target branch nodes corresponding to the reasoning results.

[0009] In some embodiments, multiple conclusion nodes include a target conclusion node and multiple basic conclusion nodes; multiple branch nodes include multiple target branch nodes and multiple basic branch nodes; after the reasoning process data and the reasoning results are displayed, it includes: in response to receiving an adjustment instruction of a basic branch node, determining whether the basic branch node has a corresponding basic conclusion node; when the basic branch node has a corresponding basic conclusion node, the basic branch node and its corresponding branch node and conclusion node are displayed separately, and a new reasoning result is generated based on the basic branch node and its corresponding branch node and conclusion node; when the basic branch node does not have a corresponding basic conclusion node, the target model is used to continue reasoning based on the basic branch node until the conclusion node corresponding to the basic branch node is obtained, the basic branch node and its corresponding branch node and conclusion node are displayed separately, and a new reasoning result is generated based on the basic branch node and its corresponding branch node and conclusion node.

[0010] In some embodiments, after the reasoning process data and the reasoning results are displayed, it also includes: in response to receiving a fusion instruction between at least two branch nodes, fusing at least two branch nodes through a target model to generate a fusion node, and continuing reasoning based on the fusion node until a conclusion node corresponding to the fusion node is obtained, the branch nodes and conclusion nodes corresponding to the fusion node are distinguished and displayed, and new reasoning results are generated based on the branch nodes and conclusion nodes corresponding to the fusion node.

[0011] In some embodiments, displaying the reasoning process data and the reasoning results also includes: in response to receiving a focus instruction for at least part of the reasoning results, determining the branch nodes corresponding to at least part of the reasoning results based on the focus instruction; and simultaneously distinguishing and displaying at least part of the reasoning results and their corresponding branch nodes.

[0012] In some embodiments, data reasoning is performed based on reasoning instructions through a target model to generate multiple branch nodes, including: performing data reasoning based on reasoning instructions through a target model to generate multiple initial branch nodes; and obtaining multiple branch nodes from the multiple initial branch nodes based on preset requirements.

[0013] In some embodiments, the inference instruction includes a requirement for the number of branch nodes in the inference process data; a preset requirement is to meet the number requirement; based on the preset requirement, multiple branch nodes are screened out from multiple initial branch nodes, including: sorting the weights of each initial node, and selecting multiple branch nodes that meet the number requirement in descending order; or the preset requirement is to meet the weight threshold; based on the preset requirement, multiple branch nodes are screened out from multiple initial branch nodes, including: obtaining the weight threshold of the target model; determining the initial nodes whose weights are greater than or equal to the weight threshold among the multiple initial branch nodes as branch nodes, and hiding the initial nodes whose weights are less than the weight threshold to obtain multiple branch nodes.

[0014] The present invention also provides an interactive device for a target model, comprising an acquisition module, a reasoning module and a display module; the acquisition module is used to acquire reasoning instructions; the reasoning module is used to perform data reasoning based on the reasoning instructions through the target model to obtain reasoning process data and reasoning results; wherein the reasoning process data includes multiple branch nodes and multiple conclusion nodes cascaded in the data reasoning process; the display module is used to display the reasoning process data and reasoning results.

[0015] The present invention also provides an electronic device, which includes: a memory and a processor coupled to each other, and the processor is used to execute program instructions stored in the memory to implement the target model interaction method of any of the above embodiments.

[0016] The present invention also provides a computer-readable storage medium storing program data, which can be executed by a processor to implement the target model interaction method as described in any of the above embodiments.

[0017] Different from the prior art, the present invention obtains reasoning instructions; performs data reasoning based on the reasoning instructions through the target model to obtain reasoning process data and reasoning results; wherein the reasoning process data includes multiple branch nodes and multiple conclusion nodes cascaded in the data reasoning process; the reasoning process data and the reasoning results are displayed so that the user knows the reasoning logic and reasoning choices of the target model, which helps the user to analyze, understand and evaluate the decision path of the target model, improves transparency and interpretability, and thus helps the user to effectively adjust the content of the reasoning instructions when the expected model results are not obtained, improves the accuracy and reliability of the target model's re-reasoning, and speeds up the efficiency of the target model to obtain the expected model output results. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flowchart of an embodiment of an interactive method of a target model provided by the present invention;

[0019] Figure 2It is a flowchart of another embodiment of the target model interaction method provided by the present invention;

[0020] Figure 3 is a schematic diagram of an implementation method of reasoning process data;

[0021] Figure 4 yes Figure 3 A schematic diagram of an implementation method after the reasoning process data is adjusted;

[0022] Figure 5 yes Figure 3 A schematic diagram of an implementation method after data fusion of the reasoning process;

[0023] Figure 6 is a schematic diagram of an implementation method for focusing on and displaying reasoning process data and reasoning results;

[0024] Figure 7 It is a schematic diagram of a framework of an embodiment of an interactive device of a target model of the present application;

[0025] Figure 8 is a structural schematic diagram of an embodiment of an electronic device provided by the present invention;

[0026] Fig. 9 It is a structural schematic diagram of an embodiment of a computer-readable storage medium provided by the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The term "and / or" in this application is only a description of the association relationship of associated objects, indicating that there may be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the objects associated before and after are in an "or" relationship. In addition, "many" in this article means two or more than two. In addition, the term "at least one" in this article means any combination of at least two of any one or more of a plurality of, for example, including at least one of A, B, and C, and can mean including any one or more elements selected from the set consisting of A, B, and C. In addition, the terms "first", "second", and "third" in this application are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features.

[0029] In order to enable those skilled in the art to better understand the technical solution of the present application, the technical solution of the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0030] See also Figure 1 , Figure 1 It is a flow chart of an embodiment of the target model interaction method provided by the present invention.

[0031] Step S11: Obtain the inference instruction.

[0032] An inference instruction refers to an instruction issued by a user to a target model to instruct the target model to output the expected inference result based on the inference instruction.

[0033] The target model can be any artificial intelligence model, and its application directions include but are not limited to one or more of multiple fields such as writing, music production, painting generation, game development, education, editing, video generation, natural language processing, medical diagnosis, etc., and it is set based on actual conditions.

[0034] In a specific application scenario, when the application direction of the target model includes painting generation, the reasoning instruction may be to generate a painting image of a certain scene or a painting image of a certain object. In a specific application scenario, when the application direction of the target model includes natural language processing, the reasoning instruction may be to generate a summary of the content of a book based on the book. In this way, the specific reasoning instructions and the application direction of the target model are not limited here.

[0035] Step S12: Perform data reasoning based on the reasoning instructions through the target model to obtain reasoning process data and reasoning results; wherein the reasoning process data includes multiple branch nodes and multiple conclusion nodes cascaded in the data reasoning process.

[0036] In the process of data reasoning, the target model will generate reasoning divergence for multiple reasoning elements and generate multiple possible reasoning divergence paths. The multiple reasoning divergence paths generated at the reasoning divergence are marked as known branch nodes. The choice of different branch nodes by the target model in the reasoning process will directly affect the reasoning result finally generated by the target model. Among them, the generation of multiple reasoning divergence paths of the target model in the reasoning process usually depends on the structural design and training of the target model. Factors such as probabilistic selection, user guidance, domain expert knowledge, environmental context, and training data will affect the decision-making judgment of the target model on multiple reasoning divergence paths. After the decision judgment, the target model will continue to reason based on different branch nodes respectively. In the process of continuing reasoning, multiple new branch nodes may also be generated. Reasoning will continue based on each new branch node until the target model obtains a reasoning conclusion that meets the reasoning instruction. The reasoning conclusion will be marked as a conclusion node. Finally, the target model selects the target conclusion node from multiple conclusion nodes to generate the final reasoning result. That is, the conclusion node is the end node of different reasoning paths obtained by the target model under the influence of probabilistic selection, domain expert knowledge, environmental context, training data and other factors.

[0037] In a specific application scenario, when the reasoning instruction requires the target model to walk out of a maze in a maze deduction game, at a fork in the maze, the target model will generate four branch nodes: turn left, turn right, go straight, and go back. The target model will continue to reason based on these four branch nodes. If there are still forks in the reasoning process, it will continue to generate corresponding branch nodes, and then continue to reason based on the new branch nodes until it reasons to the exit of the maze and obtains the conclusion node. A maze may have multiple exits or multiple clearance paths, so the number of conclusion nodes may be multiple. The final target model selects the target conclusion node from multiple conclusion nodes to generate the final clearance path.

[0038] In a specific application scenario, when the reasoning instruction requires the target model to draw a tree, the target model may generate multiple branch nodes when reasoning about the type of tree, and then continue reasoning based on different branch nodes, and generate multiple branch nodes again when reasoning about the growth location of the tree, and so on, continue reasoning until all elements of the drawing are determined and the conclusion node is obtained. Due to different selections of branch nodes, the target model may finally obtain multiple conclusion nodes. The final target model selects the target conclusion node from multiple conclusion nodes to generate the final tree drawing.

[0039] This embodiment performs data reasoning based on reasoning instructions through the target model, determines multiple branch nodes and multiple conclusion nodes cascaded in the target model during the reasoning process, obtains reasoning process data, and selects the target conclusion node from the multiple conclusion nodes to generate the final reasoning result. The selection is determined based on the training and design structure of the target model.

[0040] Step S13: Display the inference process data and inference results.

[0041] In addition to displaying the reasoning results to the user, this embodiment also displays the reasoning process data of the target model to the user, that is, multiple branch nodes and multiple conclusion nodes in the reasoning process, so that the user knows the reasoning logic and reasoning choices of the target model, which helps the user to analyze, understand and evaluate the decision path of the target model, improves transparency and explainability, and thus helps the user to effectively adjust the content of the reasoning instructions when the expected model results are not obtained, improves the accuracy and reliability of the target model's re-reasoning, and speeds up the efficiency of the target model to obtain the expected model output results.

[0042] Through the above steps, the interactive method of the target model of this embodiment obtains the reasoning instruction; performs data reasoning based on the reasoning instruction through the target model to obtain the reasoning process data and the reasoning result; wherein the reasoning process data includes multiple branch nodes and multiple conclusion nodes cascaded in the data reasoning process; the reasoning process data and the reasoning result are displayed so that the user knows the reasoning logic and reasoning choices of the target model, which helps the user to analyze, understand and evaluate the decision path of the target model, improves transparency and interpretability, and thus helps the user to effectively adjust the content of the reasoning instruction when the expected model result is not obtained, improves the accuracy and reliability of the target model's re-reasoning, and speeds up the efficiency of the target model to obtain the expected model output result.

[0043] See also Figure 2 , Figure 2 It is a flowchart of another embodiment of the target model interaction method provided by the present invention.

[0044] Step S21: Obtain the inference instruction, perform data inference based on the inference instruction through the target model, generate multiple branch nodes, and obtain multiple conclusion nodes based on the multiple branch nodes.

[0045] After receiving the user's reasoning instruction, the target model performs data reasoning based on the reasoning instruction to generate multiple initial branch nodes. Based on the preset requirements, multiple branch nodes are screened from the multiple initial branch nodes, and multiple conclusion nodes are obtained based on the reasoning of the multiple branch nodes.

[0046] In a specific application scenario, when issuing an inference instruction, the user can add a requirement for the number of branch nodes in the inference process data in the inference instruction, and the preset requirement of this application scenario is the number requirement. The step of screening out multiple branch nodes from multiple initial branch nodes based on the preset requirements specifically includes: sorting the weights of each initial node, and selecting multiple branch nodes that meet the quantity requirement in descending order. For example: when the target model generates 100 initial branch nodes by inference, and the user's quantity requirement is only 10, the weights of these 100 initial branch nodes can be sorted, and the 10 branch nodes with the largest weights can be selected in descending order. Since these 10 branch nodes have the largest weights, these 10 branch nodes can, to a certain extent, represent the key divergence information in the reasoning process, thereby helping users quickly understand the reasoning process of the target model.

[0047] In a specific application scenario, the preset requirement may also be to meet a weight threshold. The step of filtering out multiple branch nodes from multiple initial branch nodes based on the preset requirements specifically includes: obtaining the weight threshold of the target model; determining the initial nodes whose weights are greater than or equal to the weight threshold as branch nodes, and hiding the initial nodes whose weights are less than the weight threshold to obtain multiple branch nodes. By setting the weight threshold to filter the branch nodes, the obtained branch nodes can, to a certain extent, represent the key divergence information in the reasoning process, thereby helping the user to quickly understand the reasoning process of the target model. The specific preset requirements for filtering are not limited here.

[0048] In the actual reasoning process of the target model, it may face a lot of reasoning divergences, resulting in a large number of initial branch nodes. If all the initial branch nodes are displayed to the user, it will have an adverse effect on the user's judgment and information extraction due to too many nodes. Therefore, after the target model generates multiple initial branch nodes, this embodiment screens the multiple initial branch nodes based on preset requirements to obtain multiple branch nodes, thereby streamlining the nodes, retaining key nodes, and hiding general nodes, so as to improve the efficiency, effectiveness and pertinence of users in obtaining key information from multiple branch nodes.

[0049] Step S22: Mark each branch node, and record the association relationship between each branch node and the association relationship between each conclusion node and the corresponding branch node to obtain reasoning process data.

[0050] Mark each branch node to determine the label of each branch node. The label is the specific content meaning of the branch node corresponding to the divergent reasoning path. For example, when the target model is reasoning about the clearance path of a maze, it will generate 4 branch nodes at a fork in the road. Mark these 4 branch nodes, and obtain 4 branch nodes with labels of turn left, turn right, go straight, and go back. Or when the target model is creating a certain plot, two branch nodes will be generated for the plot in which a student takes an exam. Mark these two branch nodes, and obtain 2 branch nodes with labels of the student passing the exam and the student failing the exam. The specific label of each branch node is determined based on the reasoning direction of each branch node itself.

[0051] The associations between the branch nodes and the associations between the conclusion nodes and the corresponding branch nodes are recorded to obtain the reasoning process data. The reasoning process data can be presented to the user in a network, tree, mind map or other types of visualization, which is selected based on actual needs.

[0052] See also Figure 3 , Figure 3 1 is a schematic diagram of an implementation method of reasoning process data. This schematic diagram is only for illustrative purposes and is not intended to be limiting. This embodiment is described using a tree-like visualization method as an example.

[0053] When the target model of this embodiment performs reasoning based on the reasoning instruction, four branch nodes are first generated: A, B, C, and D; the target model continues to reason along the four branch nodes respectively, and obtains the reasoning process data shown in the figure. Taking branch node A as an example, the reasoning process of the target model is: reasoning along branch node A, and then generating branch node A1 and branch node A2 at the next reasoning divergence, and the target model performs reasoning along branch node A1 and branch node A2 respectively, wherein after branch node A1, branch node A1.1 is generated, and after branch node A1.1, divergent branch nodes A1.1.1 and branch node A1.1.2 are generated, and then reasoning continues along branch node A1.1.2 to generate conclusion node 1. Among them, the target model may also generate a detour after reasoning to branch node A1.1 and return to branch node A. After generating branch nodes A2 and A1.1.1, the target model did not continue reasoning in this reasoning due to the limitations of the target model's computing power or the hardware on which the target model actually relies. However, this does not mean that the reasoning paths corresponding to branch nodes A2 and A1.1.1 cannot generate corresponding conclusion nodes.

[0054] For the specific reasoning process of branch nodes B, C, and D, please refer to Figure 3 The specific reasoning logic is the same as that of branch node A, and will not be repeated here.

[0055] The cascaded association relationship is expressed by connecting multiple branch nodes, and the association relationship is also expressed by connecting each conclusion node and the corresponding branch node, so as to obtain the reasoning process data. In order to facilitate the display of the reasoning process data, the label display of each node is omitted in this schematic diagram, but in the actual display, the label of each node in the reasoning process data will be displayed synchronously with the node.

[0056] The above-mentioned multi-node visualization display method converts the linear dialogue form of the target model into a mesh structure, so that users can understand the overall reasoning logic and path of the target model, which is convenient for users to analyze, understand and evaluate the decision path of the target model, improve the transparency and explainability of model reasoning, and thus help users adjust the content of reasoning instructions to issue subsequent instructions, so that subsequent instructions are more in line with user expectations.

[0057] Step S23: Determine a target conclusion node from the multiple conclusion nodes, and generate an inference result based on the target conclusion node and its corresponding multiple branch nodes.

[0058] The target model can select the conclusion node corresponding to the reasoning path with the largest comprehensive weight as the target conclusion node by comparing the comprehensive weights between the multiple branch nodes on the entire reasoning path corresponding to each conclusion node, and generate the reasoning result based on the target conclusion node and its corresponding multiple branch nodes. Among them, the target conclusion node is the last terminal node generated by the target model in a certain reasoning path, and the reasoning conclusion is the final model output content generated by the target model after the reasoning is completed, selecting a conclusion node and the specific direction of the multiple branch nodes on the entire reasoning path corresponding to it. For example: when the target model is reasoning the path through the maze, the conclusion node is the exit to the maze, and the reasoning result is the entire path through the maze finally selected by the target model in this reasoning.

[0059] In a specific application scenario, Figure 3 For example, Figure 3The implementation method generates a total of 3 conclusion nodes: 1, 2, and 3. Among them, the reasoning path composed of conclusion node 1 and the corresponding multiple branch nodes is: branch node A, branch node A1, branch node A1.1, branch node A1.1.2, and conclusion node 1. The reasoning path composed of conclusion node 2 and the corresponding multiple branch nodes is: branch node B, branch node B1, branch node B1.2, branch node B1.2.2, and conclusion node 2. The reasoning path composed of conclusion node 3 and the corresponding multiple branch nodes is: branch node C, branch node C1, branch node C1.2, branch node D1.1.1, and conclusion node 3; or branch node D, branch node D1, branch node D1.1, branch node D1.1.1, and conclusion node 3. Among them, although the above two reasoning paths both reach conclusion node 3, the reasoning results finally generated by these two reasoning paths are not exactly the same due to the differences in the selection of some branch nodes.

[0060] Assuming that the weight of the entire reasoning path corresponding to conclusion node 2 is the largest, the target model selects conclusion node 2 as the target conclusion node, and generates reasoning results based on the target conclusion node and its corresponding multiple branch nodes, that is, generates reasoning results based on branch node B, branch node B1, branch node B1.2, branch node B1.2.2 and conclusion node 2.

[0061] Step S24: Display the reasoning process data and the reasoning results, and distinguish and display the target conclusion node and multiple target branch nodes corresponding to the reasoning results.

[0062] After obtaining the reasoning process data and the reasoning results, the target model displays the reasoning process data and the reasoning results, wherein the target conclusion node and multiple target branch nodes corresponding to the reasoning results are displayed separately to emphasize to the user the reasoning path with the largest weight of this reasoning.

[0063] The display methods include but are not limited to bold, highlight, color, flash, focus and other visual forms. Figure 3 For example, the entire reasoning path corresponding to the target conclusion node 2 may be displayed in bold to distinguish it.

[0064] Through the display of the above reasoning process data, multiple reasoning paths can be presented to the user, comprehensively identifying each key node that affects the reasoning result. At the same time, the reasoning process data also presents each unselected branch node and the possible subsequent direction of these branch nodes. By showing the user the reasoning process data with a mesh structure, it helps the user to analyze, understand and evaluate the decision path of the target model, improve transparency and explainability, and thus help the user to effectively adjust the content of the reasoning instructions when the expected model results are not obtained, improve the accuracy and reliability of the target model's re-reasoning, and speed up the efficiency of the target model to obtain the expected model output results.

[0065] Step S25: In response to receiving the adjustment instruction of the basic branch node, determining whether the basic branch node has a corresponding basic conclusion node.

[0066] After displaying the reasoning process data and the reasoning results to the user, this embodiment can also receive the user's modification and adjustment of the reasoning process data, so that the user can obtain the expected model results by directly controlling the reasoning path, thereby speeding up the reasoning efficiency.

[0067] The multiple conclusion nodes in the reasoning process data include a target conclusion node and multiple basic conclusion nodes; the multiple branch nodes include multiple target branch nodes and multiple basic branch nodes. The target conclusion node and the target branch node refer to the nodes on the reasoning path corresponding to the reasoning result finally generated by the target model in this reasoning, and the basic conclusion node and the basic branch node refer to the nodes that do not correspond to the final reasoning result. Figure 3 For example, when conclusion node 2 is the target conclusion node, conclusion node 1 and conclusion node 3 are basic conclusion nodes. Branch node B, branch node B1, branch node B1.2, branch node B1.2.2 are target branch nodes, and the other branch nodes are all basic branch nodes.

[0068] When an adjustment instruction for a basic branch node is received, it means that the user believes that the current reasoning result does not meet his expectations, and believes that the basic branch node may produce a reasoning result that meets his expectations. Therefore, an adjustment instruction is issued for the basic branch node. Among them, the user can issue the adjustment instruction by clicking or long pressing the basic branch node on the displayed reasoning process data, or by entering the label of the basic branch node. The specific method is not limited here.

[0069] This step first determines whether the basic branch node has a corresponding basic conclusion node. Due to the limitations of the target model's computing power or the hardware that the target model actually relies on, there may be some basic branch nodes that have no subsequent reasoning. In this case, first determine whether the basic branch node has a corresponding final basic conclusion node.

[0070] Step S26: When a basic branch node has a corresponding basic conclusion node, the basic branch node and its corresponding branch nodes and conclusion nodes are displayed separately, and a new reasoning result is generated based on the basic branch node and its corresponding branch nodes and conclusion nodes.

[0071] In a specific application scenario, such as Figure 3 , if the user issues an adjustment instruction for branch node A1.1.2, and branch node A1.1.2 has a corresponding conclusion node 1, then branch node A1.1.2 and its corresponding branch node A, branch node A1, branch node A1.1 and conclusion node 1 are displayed separately. The nodes on the reasoning path corresponding to the original target conclusion node can continue to be displayed separately or no longer be displayed separately.

[0072] Among them, the display differentiation methods include but are not limited to bolding, highlighting, color differentiation, flashing, focusing and other visual forms.

[0073] Step S27: When there is no corresponding basic conclusion node for a basic branch node, continue reasoning based on the basic branch node through the target model until the conclusion node corresponding to the basic branch node is obtained, distinguish and display the basic branch node and its corresponding branch nodes and conclusion nodes, and generate a new reasoning result based on the basic branch node and its corresponding branch nodes and conclusion nodes.

[0074] In a specific application scenario, such as Figure 3 If the user issues an adjustment instruction for branch node A2, and there is no corresponding basic conclusion node for branch node A2, the target model continues to reason based on branch node A2 until the conclusion node corresponding to branch node A2 is obtained, and branch node A2 and its corresponding branch nodes and conclusion nodes are displayed separately.

[0075] Please read further Figure 4 , Figure 4 yes Figure 3 A schematic diagram of an implementation method after the reasoning process data is adjusted.

[0076] Assuming that branch node A2 subsequently generates branch node A2.1, branch node A2.1.1 and conclusion node 4, branch node A, branch node A2, branch node A2.1, branch node A2.1.1 and conclusion node 4 are displayed separately.

[0077] The nodes on the reasoning path corresponding to the original conclusion node 2 can be displayed without distinction.

[0078] Among them, the display differentiation methods include but are not limited to bolding, highlighting, color differentiation, flashing, focusing and other visual forms.

[0079] The above method, after receiving the adjustment instruction of the basic branch node, selects the entire reasoning path corresponding to the basic branch node for differentiated display, so that the user can fall back to select or reselect a different reasoning path to obtain a new reasoning result. This is beneficial for the user to compare and view the differences in the reasoning results caused by the selection of different branch nodes. The reasoning path of the new reasoning result is selected by the user himself to adjust the branch node, which is more in line with the user's own expectations, so that the target model can obtain the user's expected reasoning result more quickly, thereby improving the efficiency of the target model reasoning.

[0080] Step S28: In response to receiving a fusion instruction between at least two branch nodes, at least two branch nodes are fused through the target model to generate a fusion node, and reasoning is continued based on the fusion node until a conclusion node corresponding to the fusion node is obtained, the branch nodes and conclusion nodes corresponding to the fusion node are distinguished and displayed, and new reasoning results are generated based on the branch nodes and conclusion nodes corresponding to the fusion node.

[0081] Users can also merge at least two branch nodes to generate new reasoning paths and new reasoning results. This fusion can be applied to artistic directions such as writing, painting, and composing for creative design. In terms of operation, users can use a branch node as a base point and drag other branch nodes that need to be merged to the branch node, thereby issuing fusion instructions between these multiple branch nodes.

[0082] Among them, after at least two branch nodes are fused, a new fusion node is added, and reasoning is continued based on the fusion node through the target model until the conclusion node corresponding to the fusion node is obtained, the branch nodes and conclusion nodes corresponding to the fusion node are distinguished and displayed, and new reasoning results are generated based on the fusion node and its corresponding branch nodes and conclusion nodes.

[0083] See also Figure 5 , Figure 5 yes Figure 3 Schematic diagram of an implementation method after data fusion of the reasoning process.

[0084] Assuming that branch node B1.1 is fused with branch node A1.1.2 to generate a new fused node E, the target model continues to reason based on fused node E to generate branch node E1 and conclusion node 5. The branch node A, branch node A1, branch node A1.1, branch node A1.1.2, branch node B, branch node B1, branch node B1.1, fused node E, branch node E1 and conclusion node 5 corresponding to fused node E are distinguished and displayed, and new reasoning results are generated based on the above nodes.

[0085] The nodes on the reasoning path corresponding to the original conclusion node 2 can be displayed without distinction.

[0086] In the above method, the user's interactive behavior on the node not only includes the node's fallback selection and reselection, but also includes the interactive behavior of reintegrating the linear nodes of different branches, so that the user can select at least two branch nodes, and generate a new reasoning path and corresponding fusion nodes when the algorithm logic allows. According to the above path, repeated or cyclic operations can be performed to explore a wider range of possible reasons and deduce results to meet the user's various requirements for reasoning results.

[0087] The display of reasoning process data can not only present the branch nodes that are selected or not selected in this reasoning process, but also present the subsequent direction of these branch nodes, including the direction of connecting to another unselected branch node after the path is detoured, the direction of returning to the selected node after the path is detoured, and the direction of the corresponding conclusion node generated subsequently. In addition, through node fusion, the unknown direction of the current node can also be generated for creative reasoning.

[0088] Step S29: in response to receiving a focus instruction of at least part of the reasoning result, determining a branch node corresponding to at least part of the reasoning result based on the focus instruction; and simultaneously distinguishing and displaying at least part of the reasoning result and its corresponding branch node.

[0089] This embodiment will display the reasoning process data and the reasoning results simultaneously. When the user issues a focus instruction for at least part of the reasoning results, this embodiment determines the branch nodes corresponding to at least part of the reasoning results based on the focus instruction; and simultaneously distinguishes and displays at least part of the reasoning results and their corresponding branch nodes. The user can issue a focus instruction by clicking at the end of at least part of the reasoning results.

[0090] See also Figure 6 , Figure 6 It is a schematic diagram of an implementation method for focusing on and displaying reasoning process data and reasoning results.

[0091] When the cursor 61 is focused on the inference result 60, the partial inference result focused based on the position of the cursor 61, that is, the inference result before the position of the cursor 61 is taken as the partial inference result that the user thinks needs to be focused on, and this partial inference result is distinguished and displayed, and the branch node B, branch node B1 and branch node B1.2 corresponding to the partial inference result in the inference process data are synchronously distinguished and displayed, so that the user knows that the inference logic of this partial inference result is determined by the above-mentioned corresponding branch nodes.

[0092] In a specific application scenario, when the reasoning instruction requires the target model to summarize the content of a book, the reasoning result 60 is the summarized text content. The user issues a focus instruction for part of the summarized text content. At this time, the part of the summarized text content is distinguished and displayed, and the branch nodes inferred from the part of the summarized text content in the reasoning process data are also distinguished and displayed, so that the user knows the reasoning logic of the part of the summarized text content.

[0093] In other application scenarios, when the reasoning instruction requires the target model to draw, the reasoning result 60 is a specific drawing image. When a focus instruction is issued for a drawing area such as an object or background of the specific drawing image, the part of the area is displayed separately, such as highlighted or flashed, and the branch node inferred from the part of the drawing area in the reasoning process data is also displayed separately, so that the user knows the reasoning logic of the part of the drawing area. The specific reasoning scenario is not limited here.

[0094] The above method further improves the transparency and comprehensibility of the target model's decisions, improves the user's understanding of the target model's current reasoning logic, helps the user analyze, understand and evaluate the decision path of the target model, and thus helps the user to effectively adjust the content of the reasoning instructions when the expected model results are not obtained, improves the accuracy and reliability of the target model's re-reasoning, and speeds up the efficiency of the target model in obtaining the expected model output results.

[0095] Through the above method, when receiving the reasoning instruction to perform data reasoning, this embodiment records the decision branch node every time a decision scenario that may generate a reasoning branch is encountered during the reasoning process, obtains multiple cascaded branch nodes and subsequent conclusion nodes, and generates reasoning process data and reasoning results. When the target model outputs the reasoning result to the user, the reasoning process data is also output synchronously, and the reasoning process data can be displayed to the user in a visual manner such as a mesh, tree, mind map or other types, so that the linear dialogue form created by the target model is converted into an interactive mesh structure that can be widely explored, so that the user can know the reasoning logic and reasoning selection of the target model through the reasoning process data, which helps the user to analyze, understand and evaluate the decision path of the target model, improve transparency and explainability, and thus help the user to effectively adjust the content of the reasoning instruction when the expected model result is not obtained, improve the accuracy and reliability of the target model's re-reasoning, and speed up the efficiency of the target model to obtain the expected model output result. After the display, the present embodiment also receives the user's adjustment instructions, fusion instructions, and focus instructions, so that the user can roll back, reselect, and reintegrate key nodes, thereby providing the user with the function of controlling the target model's reasoning path, improving the directionality and efficiency of interacting with the target model, and making the reasoning results of the target model produce a wider range of possible explorations, thereby generating new "creative possibilities" to meet the user's more reasoning needs. In the above adjustment process, the user can also clearly see the adjusted reasoning path based on the reasoning process data, and understand how the target model makes decisions. Setting marks and node displays in this way helps users analyze, understand, and evaluate the decision-making path of the target model, and improves transparency and explainability. This also helps to track and monitor the decision-making process of the target model.

[0096] See also Figure 7 , Figure 7 The target model interaction device of this embodiment is applied to the target model interaction method of any of the above embodiments.

[0097] The target model interaction device 70 includes an acquisition module 71, a reasoning module 72 and a display module 73. The acquisition module 71 is used to acquire the reasoning instruction; the reasoning module 72 is used to perform data reasoning based on the reasoning instruction through the target model to obtain the reasoning process data and the reasoning result; wherein the reasoning process data includes multiple branch nodes and multiple conclusion nodes cascaded in the data reasoning process; the display module 73 is used to display the reasoning process data and the reasoning result.

[0098] The above solution can display the reasoning process data and reasoning results, so that users can know the reasoning logic and reasoning choices of the target model, help users analyze, understand and evaluate the decision path of the target model, improve transparency and explainability, and speed up the efficiency of the target model to obtain the expected model output results.

[0099] The display module 73 is also used to display the reasoning process data and the reasoning results, and to distinguish and display the target conclusion node and multiple target branch nodes corresponding to the reasoning results.

[0100] The acquisition module 71 is also used to receive adjustment instructions for the basic branch node, and the reasoning module 72 is also used to determine whether the basic branch node has a corresponding basic conclusion node; when the basic branch node has a corresponding basic conclusion node, the basic branch node and its corresponding branch nodes and conclusion nodes are distinguished and displayed through the display module 73, and a new reasoning result is generated based on the basic branch node and its corresponding branch nodes and conclusion nodes through the reasoning module 72; when the basic branch node does not have a corresponding basic conclusion node, the reasoning module 72 is also used to continue reasoning based on the basic branch node through the target model until the conclusion node corresponding to the basic branch node is obtained, and the basic branch node and its corresponding branch nodes and conclusion nodes are distinguished and displayed through the display module 73, and a new reasoning result is generated based on the basic branch node and its corresponding branch nodes and conclusion nodes through the reasoning module 72.

[0101] The acquisition module 71 is also used to receive a fusion instruction between at least two branch nodes, the reasoning module 72 is also used to fuse at least two branch nodes through the target model to generate a fusion node, and continue reasoning based on the fusion node until a conclusion node corresponding to the fusion node is obtained, the display module 73 is also used to distinguish and display the branch nodes and conclusion nodes corresponding to the fusion node, and the reasoning module 72 is also used to generate new reasoning results based on the branch nodes and conclusion nodes corresponding to the fusion node.

[0102] The acquisition module 71 is also used to receive a focus instruction for at least part of the reasoning results, and the reasoning module 72 is also used to determine the branch nodes corresponding to at least part of the reasoning results based on the focus instruction; the display module 73 is also used to synchronously distinguish and display at least part of the reasoning results and their corresponding branch nodes.

[0103] The reasoning module 72 is also used to perform data reasoning based on reasoning instructions through the target model, generate multiple branch nodes, and obtain multiple conclusion nodes based on the multiple branch nodes; mark each branch node, and record the association relationship between each branch node and the association relationship between each conclusion node and the corresponding branch node to obtain reasoning process data; determine the target conclusion node from multiple conclusion nodes, and generate reasoning results based on the target conclusion node and its corresponding multiple branch nodes.

[0104] The reasoning module 72 is also used to perform data reasoning based on the reasoning instructions through the target model to generate multiple initial branch nodes; and to obtain multiple branch nodes from the multiple initial branch nodes based on preset requirements.

[0105] The reasoning module 72 is also used to sort the weights of each initial node, and select multiple branch nodes that meet the quantity requirements in descending order; or obtain the weight threshold of the target model; determine the initial nodes whose weights are greater than or equal to the weight threshold among the multiple initial branch nodes as branch nodes, and hide the initial nodes whose weights are less than the weight threshold, so as to obtain multiple branch nodes.

[0106] In some embodiments, the reasoning module 72 includes a processing submodule 721, a storage submodule 723 and a control submodule 722. The processing submodule 721 is connected to the acquisition module 71, the storage submodule 723 and the control submodule 722 respectively, and the control submodule 722 is also connected to the display module 73. The processing submodule 721 is used to perform data reasoning based on the reasoning instruction, generate multiple initial branch nodes and mark each node until the corresponding conclusion node is obtained, and generate the reasoning result. The storage submodule 723 is used to store each node and its marking information. The control submodule 722 is used to screen multiple initial branch nodes, classify and group the marking information, obtain multiple branch nodes, integrate multiple branch nodes and multiple conclusion nodes to generate reasoning process data, and send the reasoning process data and reasoning results to the display module 73.

[0107] In a specific application scenario, the acquisition module 71 obtains the user's reasoning instructions and reports the reasoning instructions to the processing submodule 721; the processing submodule 721 performs data reasoning based on the reasoning instructions, generates multiple initial branch nodes and marks each node until the corresponding conclusion node is obtained, and generates a reasoning result. In this process, the storage submodule 723 stores each node and its marking information. The control submodule 722 screens multiple initial branch nodes, classifies and groups the marking information, obtains multiple branch nodes, integrates multiple branch nodes and multiple conclusion nodes to generate reasoning process data, and sends the reasoning process data and reasoning results to the display module 73. Finally, the display module 73 displays the reasoning process data and reasoning results.

[0108] The acquisition module 71 is also used to obtain the user's adjustment instructions, fusion instructions and focus instructions, and repeat the above execution process to adjust the reasoning process data and reasoning results according to the user's specific instruction content.

[0109] The above method provides users with the function of controlling the reasoning path of the target model, improves the directionality and efficiency of interaction with the target model, and enables the reasoning results of the target model to produce a wider range of possible explorations to meet more reasoning needs of users.

[0110] Based on the same inventive concept, the present invention also proposes an electronic device, which can be executed to implement the target model interaction method of any of the above embodiments, see Figure 8 , Figure 8 It is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. The electronic device includes a processor 81 and a memory 82.

[0111] The processor 81 is used to execute the program instructions stored in the memory 82 to implement the steps of the interaction method of any target model described above. In a specific implementation scenario, the electronic device may include but is not limited to: a microcomputer, a server, and in addition, the electronic device may also include a mobile device such as a laptop computer and a tablet computer, which is not limited here.

[0112] Specifically, the processor 81 is used to control itself and the memory 82 to implement the steps of any of the above embodiments. The processor 81 can also be called a processor (Central Processing Unit). The processor 81 may be an integrated circuit chip with signal processing capabilities. The processor 81 can also be a general-purpose processor, a digital signal processor (Digital Signal Processor, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field-programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. In addition, the processor 81 can be implemented by an integrated circuit chip.

[0113] The above scheme improves the transparency and interpretability of the reasoning path of the target model, which is conducive to improving the efficiency of the target model in generating model output results that meet user expectations.

[0114] Based on the same inventive concept, the present invention also proposes a computer-readable storage medium, see Fig. 9 , Fig. 91 is a schematic diagram of the structure of an embodiment of a computer-readable storage medium provided by the present invention. The computer-readable storage medium 90 stores at least one program data 91, and the program data 91 is used to implement any of the above methods. In one embodiment, the computer-readable storage medium 90 includes: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0115] In the several embodiments provided by the present invention, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device implementation described above is only schematic, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0116] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0117] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0118] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium.

[0119] The above description is only an implementation mode of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A target model interaction method, characterized in that: include: Get the inference instruction; Performing data reasoning based on the reasoning instruction through the target model to obtain reasoning process data and reasoning results; wherein the reasoning process data includes a plurality of branch nodes and a plurality of conclusion nodes cascaded in the data reasoning process; The inference process data and the inference result are displayed.

2. The target model interaction method according to claim 1, characterized in that: The data reasoning is performed based on the reasoning instruction by the target model to obtain the reasoning process data and the reasoning result, including: Performing data reasoning based on the reasoning instruction through the target model to generate a plurality of branch nodes, and obtaining a plurality of conclusion nodes based on the reasoning of the plurality of branch nodes; Marking each of the branch nodes, and recording the association relationship between each of the branch nodes and the association relationship between each conclusion node and the corresponding branch node, to obtain the reasoning process data; A target conclusion node is determined from the multiple conclusion nodes, and the reasoning result is generated based on the target conclusion node and the multiple branch nodes corresponding to it.

3. The target model interaction method according to claim 1 or 2, characterized in that: The displaying of the reasoning process data and the reasoning result includes: The reasoning process data and the reasoning result are displayed, and a target conclusion node and a plurality of target branch nodes corresponding to the reasoning result are distinguished and displayed.

4. The target model interaction method according to claim 3, characterized in that: The multiple conclusion nodes include a target conclusion node and multiple basic conclusion nodes; The plurality of branch nodes include a plurality of target branch nodes and a plurality of basic branch nodes; After displaying the reasoning process data and the reasoning result, the method further comprises: In response to receiving an adjustment instruction of a basic branch node, determining whether the basic branch node has a corresponding basic conclusion node; When the basic branch node has a corresponding basic conclusion node, the basic branch node and its corresponding branch node and conclusion node are displayed separately, and a new reasoning result is generated based on the basic branch node and its corresponding branch node and conclusion node; When there is no corresponding basic conclusion node for the basic branch node, the target model continues to reason based on the basic branch node until the conclusion node corresponding to the basic branch node is obtained, the basic branch node and its corresponding branch nodes and conclusion nodes are distinguished and displayed, and a new reasoning result is generated based on the basic branch node and its corresponding branch nodes and conclusion nodes.

5. The target model interaction method according to claim 1 or 2, characterized in that: After displaying the reasoning process data and the reasoning result, the method further includes: In response to receiving a fusion instruction between at least two branch nodes, the at least two branch nodes are fused through the target model to generate a fusion node, and reasoning is continued based on the fusion node until a conclusion node corresponding to the fusion node is obtained, the branch nodes and conclusion nodes corresponding to the fusion node are distinguished and displayed, and new reasoning results are generated based on the branch nodes and conclusion nodes corresponding to the fusion node.

6. The target model interaction method according to claim 1 or 2, characterized in that: The displaying of the reasoning process data and the reasoning result also includes: In response to receiving a focus instruction for at least part of the reasoning result, determining a branch node corresponding to at least part of the reasoning result based on the focus instruction; At least part of the reasoning results and their corresponding branch nodes are displayed synchronously and distinguished.

7. The target model interaction method according to claim 2, characterized in that: The method of performing data reasoning based on the reasoning instruction through the target model to generate multiple branch nodes includes: Performing data reasoning based on the reasoning instruction through the target model to generate multiple initial branch nodes; Based on preset requirements, a plurality of the branch nodes are obtained by screening from the plurality of the initial branch nodes.

8. The target model interaction method according to claim 7, characterized in that: The reasoning instruction includes the quantity requirement of the branch nodes in the reasoning process data; The preset requirement is to meet the quantity requirement; the selecting multiple branch nodes from the multiple initial branch nodes based on the preset requirement includes: sorting the weights of the initial nodes, and selecting multiple branch nodes that meet the quantity requirement in descending order; or The preset requirement is to meet the weight threshold; the method of selecting multiple branch nodes from the multiple initial branch nodes based on the preset requirement includes: obtaining the weight threshold of the target model; determining the initial nodes whose weights are greater than or equal to the weight threshold among the multiple initial branch nodes as the branch nodes, and hiding the initial nodes whose weights are less than the weight threshold, so as to obtain the multiple branch nodes.

9. An interactive device for a target model, characterized in that: include: An acquisition module is used to obtain inference instructions; An inference module, used to perform data inference based on the inference instruction through a target model to obtain inference process data and inference results; wherein the inference process data includes a plurality of branch nodes and a plurality of conclusion nodes cascaded in the data inference process; The display module is used to display the reasoning process data and the reasoning result.

10. An electronic device, characterized in that: The electronic device comprises: a memory and a processor coupled to each other, wherein the processor is used to execute program instructions stored in the memory to implement the target model interaction method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program data, and the program data can be executed by a processor to implement the target model interaction method according to any one of claims 1 to 8.

Images