Incomplete data trend question and answer method based on multi-agent collaboration

Abstract

The invention discloses an incomplete data trend question and answer method based on multi-agent collaboration. The method comprises the steps of obtaining natural language query and missing original table data; constructing a sequential game model comprising an extractor and a corrector, and performing interactive reasoning and positioning on a target sub-table by using associated information in a table; aiming at time sequence data missing, constructing an incomplete information game model containing a row filler and a column filler, and respectively generating filling sub-tables based on an attribute association view and a time sequence trend view; and after respectively executing trend calculation, the coordinator performs comprehensive decision on the filling result and the trend conclusion and outputs the filling result and the trend conclusion. According to the method, the lightweight meta controller and the staged training strategy are introduced, so that the problems of difficulty in sub-table positioning and inaccuracy in missing data filling are effectively solved under the condition of no exogenous knowledge assistance, and the robustness of incomplete data trend question answering is improved.

Description

A Question Answering Method for Incomplete Data Based on Multi-Agent Collaboration Technical Field

[0001] This invention belongs to the field of table-based question-and-answer technology, specifically relating to a method for question-and-answering incomplete data trends based on multi-agent collaboration. Background Technology

[0002] Table QA technology aims to enable non-technical users to analyze structured data through natural language queries. In real-world applications, trend analysis questions are a high-frequency query type in table QA. For example, in business intelligence scenarios, users frequently ask questions such as "What are the sales trends in different regions over the past five years?" These queries require the system not only to retrieve numerical values ​​but also to analyze the trend changes of indicators over time.

[0003] Our research revealed that Existing Technology 1 proposes a Text-to-SQL (Natural Language to Structured Query Language) question intent classification framework to identify unanswerable questions due to missing metadata. This method focuses on missing information identification, suspending the semantic parsing process and indicating missing information when the current database metadata is insufficient to support the question. However, this technology lacks a mechanism for semantic reasoning based on context or data content to proactively locate potential valid data sub-tables, forcing the system to suspend the response when faced with incomplete metadata. Existing Technology 2 proposes a data cleaning framework based on retrieval-enhanced generation. This method transforms the missing value imputation task into a "retrieve first, then reason" process, retrieving relevant tuples from an expert-provided data lake and using a large language model to infer the correct value. However, in real-world business scenarios, due to the high degree of closure of private data and the lack of a general knowledge base or external data lake for matching, the imputation effect of such methods drops significantly when external auxiliary information is missing or relevant reference data cannot be retrieved, making it difficult to meet analytical needs. Existing technology 3 proposes an iterative cleaning framework based on a large language model. This method corrects errors through multiple rounds of "detection-verification". However, its feedback mechanism is mainly limited to data compliance at the micro level and lacks closed-loop feedback for macro trend analysis. In trend analysis scenarios, the accuracy of data repair depends not only on whether the values ​​are compliant, but also on whether they conform to historical time series characteristics. This technology cannot guide data filling based on the final trend pattern, which means that even if the data is cleaned, the generated trend analysis results may still be misleading due to numerical drift.

[0004] In summary, existing table-based question-answering technologies face the following three major challenges when handling trend analysis tasks in real-world scenarios: 1. Missing metadata makes it difficult to locate sub-tables: Semantic parsing methods or end-to-end models usually assume that the table metadata is complete. However, in real business data, missing metadata often leads to query conditions failing to match valid rows, resulting in data omissions or positioning errors, which affect subsequent analysis.

[0005] 2. Challenges in Missing Value Imputation for Trend Analysis: Missing value imputation faces the dilemma of "lack of external knowledge." Due to the lack of a general knowledge base to support private data, existing mainstream imputation methods are unable to accurately imput missing values ​​for trend analysis.

[0006] 3. Difficulty in trend analysis for incomplete data: Traditional serial processing lacks fault tolerance and self-correction mechanisms; the existing one-way process of "first locate, then fill, then analyze" has the problem of error propagation. Once the preceding steps deviate, the subsequent modules cannot detect and correct it, resulting in the error being amplified step by step, which seriously affects the accuracy of the final trend analysis results.

[0007] Therefore, there is an urgent need for a method that can make full use of the internal information of the data itself, and achieve accurate trend analysis of incomplete data through intelligent game and collaboration mechanisms, even in the absence of external knowledge assistance. Summary of the Invention

[0008] In view of the above, the present invention provides a method for question answering incomplete data trends based on multi-agent collaboration. This method, without the need for external knowledge base assistance, fully utilizes the internal information of the table by combining sequential game and incomplete information game multi-agent collaboration mechanism to accurately locate sub-tables of incomplete data, repair missing values, and analyze trends.

[0009] A method for answering incomplete data trend questions based on multi-agent collaboration includes the following steps: (1) obtaining a natural language trend analysis query request and the original time-series table data to be analyzed (containing metadata columns and time-series data columns, and there are cases where metadata is missing or time-series data is missing); (2) constructing a multi-agent system consisting of an extractor, a corrector, a row filler, a column filler, and a coordinator; (3) through the interaction of the extractor and the corrector, using the metadata association information of the table itself to perform reasoning, locating and extracting the target sub-table corresponding to the trend analysis query request; (4) targeting the time-series data in the target sub-table. For missing values ​​in the sequence data, the row filler based on the attribute association view and the column filler based on the time series trend view are used to play a game to generate row view filler sub-tables and column view filler sub-tables respectively; (5) Design a pluggable trend analysis tool interface, call the trend analysis tool to perform trend calculation on the row view filler sub-tables and column view filler sub-tables, and obtain the row view trend analysis results and column view trend analysis results respectively; (6) Configure the coordinator to make a comprehensive decision on the row view filler sub-tables, column view filler sub-tables, row view trend analysis results and column view trend analysis results, generate and output the final complete target sub-table and trend analysis results.

[0010] Further, the interaction process between the extractor and the corrector in step (3) is as follows: S31: The extractor acts as the first actor in the sequential game, generates an initial query statement based on the explicit constraints in the trend analysis query request according to the extraction strategy, and extracts a preliminary sub-table from the original time series table data; S32: The corrector acts as the second actor in the sequential game, observes the preliminary sub-table, and identifies the implicit association information of the metadata (including the attribute transfer relationship of the same entity under different records); S33: The corrector generates a corrected query constraint based on the correction strategy and the implicit association information; S34: The final query is executed based on the corrected query constraint, and the obtained result is determined as the target sub-table; The extraction strategy and the correction strategy are strategies that are iteratively optimized with the goal of approaching the perfect Nash equilibrium of the subgame.

[0011] Furthermore, the game process between the row filler and the column filler in step (4) is as follows: the column filler generates a first filling result from the time-series trend view based on the state of the target sub-table and the column filling strategy, forming a column view filled sub-table; the row filler generates a second filling result from the attribute column association view based on the state of the target sub-table and the row filling strategy, forming a row view filled sub-table; the column filling strategy and the row filling strategy are strategies that are iteratively optimized with the goal of approaching Bayesian Nash equilibrium.

[0012] Furthermore, in step (6), the coordinator, based on the game consensus mechanism, comprehensively evaluates the filling results of the row view and column view, and comprehensively verifies the trend analysis results of the row view and column view; based on the evaluation and verification results, the globally optimal filling value and trend analysis conclusion are determined as the final complete target sub-table and trend analysis results and output.

[0013] Furthermore, the extractor, corrector, row filler, column filler, and coordinator, as agents in the multi-agent system, are all built based on a large language model; except for the coordinator which remains in an untrained state, each of the other agents is equipped with a lightweight meta-controller based on the Actor-Critic architecture.

[0014] Furthermore, the lightweight meta-controller is obtained through reinforcement learning training. It maps game theory equilibrium conditions to reinforcement learning optimization objectives to achieve the solution. Its training process aims to drive the agent's policy towards game equilibrium. Specifically: the meta-controller is constructed as a network structure containing a policy network and a value network, configured to receive current task state information as input. The control parameters output by the policy network include the generation temperature parameters and kernel sampling parameters of the large language model, while the value network outputs the value assessment value of the current state. The parameters of the large language model are kept frozen, and the smoothness or sampling range of the output probability distribution of the large language model is adjusted only according to the control parameters, thereby guiding the generation strategy of the large language model with extremely low computational cost. During the training phase, the advantage function is calculated using the value assessment value combined with the reward from environmental feedback to evaluate the merits of the current action. The policy gradient is calculated using the action probability output by the policy network. Based on the advantage function and policy gradient, the parameters of the meta-controller are iteratively updated only, optimizing the extractor and modifier strategies with the goal of approximating the perfect Nash equilibrium of the subgame, and optimizing the row filler and column filler strategies with the goal of approximating the Bayesian Nash equilibrium.

[0015] Furthermore, the reinforcement learning training adopts a phased training strategy. Specifically: in the first phase of training, the parameters of the meta-controllers of the extractor and the corrector are updated to drive their policies to approach the perfect Nash equilibrium of the subgame; after the first phase of training is completed, the parameters of the meta-controllers of the extractor and the corrector are frozen, and their output target sub-tables are used as inputs for subsequent phases; in the second phase of training, based on the first phase model with frozen parameters, the parameters of the meta-controllers of the row filler and the column filler are updated to drive their policies to approach the Bayesian Nash equilibrium; in this phase, the coordinator remains in a non-training state.

[0016] Furthermore, the rewards used in the first and second phases of training are composed of at least one combination of the following rewards: output likelihood reward, used to measure the consistency between the agent's output action and the collaborative consensus, specifically obtained by calculating the similarity between the current agent's output and the consensus output of the multi-agent system; task-specific reward, used to evaluate specific task metrics, used in the first phase to evaluate the accuracy of sub-table localization, and in the second phase to evaluate the accuracy of missing value imputation and trend analysis; and collaborative contribution reward, used to evaluate the degree of contribution of a single agent to the final solution of the multi-agent system.

[0017] A computer device includes a memory and a processor, wherein the memory stores a computer program and the processor executes the computer program to implement the above-described incomplete data trend question answering method based on multi-agent collaboration.

[0018] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described incomplete data trend question-answering method based on multi-agent collaboration.

[0019] This invention effectively overcomes the dependence of traditional trend analysis on data integrity and external knowledge by introducing a multi-agent collaborative mechanism. Specifically, this invention utilizes a sequential game between extractors and correctors to deeply mine the implicit metadata relationships within tables, solving the problem of sub-table location when metadata is missing. By constructing an incomplete information game between row and column fillers, it achieves accurate missing value repair by collaboratively utilizing a dual view of temporal trends and attribute constraints without external assistance. Furthermore, this invention innovatively combines a lightweight meta-controller based on reinforcement learning and a phased parameter freezing training strategy to achieve refined policy control of large models at extremely low cost, driving the multi-agent system to efficiently approach game equilibrium. On this basis, this invention effectively blocks the error propagation effect in traditional serial processes through the dual-view analysis and comprehensive decision-making mechanism of the coordinator, significantly improving the robustness of trend analysis in complex scenarios. Attached Figure Description

[0020] Figure 1 is a schematic diagram of the incomplete data trend question answering method based on multi-agent collaboration of the present invention. Detailed Implementation

[0021] To describe the present invention in more detail, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] This embodiment provides a multi-agent collaborative method for incomplete data trend question answering. This method aims to solve the technical challenge in real-world business scenarios where missing table metadata and time-series data prevent linear analysis processes from accurately extracting data and analyzing trends. The following will illustrate this method using a specific sports data analysis scenario. The entire method flow is shown in Figure 1: Step S11: Obtain the natural language trend analysis query request and the original time-series table data to be analyzed.

[0023] In this embodiment, the scenario assumes that the user wants to analyze the performance trend of a specific object (athlete "Broady") over a period of time. Input query: For example, "Do you think Broady's rank was stable during t to t+2?", raw table data: The raw time-series table read by the system contains metadata columns (such as Time, Entity ID, Name) and time-series data columns (such as Rank, Point). Due to data quality limitations, the table contains the following data records and missing information within the time period from t to t+2: At time t, the record has ID "A2004", Rank "16", and Point "95"; however, the Name attribute of this row is missing (empty value); at time t+1, the record has ID "A2004", Name "Broady", and Point "98", but the Rank value of this row is missing (empty value); at time t+2, the record has Name "Broady", Rank "13", and Point "99", but the ID attribute of this row is missing (empty value). Therefore, the original table suffers from both missing metadata (Name missing in row 1, ID missing in row 3) and missing time-series data (Rank missing in row 2), directly hindering the use of traditional methods to locate the complete target sub-table and calculate continuous trends.

[0024] Step S12: Based on the trend analysis query request, construct a sequential game model containing an extractor and a modifier. The system initializes the extractor and modifier agents, both of which are built based on the large language model and configured with a pre-trained meta-controller. The system uses the meta-controller to output control parameters according to the current task state and dynamically adjust the generation strategy of the large language model.

[0025] Step S13: Through the interaction between the extractor and the corrector, reasoning is performed using the table's own metadata association information to locate and extract the target sub-table corresponding to the query request. This step aims to solve the difficulty in locating the target sub-table caused by missing metadata. The specific interaction process is as follows: S31 (Extractor Action): As the first player in the sequential game, the extractor generates an initial query statement based on the explicit constraints (Name='Broady') in the query request according to the extraction strategy. In this embodiment, the extractor generates SELECT * FROM Table WHERE Name='Broady'; after executing this query, only rows t+1 and t+2 are extracted, omitting row t (because Name is empty in row t).

[0026] S32 (Modifier Observation): The modifier, acting as a follower in the sequential game, observes the initial sub-table to identify implicit associations exposed due to missing metadata. In this example, the modifier discovers that ID=A2004 and Name=Broady co-occur in row t+1; it identifies the implicit association: in this table, the attribute transmission relationship of the same entity Broady across different records is characterized by ID consistency, meaning A2004 refers to Broady.

[0027] S33 (Modified Constraints): The modifier generates modified query constraints based on the implicit association information using the modification strategy. In this example, the modifier generates the modified condition WHERE Name='Broady' OR ID='A2004'.

[0028] S34 (Extract the final sub-table): Execute the final query based on the modified query constraints, and determine the result as the target sub-table. In this embodiment, the modified SQL (Structured Query Language) is executed, successfully extracting the target sub-table containing rows t, t+1, and t+2.

[0029] The extraction and modification strategies are iteratively optimized strategies aimed at achieving a perfect Nash equilibrium in the subgame. At this point, the subtable localization phase based on sequential game theory concludes, and the system obtains the complete target subtable.

[0030] Step S14: For missing time-series data values ​​in the target sub-table, construct an incomplete information game model including row fillers and column fillers. The system identifies that the Rank column of the second row (time t+1) in the target sub-table has missing values. At this time, the row filler and column filler are initialized. Both agents are built based on the large language model and configured with a pre-trained meta-controller. The system uses the meta-controller to output control parameters according to the current task state and dynamically adjust the generation strategy of the large language model.

[0031] Step S15: A game is played between the row filler based on the attribute-related view and the column filler based on the time-series trend view to generate row view filled sub-tables and column view filled sub-tables respectively. The specific game process for this step is as follows: S51 (Column View Filling): The column filler generates the first filling result from the time-series trend view based on the state of the target sub-table and the column filling strategy. In this embodiment: The column filler observes the vertical data of the Rank column, infers the missing value should be 14 based on the time-series trend, and generates the column view filled sub-table.

[0032] S52 (Row View Fill): The row filler generates a second fill result from the attribute column associated view based on the state of the target sub-table and the row fill strategy. In this example, the row filler observes the horizontal data Point=98 in row t+1, compares it with other rows where 95 points rank 16th and 99 points rank 13th, infers that 98 points should correspond to 14th place, and generates a row view to fill the sub-table.

[0033] The column filling strategy and row filling strategy are iterative optimization strategies aimed at achieving a Bayesian Nash equilibrium. At this point, the collaborative filling phase ends, and the system obtains two candidate complete tables: a column view filled sub-table and a row view filled sub-table.

[0034] Step S16: Design a pluggable trend analysis tool interface, call the trend analysis tool to perform trend calculations on the row view and column view filled sub-tables, and obtain the row view trend analysis results and column view trend analysis results respectively. The system calls the built-in trend analysis tool (such as the Mann-Kendall test algorithm or linear regression model, etc.) to calculate the trend of the ranking sequence [16, 14, 13] based on the column view complete sub-table, and the result is a decrease in value (improved performance); calculate the trend of the ranking sequence [16, 14, 13] based on the row view sub-table, and the result is a decrease in value (improved performance).

[0035] S17: Configure the coordinator, which then makes a comprehensive decision on the row view imputation sub-table, column view imputation sub-table, row view trend analysis results, and column view trend analysis results to generate and output the final complete target sub-table and the final trend analysis results. In this embodiment, the coordinator is built based on a large language model. It receives the imputed sub-table and trend analysis results as input prompts and utilizes the semantic understanding and logical reasoning capabilities of the large language model to perform the following operations: Based on a game-theoretic consensus mechanism, the coordinator evaluates the imputation results, finds that the row view and column view have reached a consensus, and decides that the final imputation value is 14. The coordinator performs a consistency check on the trend analysis results, confirms that the conclusions of the two views are consistent, and outputs the final complete table and the answer: "Broady's performance is stable and its ranking is trending upwards."

[0036] The efficient execution of the above process relies on the model architecture and training strategy of this invention. In this embodiment, the extractor, corrector, row filler, and column filler, as agents in the multi-agent system, are all built based on the large language model. Except for the coordinator, each agent is equipped with a lightweight meta-controller based on the Actor-Critic architecture. The meta-controller is obtained through reinforcement learning training, mapping the equilibrium conditions of game theory to the optimization objective of reinforcement learning to achieve the solution. Its construction and operation mechanism includes the following steps: Constructing the meta-controller: The meta-controller is constructed as a structure containing a policy network and a value network, and configured to receive current task state information as input, including current task state information, query statement encoding, current table encoding, etc.; the policy network outputs control parameters, including at least the generation temperature parameter (Temperature) and kernel sampling parameter (Top-P) of the large language model; the value network outputs the value evaluation value of the current state, used to evaluate the merits of the current situation.

[0037] Freeze large model parameters: During the training and inference phases, keep the parameters of the large language model of the agent configured with the meta-controller frozen, and adjust the smoothness or sampling range of the output probability distribution of the large language model only according to the control parameters output by the policy network.

[0038] Iterative training of the execution strategy: During the training phase, the advantage function is calculated by combining the value evaluation value output by the value network with the reward from environmental feedback to evaluate the merits of the current action; the policy gradient is calculated using the policy network, and the parameters of the meta-controller are iteratively updated only, aiming to optimize the policies of the extractor and modifier to approximate the perfect Nash equilibrium of the subgame, and to optimize the policies of the row filler and column filler to approximate the Bayesian Nash equilibrium. Specifically, the advantage function is calculated using the generalized advantage estimation method. The calculation formula is as follows: ,in The timing difference error is calculated as follows: , For intelligent agents in The total rewards earned at any given moment (including output likelihood rewards, task-specific rewards, and collaboration contribution rewards). and The value network is for the current state. and the next state Output value assessment The discount factor is used to balance short-term and long-term returns. The smoothing parameter is used to balance bias and variance. The length of the trajectory.

[0039] To address the stability issue in training multi-agent systems, this embodiment employs a phased training strategy, including: the first phase of training updates the parameters of the meta-controllers of the extractor and the corrector to drive their strategies to approximate the perfect Nash equilibrium of the subgame; after the first phase of training is completed, the parameters of the meta-controllers of the extractor and the corrector are frozen, and their output target sub-tables are used as inputs for subsequent phases.

[0040] The second-stage training is based on the first-stage model after parameter freezing. The parameters of the meta-controllers of the row filler and column filler are updated to drive their policies to approach Bayesian Nash equilibrium. During this stage, the coordinator remains in an untrained state.

[0041] The rewards used in the first and second phases of training consist of the following combination of components: Collaborative Contribution Rewards ( ): Used to evaluate the contribution of a single agent to the final solution. Specifically, it can be obtained by normalizing the score of the agent's reasoning path by a large language model judge.

[0042] Output likelihood reward ( ): Used to measure the consistency between the agent's output and the collaborative consensus. Specifically, it can be obtained by calculating the cosine similarity between the current agent's output vector and the average consensus vector of the multi-agent system. The calculation formula is: ,in The embedding vector output by the current agent. For the embedding vector of collaborative consensus in multi-agent systems, This represents the cosine similarity.

[0043] Mission-specific rewards In the first stage, the overlap rate between the extracted sub-table and the real sub-table labels can be calculated as a reward. The calculation formula is as follows: in: This represents the set of sub-table row indexes extracted by the agent. The set of row indices representing the standard answer.

[0044] In the second stage, the missing value imputation accuracy and trend analysis accuracy can be calculated as rewards, using the following formula: ,in To balance the weighting of numerical accuracy and trend consistency rewards (e.g., set to 0.5), the numerical accuracy reward... The method used to measure the absolute accuracy of the filled value is the normalized absolute error form, and the calculation formula is as follows: in: The filler value output by the agent. Labels for real values. To prevent small constants with a denominator of zero (such as...) ).

[0045] Trend Rewards To measure the accuracy of the overall evolution trend after filling, this embodiment directly obtains the time series sequence corresponding to the complete target sub-table located in the first stage. To reduce computational complexity, it is assumed that the data exhibits a linear trend, and the least squares method is used to linearly fit the sequence and extract the global slope. As a trend feature, the calculation formula is: ,in It is the slope obtained by linearly fitting the entire target sub-table sequence containing the imputed values. It is the slope obtained by linearly fitting the entire target sub-table sequence containing the true values. It is the hyperbolic tangent function, used to map the absolute error of the slope to... Interval.

[0046] It should be noted that the above linear fitting is only one specific way to calculate trend reward. In other embodiments, based on the distribution characteristics of time series data (such as exponential growth, periodic fluctuations, etc.), the trend feature extraction methods may also include, but are not limited to, nonlinear fitting parameters, correlation coefficients, etc.

[0047] Based on the phased training strategy, the comprehensive reward function for each phase is constructed using the following weighted aggregation method: Phase 1 Comprehensive Reward The calculation method is as follows: ,in The first phase's reward component weighting coefficients; the second phase's comprehensive reward... The calculation method is as follows: ,in This refers to the weighting coefficients for the reward items in the second phase.

[0048] The above description of the embodiments is provided to enable those skilled in the art to understand and apply the present invention. Those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made to the present invention by those skilled in the art based on the disclosure thereof should be within the scope of protection of the present invention.

Claims

1. A question-answering method for incomplete data based on multi-agent collaboration, characterized in that, The steps include: (1) obtaining natural language trend analysis query requests and the original time series table data to be analyzed; (2) constructing a multi-agent system consisting of an extractor, a corrector, a row filler, a column filler, and a coordinator. (3) Through the interaction between the extractor and the corrector, reasoning is performed using the metadata association information of the table itself to locate and extract the target sub-table corresponding to the trend analysis query request; (4) For the missing time series data values ​​in the target sub-table, the row filler based on the attribute association view and the column filler based on the time series trend view are used to play a game to generate the row view fill sub-table and the column view fill sub-table respectively. (5) Design a pluggable trend analysis tool interface, call the trend analysis tool to perform trend calculations on the row view fill sub-table and the column view fill sub-table, and obtain the row view trend analysis results and the column view trend analysis results respectively; (6) Configure the coordinator to make comprehensive decisions on the row view fill sub-table, column view fill sub-table, row view trend analysis results and column view trend analysis results, generate and output the final complete target sub-table and trend analysis results.

2. The incomplete data trend question answering method based on multi-agent collaboration according to claim 1, characterized in that, The interaction process between the extractor and the corrector in step (3) is as follows: S31: The extractor acts as the first actor in the sequential game, generates an initial query statement based on the explicit constraints in the trend analysis query request according to the extraction strategy, and extracts a preliminary sub-table from the original time series table data; S32: The corrector acts as the last actor in the sequential game, observes the preliminary sub-table, and identifies the implicit association information of the metadata therein; S33: The corrector generates a corrected query constraint based on the correction strategy and the implicit association information; S34: The final query is executed based on the corrected query constraint, and the obtained result is determined as the target sub-table; The extraction strategy and the correction strategy are strategies that are iteratively optimized with the goal of approaching the perfect Nash equilibrium of the subgame.

3. The incomplete data trend question answering method based on multi-agent collaboration according to claim 1, characterized in that, The game process between the row filler and the column filler in step (4) is as follows: the column filler generates the first filling result from the time series trend view based on the state of the target sub-table and the column filling strategy, forming a column view to fill the sub-table; The row filler generates a second filling result from the attribute column association view based on the state of the target sub-table and the row filling strategy, forming a row view filled sub-table; the column filling strategy and the row filling strategy are strategies that are iteratively optimized with the goal of approaching Bayesian Nash equilibrium.

4. The incomplete data trend question answering method based on multi-agent collaboration according to claim 1, characterized in that: In step (6), the coordinator, based on the game consensus mechanism, comprehensively evaluates the filling results of the row view and column view, and comprehensively verifies the trend analysis results of the row view and column view. Based on the evaluation and verification results, the globally optimal fill value and trend analysis conclusions are determined as the final complete target sub-table and trend analysis results, and then output.

5. The incomplete data trend question answering method based on multi-agent collaboration according to claim 1, characterized in that: The extractor, corrector, row filler, column filler, and coordinator, as agents in the multi-agent system, are all built based on a large language model; except for the coordinator which remains in an untrained state, each of the other agents is equipped with a lightweight meta-controller based on the Actor-Critic architecture.

6. The incomplete data trend question answering method based on multi-agent collaboration according to claim 5, characterized in that: The lightweight meta-controller is trained through reinforcement learning, mapping game theory equilibrium conditions to reinforcement learning optimization objectives to achieve the solution. Its training process aims to drive the agent's policy towards game equilibrium. Specifically: the meta-controller is constructed as a network structure containing a policy network and a value network, configured to receive current task state information as input. The control parameters output by the policy network include the generation temperature parameters and kernel sampling parameters of the large language model, while the value network outputs the value assessment value of the current state. The parameters of the large language model are kept frozen, and the smoothness or sampling range of the output probability distribution of the large language model is adjusted only according to the control parameters, thereby guiding the generation strategy of the large language model with extremely low computational cost. During the training phase, the advantage function is calculated using the value assessment value combined with the reward from environmental feedback to evaluate the merits of the current action. The policy gradient is calculated using the action probability output by the policy network. Based on the advantage function and policy gradient, the parameters of the meta-controller are iteratively updated only, optimizing the extractor and modifier strategies with the goal of approximating the perfect Nash equilibrium of the subgame, and optimizing the row filler and column filler strategies with the goal of approximating the Bayesian Nash equilibrium.

7. The incomplete data trend question answering method based on multi-agent collaboration according to claim 6, characterized in that, The reinforcement learning training adopts a phased training strategy. Specifically, in the first phase of training, the parameters of the meta-controllers of the extractor and the corrector are updated to drive their strategies to approach the perfect Nash equilibrium of the subgame. After the first phase of training is completed, the parameters of the meta-controllers of the extractor and the corrector are frozen, and the target sub-tables output by them are used as the input for the subsequent phases. In the second stage of training, based on the first stage model after parameter freezing, the meta controllers of the row filler and column filler are updated to drive their policies to approach the Bayesian Nash equilibrium. During this phase, the coordinator remains in a non-training state.

8. The incomplete data trend question answering method based on multi-agent collaboration according to claim 7, characterized in that: The rewards used in the first and second phases of training consist of a combination of at least one of the following: output likelihood reward, used to measure the consistency between the agent's output action and the collaborative consensus, specifically obtained by calculating the similarity between the current agent's output and the consensus output of the multi-agent system; task-specific reward, used to evaluate specific task metrics, used in the first phase to evaluate the accuracy of sub-table localization, and in the second phase to evaluate the accuracy of missing value imputation and trend analysis; and collaborative contribution reward, used to evaluate the degree of contribution of a single agent to the final solution of the multi-agent system.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: The processor is used to execute the computer program to implement the incomplete data trend question answering method based on multi-agent collaboration as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the incomplete data trend question answering method based on multi-agent collaboration as described in any one of claims 1 to 8.

Images