Intelligent Question Answering System and Method Based on Automated Iterative Reasoning
By constructing an intelligent question-answering system based on automatic iterative reasoning, the problems of omission, duplication, and premature termination in intelligent question-answering systems in social governance are solved, the recoverability of tasks and the interpretability of results are achieved, and the efficiency and accuracy of the question-answering process are improved.
Patent Information
- Application Number
- CN202611395436.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-09
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies in intelligent question-answering systems for social governance are prone to problems such as premature termination before the evidence meets the constraints, repeated execution, difficulty in identifying conflicts caused by inconsistent execution results, inability to replan when execution fails, and inability to recover tasks due to historical interruptions.
An intelligent question-answering system based on automatic iterative reasoning is adopted. It forms a closed-loop intelligent question-answering process through a question state construction module, a subtask selection module, an execution scheduling module, an observation result standardization module, a state feedback generation module, an iterative state update module, and a dynamic termination judgment module. This ensures that after each round of execution, completion changes, evidence increments, conflicts, and anomaly feedback are explicitly generated, and the state is updated based on the feedback to dynamically terminate the control reasoning loop.
It reduces omissions, invalid repetitions, and premature termination in complex question-and-answer processes, achieves task recoverability and result interpretability, and improves the efficiency and accuracy of the question-and-answer process.
Smart Images

Figure CN123115342A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of social governance information processing technology, and more specifically, to an intelligent question-answering system and method based on automatic iterative reasoning. Background Technology
[0002] With the development of large language models and intelligent agent technology, computer systems can transform natural language tasks into reasoning actions and invoke external execution capabilities such as knowledge retrieval, rule validation, computational programs, or business services. For complex tasks that cannot be completed in one go, the system typically executes the "reasoning-action-observation" process alternately, using the result of the previous action as the context for the next round of reasoning.
[0003] In the field of social governance, public reporting, grid patrols, hotline handling, and inter-departmental collaboration generate a large number of natural language processing (NLP) questions. These questions, while superficially in a question-and-answer format, often require the system to continuously complete multiple interdependent sub-tasks, such as event element completion, event classification, risk identification, similar event retrieval, responsible department recommendation, and integration of handling criteria. For example, when a user inputs "A shop in a residential area has been occupying the road for a long time and has had a dispute with residents; how should this be handled?", the system cannot simply answer based on a single text generation. It must also identify the location and perpetrators, determine the event category, verify risk factors, query relevant handling information, and determine the responsible department. The facts or anomalies returned by the previous sub-task will influence the selection of subsequent sub-tasks.
[0004] Social governance issues also have characteristics such as incomplete input information, non-standard event descriptions, fragmented departmental knowledge, risk levels changing with new evidence, and the need to keep records of handling results. If the system simply appends each execution result as ordinary dialogue text, it cannot clearly identify which subtasks have been completed, which pieces of evidence conflict, or which execution calls failed, and it is also difficult to determine whether the current iteration has truly promoted problem resolution.
[0005] While existing technologies can generate multiple rounds of inference or iterative calls, they still have at least the following problems.
[0006] First, it is easy to end prematurely before the evidence meets the constraints, or to repeat execution under low gain conditions.
[0007] Second, when the return structures of different execution capabilities are inconsistent, the reasoning module has difficulty distinguishing between new facts, duplicate facts, conflicting facts and failure feedback, and cannot stably calculate the effective changes produced by the current execution relative to the previous state.
[0008] Third, if execution fails, evidence conflicts, or actions are invalid, the original plan will continue to be used, but when the maximum number of rounds is reached, it can only be forcibly truncated.
[0009] Fourth, it may cause the execution request to be retained in the history without corresponding observations, or the occurrence of isolated observations, affecting the state judgment. This may cause the model interface to reject the context, making the task unrecoverable.
[0010] Therefore, how to adapt to scenarios such as social governance incident handling, risk assessment, and inter-departmental collaboration, while maintaining the scalability of execution units and the recoverability of reasoning processes, is an urgent problem to be solved. Summary of the Invention
[0011] In view of the shortcomings of the prior art, the purpose of this invention is to provide an intelligent question-answering system and method based on automatic iterative reasoning, which can reduce omissions, invalid repetitions and premature termination in complex question-answering.
[0012] To achieve the above objectives, in a first aspect, the present invention provides an intelligent question-answering system based on automatic iterative reasoning, including a question state construction module, a subtask selection module, an execution scheduling module, an observation result standardization module, a state feedback generation module, an iterative state update module, a dynamic termination determination module, and an answer generation module; The question state construction module is used to construct the initial inference state based on the user's question. The inference state includes at least the question-answering goal, constraint set, candidate subtask set, uncompleted subtask set, evidence set, execution history, and resource budget. The subtask selection module is used to determine the target subtask for the next round from the candidate subtask set or the set of subtasks to be completed based on the current round inference state. The target subtask includes at least the information requirement to be met, the subtask type, the execution parameters, and the expected feedback type. The execution scheduling module is used to select and trigger the execution unit according to the target sub-task. The execution unit is at least one of the following: information acquisition unit, knowledge retrieval unit, entity extraction unit, event classification unit, risk assessment unit, department recommendation unit, calculation unit, rule verification unit, or external service unit. The observation results standardization module is used to convert the heterogeneous results returned by the execution unit into standardized observation results, i.e., a unified observation structure, which includes at least the execution request identifier, execution status, fact payload, source identifier, generation time, confidence information, exception information, and feedback type. The status feedback generation module is used to associate standardized observation results with the current round of reasoning status and target sub-tasks, and extract information on sub-task completion changes, evidence increments, evidence conflicts, execution anomalies and repeated executions to form the status feedback quantity for the next round. The iterative state update module is used to update the set of completed subtasks, the set of subtasks to be completed, the set of evidence, the set of conflicts, the execution history, and the resource budget based on the state feedback, so as to form the inference state for the next round. The dynamic termination judgment module is used to combine and judge the status indicators and output a successful termination instruction, a continue iteration instruction, a reselection of subtasks instruction, or a downgraded termination instruction. Specifically, a successful termination instruction is output when the subtask completion reaches the completion threshold, the evidence sufficiency reaches the sufficiency threshold, and there are no unresolved items or high-level conflicts. A continue iteration instruction is output when the task has not yet met the successful termination conditions, but the current round generates positive iteration gain, the subtask redundancy is low, and the resource budget is sufficient to support the next round of execution. A reselection of subtasks instruction is output when the current round fails, a high-level conflict occurs, the current subtask does not reduce the number of subtasks to be completed, or the subtask redundancy increases but has not yet met the downgraded termination conditions. A downgraded termination instruction is output when the information gain for a consecutive preset number of rounds is lower than the gain threshold, the redundancy reaches the redundancy threshold, the resource budget is exhausted, or an unrecoverable anomaly occurs. The subtask selection module responds to the continue iteration instruction by selecting the next target subtask from the updated set of subtasks to be completed; responds to the reselect subtask instruction by updating the candidate subtask set based on execution exceptions, evidence conflicts, or low gain reasons and then selecting the next target subtask, until the dynamic termination judgment module outputs a successful termination instruction or a downgraded termination instruction; the answer generation module generates question-and-answer results associated with the evidence set based on the reasoning state at the time of termination.
[0013] As a further optimization, the problem status construction module parses the user's problem into multiple information requirements and generates candidate subtasks based on the information requirements. It records the priority, completion mark, dependency relationship, supporting evidence and conflict mark for each candidate subtask. The subtask selection module prioritizes candidate subtasks whose dependency conditions have been met and which are expected to reduce the number of subtasks to be completed.
[0014] As a further preferred option, the status feedback generation module compares the current standardized observation results with the historical evidence set, divides the fact load into new facts, corrective facts, conflicting facts and duplicate facts, and generates completion status feedback indicating whether the target subtask has changed from incomplete to completed, remains incomplete or changed to abnormal.
[0015] As a further preferred option, the iterative state update module uses a state difference function to generate a state increment relative to the current round of reasoning state, and uses a state merging function to write the state increment into the next round of reasoning state, so that the selection of the target sub-task in the next round takes the updated set of sub-tasks to be completed, the set of evidence, and the anomaly information as input.
[0016] As a further preferred option, the dynamic termination judgment module outputs a successful termination command when the subtask completion is not lower than the completion threshold, the evidence sufficiency is not lower than the sufficiency threshold, the number of subtasks to be completed is zero, and there are no unresolved high-level conflicts; outputs a downgrade termination command when the iteration gain of a consecutive preset number of rounds is not higher than the gain threshold, the subtask repetition is not lower than the repetition threshold, or the remaining budget is insufficient to support the expected cost of the next round; and outputs a reselection subtask command when there is an execution anomaly, evidence conflict, or the current target subtask cannot reduce the number of subtasks to be completed.
[0017] As a further preferred embodiment, the system also includes a history consistency repair module. Before providing the execution history to the subtask selection module, the history consistency repair module verifies whether each execution request identifier has a corresponding observation result; excludes incomplete execution rounds that lack corresponding observation results and isolated observation results that lack corresponding execution requests from the current inference context copy; and forms a paired complete inference context without changing the original persistent state.
[0018] As a further preferred embodiment, the system also includes a state persistence module, which saves the inference state, execution request and standardized observation results of each round according to the task session identifier, and continues to iterate from the most recent complete pair of inference states when the task resumes.
[0019] As a further optimization, the execution scheduling module selects execution units based on the expected feedback type of the target subtask, the description of the execution unit's capabilities, historical success rate, expected latency, and invocation cost; when the execution unit invocation fails, it still generates standardized observation results with execution request identifiers and exception information, and provides the exception information as status feedback to the subtask selection module.
[0020] Secondly, the present invention provides an intelligent question-answering method based on automatic iterative reasoning applied to the above-mentioned system, comprising: S1. Construct an initial reasoning state based on the user's question, including the question-answering goal, constraint set, candidate subtask set, uncompleted subtask set, evidence set, execution history, and resource budget; S1. Construct an initial reasoning state based on the user's question, including the question-answering goal, constraint set, candidate subtask set, uncompleted subtask set, evidence set, execution history, and resource budget; S2. Select the target subtask based on the current inference state, and select an execution unit to execute the target subtask; S3. Convert the execution results into standardized observations with execution request identifiers; S4. Based on standardized observation results, target sub-tasks and the previous round of inference state, extract information on sub-task completion changes, evidence increments, evidence conflicts, execution anomalies and repeated executions to form state feedback quantities, and update the current inference state according to the state feedback quantities. S5. Make a combined judgment based on the updated subtask completion rate, evidence sufficiency, iteration gain, subtask redundancy, number of subtasks to be completed, execution anomaly information, and resource budget; S6. When the combined judgment result is to continue iteration, select the next target subtask from the updated set of subtasks to be completed and return to step S2; when the combined judgment result is to reselect a subtask, update the candidate subtask set according to the state feedback quantity and return to step S2; when the combined judgment result is successful termination or downgraded termination, generate the question and answer result according to the current reasoning state.
[0021] Thirdly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described intelligent question-answering method based on automatic iterative reasoning.
[0022] Implementing the intelligent question-answering system and method based on automatic iterative reasoning provided by this invention has the following beneficial effects: This invention integrates artificial intelligence reasoning, task planning, knowledge-enhanced question answering, and social governance information processing technologies to form a closed-loop intelligent question answering system centered around subtask selection, state feedback, iterative updates, and dynamic termination. Specifically, it constructs a reasoning state based on the user's question, including the objective, constraints, candidate subtasks, uncompleted subtasks, evidence set, execution history, and resource budget. In each iteration, a target subtask is selected based on the current reasoning state, and the execution unit matching the target subtask is invoked to obtain standardized observation results. From the standardized observation results, information on subtask completion changes, evidence increments, evidence conflicts, execution anomalies, and duplicate executions is extracted to form state feedback quantities and update the reasoning state. Then, based on the updated subtask completion degree, evidence sufficiency, iteration gain, duplication rate, anomaly information, and resource budget, the system dynamically outputs instructions for successful termination, continued iteration, reselection of subtasks, or downgraded termination. The system also verifies the completeness of the pairing between execution requests and observation results and saves the reasoning state according to the session identifier. This forms an automatic closed loop of "subtask selection—execution—state feedback—iterative update—dynamic termination," which can reduce omissions, invalid repetitions, and premature termination in complex question answering.
[0023] Compared with the prior art, the present invention has the following advantages: (1) Transform complex problems into a set of subtasks that are selectable, completeable and reselectable, so that the question-and-answer process is transformed from a one-time generation to an automatic iterative process constrained by the current state.
[0024] (2) After each round of execution, five types of status feedback are explicitly generated: completion change, evidence increment, conflict, abnormality and repetition. The feedback is used to update the selection criteria for the next round of sub-tasks, solving the problem of "only termination judgment and lack of status feedback loop" in the existing process.
[0025] (3) By distinguishing between new, error correction, conflict and repeated feedback through state increment, it is possible to quantify whether each round of execution truly promotes the completion of the task; continuous low gain and high repetition state can be identified in a timely manner, reducing invalid calls and resource waste.
[0026] (4) By standardizing the observation structure, the return differences of different execution units are shielded, so that the same state feedback and termination determination mechanism can be reused for tasks such as information acquisition, calculation, classification, diagnosis and rule verification.
[0027] (5) By performing a pairing check between the execution request and the observation results, incomplete rounds are eliminated without destroying the original audit records, so that the task after abnormal interruption can continue to run from the paired complete state.
[0028] (6) Degradation termination does not output pseudo-complete results, but explicitly gives verified facts, unresolved items, conflicting items and termination reasons, thereby improving the interpretability and security of the results.
[0029] This invention is particularly suitable for the integration of social governance incident acceptance, incident classification, risk assessment, recommendation of responsible departments, and handling basis, and can also be used in other question-and-answer scenarios that require multi-step information processing. Attached Figure Description
[0030] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a block diagram of the intelligent question-answering system based on automatic iterative reasoning provided by the present invention; Figure 2 This is a flowchart of the intelligent question-answering method based on automatic iterative reasoning provided by the present invention; Figure 3 This is a schematic diagram of the data processing relationship between state feedback generation, inference state update, and dynamic termination determination provided by the present invention. Detailed Implementation
[0031] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0032] This invention, through research, reveals that intelligent question answering for social governance not only requires the ability to "call multiple capabilities," but also an automatic iterative reasoning mechanism based on state feedback. The ReAct method interleaves reasoning trajectories with environment-oriented actions, enabling the model to update its plan and handle anomalies based on observations, laying the foundation for a common technical framework for agent loops. However, while existing technologies can form multi-round reasoning or iterative calls, they still have at least the following problems: First, the termination condition is usually the model outputting an ending action, a general judgment that "information is sufficient," or reaching a fixed maximum number of rounds; this approach does not unify the coverage of the current task objective, the validity of the observation results, the actual gain of continuous iteration, and the degree of action repetition into verifiable state variables, which can easily lead to premature termination before the evidence meets the constraints, or repeated execution in a low-gain state; Second, observation results are often directly appended to the message history as raw text, lacking a unified execution state, source identifier, anomaly type, and request identifier; when the return structures of different execution capabilities are inconsistent, the reasoning module has difficulty distinguishing between newly added facts, repeated facts, and conflicting facts. First, the lack of factual and failure feedback makes it difficult to reliably calculate the effective changes in the current execution round relative to the previous round's state. Second, some systems only have two branches: successful termination and continued execution. When execution failure, evidence conflict, or invalid actions occur, the system may continue to use the original plan and cannot treat "replanning" as an independent control state. When the maximum number of rounds is reached, it can only be forcibly truncated and cannot output downgraded results with unresolved items and evidence boundaries. Third, streaming interruptions, execution unit timeouts, or concurrency anomalies may cause execution requests to be retained in the history without corresponding observations, or isolated observations may appear. If incomplete history is directly sent to the next round of reasoning, it will not only affect the state judgment but may also cause the model interface to reject the context, making the task unrecoverable. Therefore, an automatic iterative reasoning system is needed, which uses intelligent question answering as the interaction entry point but is not limited to single-round answer generation or fixed toolchains, and can control the reasoning loop around subtask selection, state feedback, iterative updates, and dynamic termination. This system should be adaptable to scenarios such as social governance event handling, risk assessment, and departmental collaboration, while maintaining scalable execution units and recoverable reasoning processes.
[0033] Based on the above research, this invention breaks down the above complex technical problems into the following problems: (1) How to organize the task objectives, constraints, unresolved items, evidence and execution history into a continuously updated reasoning state so that the observation results of each round can form a clear state increment; (2) How to extract sub-task completion changes, evidence increments, conflicts, anomalies and duplication information based on the execution results of each round so that the execution feedback can update the selection basis for the next round of sub-tasks; (3) How to standardize heterogeneous execution results and maintain a one-to-one correspondence between execution requests and observation results so that the state feedback can be calculated consistently across execution units; (4) How to repair the reasoning context and recover from the available state when execution fails, streaming is interrupted or history is incomplete, so as to avoid incomplete rounds polluting subsequent reasoning; (5) How to retain evidence basis, unresolved items and termination reasons when terminating so that the output results have traceable and interpretable boundaries.
[0034] To address the aforementioned shortcomings, this invention provides an intelligent question-answering system and method based on automatic iterative reasoning, the technical solution of which is as follows: 1. Overall Approach; This invention does not leave the decision of "whether to continue" entirely to the generative model to decide freely using natural language. Instead, it sets up a computable state feedback layer and a dynamic termination control layer outside the inference loop. The generative model or other planner is responsible for proposing candidate inference actions; the execution unit is responsible for generating environmental feedback; the state feedback layer converts the environmental feedback into structured state increments; and the dynamic termination control layer generates control instructions based on the current state and the changes in the most recent rounds. This makes generative inference subject to deterministic state machine constraints.
[0035] In this invention, a "task" is a target description that requires one or more processing steps to obtain a result; an "execution unit" is a software component or external service capable of receiving structured parameters and returning execution results, including but not limited to information acquisition, data computation, entity or event classification, risk assessment, rule verification, and knowledge graph query. The execution unit is merely a means of generating status feedback and does not constitute the application center limited by this invention.
[0036] This invention allows the subtask selection module to be implemented using a large language model, a rule planner, a search algorithm, or a combination thereof. This invention is applicable as long as the module can select a structured target subtask based on the current inference state and accept control instructions from the dynamic termination decision module.
[0037] 2. System block diagram; Figure 1 The overall structure of the system of the present invention is shown. Solid lines represent the main data flow, and dashed lines represent the feedback information sent from the state feedback generation module to the iterative state update module. This differs from the flow that only shows the dynamic termination loop. Figure 1It is explicitly stated that the execution result must first be converted into a state feedback containing information on completion changes, evidence increments, conflicts, anomalies, and duplications; after the state feedback updates the inference state, the updated subtasks to be completed, evidence, and budget are provided in reverse to the subtask selection module. The dynamic termination decision is executed only after the state update, thus forming a verifiable closed loop.
[0038] 3. Module Function Description; 3.1 Problem Status Construction Module; The question state construction module identifies user questions, parses objectives, and extracts constraints to form a question-and-answer objective, a set of constraints, and a set of information needs. It then maps these information needs to candidate subtasks. Each candidate subtask is the smallest executable unit that drives the completion of the question-and-answer objective; examples include event element extraction, knowledge retrieval, event classification, risk assessment, or recommendation of responsible departments. The module initializes the set of completed subtasks, the set of subtasks to be completed, the evidence set, the execution history, the iteration rounds, and the resource budget. It also sets completion markers, dependencies, and conflict markers for each candidate subtask.
[0039] 3.2 Subtask Selection Module; The subtask selection module reads the current inference state, not just the original user question. Based on the priority of the subtasks to be completed, dependency satisfaction, expected information gain, available execution units, and historical failure information, the module selects one or more target subtasks, forming an execution request that includes information requirements, subtask type, execution parameters, expected feedback type, and a subtask signature. The subtask signature, generated jointly by the subtask type, target information requirements, and normalization parameters, is used for subsequent repetition detection.
[0040] When a continued iteration instruction is received, the module prioritizes subtasks whose dependency conditions are met and which can reduce the number of subtasks to be completed. When a reselection instruction is received, the module lists the most recent failure reasons, conflict evidence, or duplicate subtasks as negative constraints, prohibits the replay of invalid calls as is, and selects alternative execution units, adjusts parameters, or splits information requirements.
[0041] 3.3 Execution scheduling module and execution unit set; The execution scheduling module matches the execution unit's capability description with the expected feedback type of the candidate inference action, and can make a selection based on historical success rate, expected latency, and call cost. The module generates a unique execution request identifier for each execution and writes the request identifier, action signature, and execution parameters into the execution history.
[0042] When execution is successful, the execution unit returns facts, values, classifications, or verification conclusions; when execution fails, the execution scheduling module does not cause the entire inference process to exit abnormally directly, but generates a failure result containing the same execution request identifier, failure type, error code, and retry flag, making the failure itself a state feedback that can be replanned.
[0043] 3.4 Standardization module for observation results; The observation results standardization module maps heterogeneous execution results to a unified observation structure. This unified observation structure includes at least: execution request identifier, execution status, fact payload, source identifier, generation time, confidence information, exception information, and feedback type. The fact payload is further broken down into new facts, corrective facts, conflicting facts, and duplicate facts.
[0044] The source identifier indicates whether the observation comes from local computation, internal knowledge sources, external services, or human verification. A lower source confidence level is set for results whose source cannot be verified; for results that fail to execute, the fact payload is set to empty while retaining the anomaly information.
[0045] 3.5. State feedback generation and iterative state update module; The status feedback generation module first associates standardized observations with the target subtask based on the execution request identifier. Then, it compares these observations with the previous round of inference status, extracting five types of feedback: changes in the completion of the target subtask, evidence increment, evidence conflict, execution anomalies, and subtask duplication information. Status feedback is not equivalent to the raw returned text; rather, it is a structured description of whether the current execution round has been completed, whether evidence has been added or corrected, whether conflicts have been introduced, and whether failures or duplications have occurred.
[0046] The iterative state update module updates the set of completed subtasks, the set of subtasks to be completed, the set of evidence, the set of conflicts, the exception information, and the execution history based on the state feedback. It also calculates the subtask completion rate, evidence sufficiency, iteration gain, subtask redundancy, number of subtasks to be completed, and remaining resource budget. The updated inference state serves as the common input for the next round of subtask selection and dynamic termination decision, ensuring that the execution feedback can actually change the next step, rather than just appearing as supplementary text.
[0047] 3.6 Dynamic Termination Determination Module; The dynamic termination determination module performs combined judgments on status indicators and outputs four types of control instructions.
[0048] Successful termination: The subtask completion rate reaches the completion threshold, the evidence sufficiency reaches the sufficiency threshold, and there are no unresolved items or high-level conflicts.
[0049] Continue iterating: The task has not yet met the conditions for successful termination, but the current round generates positive iterative gains, the subtask redundancy is low, and the resource budget is sufficient to support the next round of execution.
[0050] Subtask reselection occurs when the current round of execution fails, a high-level conflict occurs, the current subtask has not reduced the number of pending subtasks, or the subtask redundancy has increased but has not yet met the downgrade termination condition. Subtask reselection is not a simple re-invocation, but rather a change in the subtask type, execution unit, parameters, or problem decomposition method.
[0051] Degradation Termination: Termination occurs when the information gain falls below a preset number of rounds, the repetition rate reaches a repetition threshold, resource budget is exhausted, or an unrecoverable anomaly occurs. During degradation termination, the currently verified facts, unresolved items, conflicting items, and termination reason are still output, instead of being disguised as a complete success result.
[0052] 3.7 Historical Consistency Repair and State Persistence Module; The historical consistency repair module scans the execution history before generating each candidate inference action. For an action record containing one or more execution requests, the complete round is retained only if all execution request identifiers have corresponding observations; action records lacking observations and their incomplete results are excluded from the current inference context; isolated observations without corresponding execution requests are also excluded.
[0053] The above-mentioned repair method preferably applies only to the context copy provided to the subtask selection module, and does not directly delete the original persistent records, in order to preserve audit trails. The state persistence module saves the state of each round according to the task session identifier, and avoids repeatedly establishing state storage connections when the same task is first resumed through concurrent initialization mutual exclusion and asynchronous writing.
[0054] 3.8 Result Generation Module; The results generation module generates structured task results based on the reasoning state at the time of termination. Upon successful termination, the task results include the final conclusion, supporting evidence, key execution trajectories, and constraint satisfaction status. Upon degraded termination, the task results, in addition to the currently verifiable conclusions, also include unresolved items, conflicting items, attempted actions, and the reason for termination. A correspondence is established between factual statements and evidence identifiers in the results.
[0055] 4. System Flow Description; like Figure 2 As shown, the intelligent question-answering method based on automatic iterative reasoning for the above system includes the following steps: Step S101: Receive task input, break down the task objective into requirement units, extract constraints, and initialize the zeroth round of reasoning state.
[0056] Step S102: Perform pairing integrity verification on the persistent execution history to form a context copy that can be used for this round of inference.
[0057] Step S103: Generate a set of candidate subtasks, and select the target subtask from the subtasks to be completed to form the corresponding subtask signature.
[0058] Step S104: Select an execution unit based on the expected feedback type, generate an execution request identifier, and execute the candidate inference action.
[0059] Step S105: Convert the execution result or execution exception into a standardized observation associated with the execution request identifier.
[0060] Step S106: Associate the standardized observation results with the target sub-task and the previous round of reasoning state, and extract information on sub-task completion changes, new or corrected evidence, evidence conflicts, execution anomalies and repeated executions to form state feedback quantities.
[0061] Step S107 updates the set of completed subtasks, the set of uncompleted subtasks, the set of evidence, the set of conflicts, the anomaly information, the execution history, and the resource budget based on the state feedback. This step transforms the observations of this round into inference states that can be directly read for the next round of subtask selection.
[0062] Step S108: Calculate the subtask completion rate, evidence sufficiency, iteration gain, subtask repetition rate, number of subtasks to be completed, and number of consecutive low-gain rounds.
[0063] Step S109: Output the following instructions based on the dynamic termination rules: successful termination, continue iteration, reselect subtask, or downgrade termination. Continue iteration and return to step S103, selecting the next target subtask based on the updated inference state; for reselecting subtasks, first write the reasons for failure, conflict, or low gain into the constraints, and then regenerate the candidate subtask set.
[0064] Step S110: If the termination is successful, a complete question and answer result is generated; if the termination is downgraded, a partial question and answer result with evidence boundaries, unresolved items, and termination reasons is generated.
[0065] The data processing relationship between state feedback generation, inference state update, and dynamic termination determination can be further represented as follows: Figure 3 As shown.
[0066] 4.1 Symbols, Variables, and Decision Formulas; To ensure the reproducibility of state feedback and dynamic termination determination, the following symbols are defined as shown in Table 1. Indicates the iteration round, using Indicates the demand unit number.
[0067] Table 1: Symbols and Variables
[0068] State updates are performed by the state merging function. and state difference function Finish: (1) (2) The completion rate of a subtask is determined by the weighted average of the completed subtasks: (3) The sufficiency of evidence can be determined by the credibility of the source. Cross-source consistency Success of execution and constraint satisfaction Weighted average, minus conflict penalty. : (4) in, And each weight is not less than 0.
[0069] The iterative gain comprehensively considers the increment of subtask completion, the increment of evidence sufficiency, and the proportion of newly added facts. Subtask redundancy: (5) The subtask repetition rate can be taken as the maximum similarity between the current subtask signature and the historical subtask signatures: (6) in, , The output range is 0 to 1; The conditions for successful termination are: (7) Continuous low gain state according to accumulation .when , or When a downgrade is triggered, termination occurs; when there is a recoverable execution exception, evidence conflict, or If no new facts are added to the current subtask, the subtask is reselected; if the remaining states do not satisfy equation (7) and still have positive gain, the iteration continues. It should be noted that the priority order is: successful termination → downgrade when there is an unrecoverable exception / insufficient budget / reaching continuous low gain or high repetition conditions → reselect when there is a recoverable exception, conflict or single round of low gain → continue for the rest, and handle overlapping conditions.
[0070] 5. Key technical parameters are shown in Table 2; Table 2: Key Technical Parameters
[0071] The above parameters can be adaptively adjusted according to the task risk level. Higher risk tasks require higher adjustment levels. and This reduces the allowable conflict tolerance; low-latency tasks can reduce... or Adjusting the parameters does not change the status feedback and dynamic termination control main link of this invention.
[0072] 6. A description of consistency with existing system implementations; This invention is based on the existing intelligent agent system's automatic looping, registration execution unit, tool result messages, session state saving, and historical integrity restoration. Furthermore, it explicitly sets up state feedback generation, iterative state updates, and dynamic termination control. The correspondence between the two is as follows.
[0073]
[0074] The first five items are engineering foundations that can be directly provided by existing projects; the structured feedback indicators and the four-branch controller for successful termination, continued iteration, reselection of subtasks, and degraded termination in the last item are patented enhancements on existing loops. This distinction does not change the integrity of the technical solution of this invention, while avoiding the representation of scoring variables that have not yet been explicitly encoded as existing code facts.
[0075] It should be noted that the key points of this invention are: First, it uses a structured reasoning state containing candidate subtasks, completed subtasks, and uncompleted subtasks as a unified carrier for automatic iteration, rather than simply storing natural language messages. Second, it associates the target subtask, standardized observation results, and the previous round of reasoning state to extract completion changes, evidence increments, conflicts, anomalies, and repetition information, and uses state feedback to drive updates to the subtask set, evidence set, conflict information, and resource budget. Third, it reselects the next target subtask based on the updated reasoning state and makes a combined judgment based on completion, sufficiency, iteration gain, repetition, uncompleted subtasks, anomaly information, and resource budget, outputting four types of control instructions: successful termination, continued iteration, reselection of subtask, or downgraded termination. Fourth, it identifies stagnant states using continuous low gain and action repetition, and proactively downgrades to termination before the budget is exhausted, avoiding reliance solely on a fixed maximum number of iterations. Fifth, it triggers replanning when execution fails, evidence conflicts occur, or actions are invalid, continuing iteration by changing the action type, execution unit, parameters, or task decomposition method. Sixth, it identifies and pairs execution requests and observation results; it repairs incomplete history before inference and restores from the most recently paired complete state. Seventh, it explicitly carries evidence boundaries, unresolved items, and termination reasons in the downgraded termination results, making partial results distinguishable from complete results. Eighth, intelligent question answering serves only as a form of user interaction and result presentation; the independent protection main line remains the closed-loop relationship between subtask selection, status feedback, iterative updates, and dynamic termination, and is not limited by a particular search service or business tool.
[0076] The following section provides further explanation based on specific implementation details.
[0077] Example 1: This embodiment provides an intelligent question answering system based on automatic iterative reasoning, including a question state construction module, a subtask selection module, an execution scheduling module, an observation result standardization module, a state feedback generation module, an iterative state update module, a dynamic termination determination module, and an answer generation module; The question state construction module is used to construct the initial inference state based on the user's question. The inference state includes at least the question-answering goal, constraint set, candidate subtask set, uncompleted subtask set, evidence set, execution history, and resource budget. The subtask selection module is used to determine the target subtask for the next round from the candidate subtask set or the set of subtasks to be completed based on the current round inference state. The target subtask includes at least the information requirement to be met, the subtask type, the execution parameters, and the expected feedback type. The execution scheduling module is used to select and trigger the execution unit according to the target sub-task. The execution unit is at least one of the following: information acquisition unit, knowledge retrieval unit, entity extraction unit, event classification unit, risk assessment unit, department recommendation unit, calculation unit, rule verification unit, or external service unit. The observation results standardization module is used to convert the heterogeneous results returned by the execution unit into standardized observation results, i.e., a unified observation structure, which includes at least the execution request identifier, execution status, fact payload, source identifier, generation time, confidence information, exception information, and feedback type. The status feedback generation module is used to associate standardized observation results with the current round of reasoning status and target sub-tasks, and extract information on sub-task completion changes, evidence increments, evidence conflicts, execution anomalies and repeated executions to form the status feedback quantity for the next round. The iterative state update module is used to update the set of completed subtasks, the set of subtasks to be completed, the set of evidence, the set of conflicts, the execution history, and the resource budget based on the state feedback, so as to form the inference state for the next round. The dynamic termination judgment module is used to combine and judge the status indicators and output a successful termination instruction, a continue iteration instruction, a reselection of subtasks instruction, or a downgraded termination instruction. Specifically, a successful termination instruction is output when the subtask completion reaches the completion threshold, the evidence sufficiency reaches the sufficiency threshold, and there are no unresolved items or high-level conflicts. A continue iteration instruction is output when the task has not yet met the successful termination conditions, but the current round generates positive iteration gain, the subtask redundancy is low, and the resource budget is sufficient to support the next round of execution. A reselection of subtasks instruction is output when the current round fails, a high-level conflict occurs, the current subtask does not reduce the number of subtasks to be completed, or the subtask redundancy increases but has not yet met the downgraded termination conditions. A downgraded termination instruction is output when the information gain for a consecutive preset number of rounds is lower than the gain threshold, the redundancy reaches the redundancy threshold, the resource budget is exhausted, or an unrecoverable anomaly occurs. The subtask selection module responds to the continue iteration instruction by selecting the next target subtask from the updated set of subtasks to be completed; responds to the reselect subtask instruction by updating the candidate subtask set based on execution exceptions, evidence conflicts, or low gain reasons and then selecting the next target subtask, until the dynamic termination judgment module outputs a successful termination instruction or a downgraded termination instruction; the answer generation module generates question-and-answer results associated with the evidence set based on the reasoning state at the time of termination.
[0078] Specifically, the problem status construction module parses the user's problem into multiple information requirements and generates candidate subtasks based on the information requirements. It records the priority, completion mark, dependency relationship, supporting evidence and conflict mark for each candidate subtask. The subtask selection module prioritizes candidate subtasks whose dependency conditions have been met and which are expected to reduce the number of subtasks to be completed.
[0079] Specifically, the status feedback generation module compares the current standardized observation results with the historical evidence set, divides the fact load into new facts, corrective facts, conflicting facts, and duplicate facts, and generates completion status feedback indicating whether the target subtask has changed from incomplete to completed, remains incomplete, or changed to abnormal.
[0080] Specifically, the iterative state update module uses the state difference function to generate a state increment relative to the current round of reasoning state, and uses the state merging function to write the state increment into the next round of reasoning state, so that the selection of the target sub-task in the next round takes the updated set of sub-tasks to be completed, the set of evidence, and the abnormal information as input.
[0081] Specifically, the dynamic termination judgment module outputs a successful termination command when the subtask completion rate is not lower than the completion threshold, the evidence sufficiency is not lower than the sufficiency threshold, the number of subtasks to be completed is zero, and there are no unresolved high-level conflicts; it outputs a downgrade termination command when the iteration gain of a consecutive preset number of rounds is not higher than the gain threshold, the subtask repetition is not lower than the repetition threshold, or the remaining budget is insufficient to support the expected cost of the next round; and it outputs a reselection subtask command when there is an execution anomaly, evidence conflict, or the current target subtask cannot reduce the number of subtasks to be completed.
[0082] Specifically, the system also includes a history consistency repair module. Before providing the execution history to the subtask selection module, the history consistency repair module verifies whether each execution request identifier has a corresponding observation result; it excludes incomplete execution rounds that lack corresponding observation results and isolated observation results that lack corresponding execution requests from the current inference context copy, and forms a paired complete inference context without changing the original persistent state.
[0083] Specifically, the system also includes a state persistence module, which saves the inference state, execution request and standardized observation results of each round according to the task session identifier, and continues to iterate from the most recent complete pair of inference states when the task resumes.
[0084] Specifically, the execution scheduling module selects execution units based on the expected feedback type of the target subtask, the description of the execution unit's capabilities, historical success rate, expected latency, and invocation cost. When an execution unit invocation fails, it still generates standardized observation results with execution request identifiers and exception information, and provides the exception information as status feedback to the subtask selection module.
[0085] Example 2: This embodiment provides an intelligent question-answering method based on automatic iterative reasoning applied to the above system, including: S1. Construct an initial reasoning state based on the user's question, including the question-answering goal, constraint set, candidate subtask set, uncompleted subtask set, evidence set, execution history, and resource budget; S1. Construct an initial reasoning state based on the user's question, including the question-answering goal, constraint set, candidate subtask set, uncompleted subtask set, evidence set, execution history, and resource budget; S2. Select the target subtask based on the current inference state, and select an execution unit to execute the target subtask; S3. Convert the execution results into standardized observations with execution request identifiers; S4. Based on standardized observation results, target sub-tasks and the previous round of inference state, extract information on sub-task completion changes, evidence increments, evidence conflicts, execution anomalies and repeated executions to form state feedback quantities, and update the current inference state according to the state feedback quantities. S5. Make a combined judgment based on the updated subtask completion rate, evidence sufficiency, iteration gain, subtask redundancy, number of subtasks to be completed, execution anomaly information, and resource budget; S6. When the combined judgment result is to continue iteration, select the next target subtask from the updated set of subtasks to be completed and return to step S2; when the combined judgment result is to reselect a subtask, update the candidate subtask set according to the state feedback quantity and return to step S2; when the combined judgment result is successful termination or downgraded termination, generate the question and answer result according to the current reasoning state.
[0086] Example 3: This embodiment provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described intelligent question-answering method based on automatic iterative reasoning.
[0087] Example 4: Automatic analysis of social governance events; The system receives a user question: "A shop in a certain community has been encroaching on the road for a long time and has caused disputes with residents. Which department should handle this, and what are the risks?" The problem status construction module identifies that this problem cannot be reliably completed in a single generation and breaks it down into four candidate sub-tasks: event element completion, event classification, risk assessment, and recommendation of responsible department. It also records missing locations, time conflicts, and other content as constraints or information requirements to be completed.
[0088] Let the weights of the four demand units be 0.25, 0.20, 0.35, and 0.20 respectively. , , , , Initial resource budget Each implementing unit.
[0089] In the first round, the subtask selection module selects "Event Element Completion" and executes the event information extraction unit. Standardized observation results return the event time, location, subject, and behavioral description. The status feedback generation module generates feedback such as "Event Element Completion changed from incomplete to complete," "Four new facts added," "No conflict," and "Execution successful." Based on this, the iterative status update module moves the subtask into the completed set, obtaining... Since the source of the result is clear and the execution was successful, When the iteration gain is positive, the dynamic termination determination module outputs to continue iteration.
[0090] In the second round, the system selects event classification subtasks based on the updated set of subtasks to be completed, calls the event classification unit, and obtains the classification results. These results are consistent with existing evidence, and the status feedback causes the event classification subtasks to be marked as completed. , The system continues to iterate.
[0091] In the third round, the comprehensive risk assessment unit returned a risk level, but the location field used conflicted with the high-confidence event record. The status feedback generation module generated feedback of "risk subtask still incomplete" and "location evidence conflict," and the iterative status update module wrote the conflict item into the inference status and reduced it. The dynamic termination judgment module outputs a reselected subtask. The subtask selection module does not repeat the original risk subtask, but instead adds a new subtask and prioritizes address verification subtasks.
[0092] In the fourth round, the address verification results corrected the administrative divisions used in the risk calculation, resolving the conflict. In the fifth round, the risk calculation was re-executed and the risk requirement units were met. In the sixth round, the responsible departments were matched, and all four requirement units were met. , , If equation (7) is satisfied, the system will successfully terminate and generate the judgment result.
[0093] In this embodiment, event information extraction, event classification, comprehensive risk assessment, and department recommendation are all execution units that can be registered in existing projects. The core of the system lies in converting the results into comparable status feedback after each round of execution, using the feedback to update the set of sub-tasks for the next round, and controlling the next step by dynamically determining the results, rather than controlling the process by a question-and-answer interface or a fixed toolchain.
[0094] Example 5: Stagnation detection in fault diagnosis; The system receives device anomaly diagnosis issues and sets "obtain logs," "identify the abnormal component," and "verify the repair solution" as candidate subtasks. In two consecutive rounds, the subtask selection module requested essentially the same log range due to different parameter expressions. After subtask signature normalization, the similarity between the two rounds reached [value missing]. Furthermore, the proportion of new facts is close to 0, leading to .
[0095] During the initial low gain, the dynamic termination decision module outputs a reselection subtask, requesting a change to the diagnostic subtask and adding a negative constraint to the requested log range. If low gain persists for two consecutive rounds after reselection, then... If the system outputs "Terminate," it will no longer execute indefinitely. The degradation result lists the verified abnormal time period, the unidentified abnormal component, the attempted log subtasks, and the termination reason for "continuous low gain."
[0096] Example 6: Interruption recovery and historical pairing repair; A streaming task experienced a connection interruption when the execution request had been written to the state but the execution result had not yet been written back. After the task restarted, the history consistency repair module found that the execution request identifier had no corresponding observation, so it excluded the incomplete round from the current inference context copy and loaded the complete state of its previous round.
[0097] If the original execution unit supports idempotent retries, the system can re-execute it with a new execution request identifier under the reselection subtask instruction; if retries are not supported, an alternative execution unit is selected. The original incomplete rounds are still retained in the persistent audit log, but they do not participate in the calculation of the model context and state indicators in the next round, thereby avoiding isolated execution requests that could lead to subsequent inference failures.
[0098] It should be noted that the subtask completion degree of the present invention can be calculated either by the weight of the requirement unit or by the weighted proportion of the completed nodes in the task directed graph; the sufficiency of evidence can be calculated by rules or by the output of the verification model and then normalized to the interval; the action repetition degree can be calculated by string similarity, parameter vector similarity or semantic vector similarity.
[0099] It should be noted that the dynamic termination determination module of the present invention can be implemented using deterministic rules, finite state machines, decision trees or calibrated classification models, but at least four distinguishable control results should be retained: successful termination, continued iteration, reselection of subtasks and degraded termination, and triggered by the updated state feedback index.
[0100] It should be noted that the state persistence of this invention can employ relational databases, key-value databases, or graph state storage. For long-running tasks, only the most recent state can be stored. A complete record of each pair is provided to the subtask selection module, while the complete execution trajectory is preserved in the persistent state. The pruning operation preferably uses the complete execution request-observation round as the smallest unit to avoid truncating from the middle of the round.
[0101] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. An intelligent question-answering system based on automatic iterative reasoning, characterized in that, It includes a problem state construction module, a subtask selection module, an execution scheduling module, an observation result standardization module, a state feedback generation module, an iterative state update module, a dynamic termination determination module, and an answer generation module; The question state construction module is used to construct an initial reasoning state based on the user's question. The reasoning state includes at least the question-answering goal, constraint set, candidate subtask set, subtask to be completed set, evidence set, execution history, and resource budget. The subtask selection module is used to determine the target subtask for the next round from the candidate subtask set or the subtask set to be completed based on the current round reasoning state. The target subtask includes at least the information requirement to be met, the subtask type, the execution parameters, and the expected feedback type. The execution scheduling module is used to select an execution unit and trigger execution according to the target sub-task. The execution unit is at least one of the following: information acquisition unit, knowledge retrieval unit, entity extraction unit, event classification unit, risk assessment unit, department recommendation unit, calculation unit, rule verification unit, or external service unit. The observation result standardization module is used to convert the heterogeneous results returned by the execution unit into standardized observation results, which include at least execution request identifier, execution status, fact payload, source identifier, generation time, confidence information, anomaly information and feedback type, i.e., a unified observation structure; The state feedback generation module is used to associate the standardized observation results with the current round of reasoning state and the target sub-task, and extract information such as sub-task completion changes, evidence increments, evidence conflicts, execution anomalies and repeated executions to form the state feedback quantity for the next round. The iterative state update module is used to update the set of completed subtasks, the set of subtasks to be completed, the set of evidence, the set of conflicts, the execution history, and the resource budget according to the state feedback quantity, so as to form the next round of reasoning state; The dynamic termination judgment module is used to combine and judge the status indicators and output a successful termination instruction, a continue iteration instruction, a reselection of subtasks instruction, or a downgraded termination instruction. Specifically, a successful termination instruction is output when the subtask completion reaches the completion threshold, the evidence sufficiency reaches the sufficiency threshold, and there are no unresolved items or high-level conflicts. A continue iteration instruction is output when the task has not yet met the successful termination conditions, but the current round generates positive iteration gain, the subtask redundancy is low, and the resource budget is sufficient to support the next round of execution. A reselection of subtasks instruction is output when the current round fails, a high-level conflict occurs, the current subtask does not reduce the number of subtasks to be completed, or the subtask redundancy increases but has not yet met the downgraded termination conditions. A downgraded termination instruction is output when the information gain for a consecutive preset number of rounds is lower than the gain threshold, the redundancy reaches the redundancy threshold, the resource budget is exhausted, or an unrecoverable anomaly occurs. The subtask selection module responds to the continue iteration instruction by selecting the next target subtask from the updated set of subtasks to be completed; responds to the reselect subtask instruction by updating the candidate subtask set according to the reasons of execution exception, evidence conflict or low gain and then selecting the next target subtask, until the dynamic termination judgment module outputs a successful termination instruction or a downgraded termination instruction; the answer generation module generates a question and answer result associated with the evidence set based on the reasoning state at the time of termination.
2. The system according to claim 1, characterized in that, The problem status construction module parses the user's problem into multiple information requirements and generates candidate subtasks based on the information requirements. It records the priority, completion mark, dependency relationship, supporting evidence and conflict mark for each candidate subtask. The subtask selection module prioritizes candidate subtasks whose dependency conditions have been met and which are expected to reduce the number of subtasks to be completed.
3. The system according to claim 1, characterized in that, The state feedback generation module compares the current standardized observation results with the historical evidence set, divides the fact load into new facts, corrective facts, conflicting facts and duplicate facts, and generates completion state feedback indicating whether the target subtask has changed from incomplete to completed, remains incomplete or changed to abnormal.
4. The system according to claim 1, characterized in that, The iterative state update module uses a state difference function to generate a state increment relative to the current round of reasoning state, and uses a state merging function to write the state increment into the next round of reasoning state, so that the selection of the target sub-task in the next round takes the updated set of sub-tasks to be completed, the set of evidence, and the anomaly information as input.
5. The system according to claim 1, characterized in that, The dynamic termination determination module outputs a successful termination command when the subtask completion is not lower than the completion threshold, the evidence sufficiency is not lower than the sufficiency threshold, the number of subtasks to be completed is zero, and there are no unresolved high-level conflicts; it outputs a downgrade termination command when the iteration gain of a consecutive preset number of rounds is not higher than the gain threshold, the subtask repetition is not lower than the repetition threshold, or the remaining budget is insufficient to support the expected cost of the next round; and it outputs a reselection subtask command when there is an execution anomaly, evidence conflict, or the current target subtask cannot reduce the number of subtasks to be completed.
6. The system according to claim 1, characterized in that, It also includes a history consistency repair module, which verifies whether each execution request identifier has a corresponding observation before providing the execution history to the subtask selection module; excludes incomplete execution rounds that lack corresponding observations and isolated observations that lack corresponding execution requests from the current inference context copy, and forms a paired complete inference context without changing the original persistent state.
7. The system according to claim 1, characterized in that, It also includes a state persistence module, which saves the inference state, execution request and standardized observation results of each round according to the task session identifier, and continues to iterate from the most recent complete pair of inference states when the task resumes.
8. The system according to claim 1, characterized in that, The execution scheduling module selects execution units based on the expected feedback type of the target subtask, the description of the execution unit's capabilities, historical success rate, expected latency, and invocation cost. When an execution unit invocation fails, it still generates standardized observation results with execution request identifiers and exception information, and provides the exception information as status feedback to the subtask selection module.
9. An intelligent question-answering method based on automatic iterative reasoning, characterized in that, include: S1. Construct an initial reasoning state based on the user's question, including the question-answering goal, constraint set, candidate subtask set, uncompleted subtask set, evidence set, execution history, and resource budget; S2. Select a target subtask based on the current inference state, and select an execution unit to execute the target subtask; S3. Convert the execution results into standardized observations with execution request identifiers; S4. Based on standardized observation results, target sub-tasks and the previous round of inference state, extract sub-task completion changes, evidence increments, evidence conflicts, execution anomalies and repeated execution information to form a state feedback quantity, and update the current inference state according to the state feedback quantity. S5. Make a combined judgment based on the updated subtask completion rate, evidence sufficiency, iteration gain, subtask redundancy, number of subtasks to be completed, execution anomaly information, and resource budget; S6. When the combined judgment result is to continue iteration, select the next target subtask from the updated set of subtasks to be completed and return to step S2; when the combined judgment result is to reselect a subtask, update the candidate subtask set according to the state feedback quantity and return to step S2; when the combined judgment result is successful termination or downgraded termination, generate the question and answer result according to the current reasoning state.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent question-answering method based on automatic iterative reasoning as described in claim 9.
