Intelligent decision-making method, device and equipment based on entropy guidance and medium

Abstract

The invention relates to the technical field of data processing, and discloses an intelligent decision-making method, device and equipment based on entropy guidance and a medium, which can be applied to the financial and medical fields, and the method comprises the following steps: obtaining a semantic representation object, driving a preset language model to reasone and calculating information entropy; respectively constructing an entropy monitoring data object and a reasoning control parameter to carry out multi-step reasoning, and fusing to obtain a candidate reasoning result set; retrieving the knowledge graph, generating a knowledge enhancement context, inputting the knowledge enhancement context into a preset language model for reasoning, obtaining a knowledge enhancement reasoning result set, generating multiple pieces of decision information, and combining the decision information to obtain an intelligent decision output object. In the invention, aiming at the problem that the accuracy of the intelligent decision result output by the existing intelligent question-answering system is low, the knowledge graph can be retrieved through the candidate reasoning result set obtained by calculation, then the preset language model is input for reasoning, and finally multiple pieces of decision information are combined to obtain the intelligent decision output object, so that the accuracy of the intelligent decision result output by the intelligent question-answering system is improved. And the accuracy of the intelligent decision-making result is improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an entropy-guided intelligent decision-making method, apparatus, device, and medium. Background Technology

[0002] Currently, in the financial and healthcare sectors, intelligent question-answering systems are widely used for tasks such as policy term consultation, product portfolio recommendation, risk level assessment, and claims rule determination. Some existing solutions have initially achieved multi-turn dialogue and scene recognition by linking pre-trained language models with domain knowledge bases.

[0003] However, when dealing with complex, fuzzy reasoning tasks involving multivariate relationships, existing technologies typically employ preset fixed reasoning depths or static reasoning strategies, which cannot adaptively and dynamically adjust the reasoning depth and reasoning path, making it difficult to effectively guarantee the accuracy of intelligent decision-making results. Summary of the Invention

[0004] This invention provides an entropy-guided intelligent decision-making method, apparatus, device, and medium to solve the technical problem of low accuracy of intelligent decision-making results output by existing intelligent question-answering systems.

[0005] Firstly, an entropy-guided intelligent decision-making method is provided, including: The raw request data input by the user is preprocessed to organize the task type and context, and obtain the user request object; The user request object is semantically recognized and encoded to obtain a semantic representation object; Based on the semantic representation object, a preset language model is driven to reason and calculate information entropy, and entropy monitoring data objects and reasoning control parameters are constructed respectively. Multi-step reasoning is performed based on the entropy monitoring data object and reasoning control parameters, and the outputs of each reasoning path are fused to obtain a set of candidate reasoning results; The knowledge graph is retrieved from the candidate reasoning result set and a knowledge-enhanced context is generated. The knowledge-enhanced context is then input into the preset language model for reasoning to obtain a knowledge-enhanced reasoning result set. Multiple decision information is generated based on the knowledge-enhanced reasoning result set, and the multiple decision information is combined to obtain an intelligent decision output object.

[0006] Secondly, an entropy-guided intelligent decision-making device is provided, comprising: The data acquisition module is used to preprocess the raw request data input by the user in order to organize the task type and context and obtain the user request object; The semantic recognition module is used to perform semantic recognition on the user request object and encode it to obtain a semantic representation object; The data reasoning module is used to drive a preset language model to reason and calculate information entropy based on the semantic representation object, and to construct entropy monitoring data objects and reasoning control parameters respectively. The data fusion module is used to perform multi-step inference based on the entropy monitoring data object and inference control parameters, and to fuse the outputs of each inference path to obtain a set of candidate inference results. The knowledge enhancement module is used to retrieve a knowledge graph through the candidate reasoning result set and generate a knowledge enhancement context, and input the knowledge enhancement context into the preset language model for reasoning to obtain a knowledge enhancement reasoning result set; The decision output module is used to generate decision information based on the knowledge-enhanced reasoning result set, and combine the decision information to obtain an intelligent decision output object.

[0007] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the entropy-guided intelligent decision-making method described above.

[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the aforementioned entropy-guided intelligent decision-making method.

[0009] In the aforementioned scheme implemented by the entropy-guided intelligent decision-making method, device, equipment, and medium, the original request data input by the user can be preprocessed to organize the task type and context, obtaining the user request object; the user request object is semantically recognized and encoded to obtain a semantic representation object; based on the semantic representation object, a preset language model is driven to reason and calculate information entropy, constructing entropy monitoring data objects and reasoning control parameters respectively; multi-step reasoning is performed based on the entropy monitoring data objects and reasoning control parameters, and the outputs of each reasoning path are fused to obtain a candidate reasoning result set; the candidate reasoning result set is used to retrieve a knowledge graph and generate a knowledge-enhanced context, which is then input into the preset language model for reasoning to obtain a knowledge-enhanced reasoning result set; multiple decision information is generated based on the knowledge-enhanced reasoning result set, and the multiple decision information is combined to obtain an intelligent decision output object. In this invention, addressing the problem of low accuracy in intelligent decision-making results output by existing intelligent question-answering systems, the calculated candidate reasoning result set can be used to retrieve a knowledge graph, then input into a preset language model for reasoning, and finally, multiple decision information can be combined to obtain an intelligent decision output object, thus improving the accuracy of intelligent decision-making results. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating an entropy-guided intelligent decision-making method according to an embodiment of the present invention. Figure 2 yes Figure 1 A detailed implementation flow diagram of step S30 Figure 1 ; Figure 3 yes Figure 1 A detailed implementation flow diagram of step S30 Figure 2 ; Figure 4 yes Figure 1 A detailed implementation flow diagram of step S30 Figure 3 ; Figure 5 yes Figure 1 A detailed implementation flow diagram of step S40 Figure 4 ; Figure 6 yes Figure 1 Another specific implementation flow diagram of step S50 Figure 5 ; Figure 7 yes Figure 1 A detailed implementation flow diagram of step S60 Figure 6 ; Figure 8 This is a schematic diagram of an entropy-guided intelligent decision-making device according to an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 10 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Please see Figure 1 As shown, Figure 1 A flowchart illustrating an entropy-guided intelligent decision-making method provided in an embodiment of the present invention includes the following steps: S10: Preprocess the raw request data input by the user to organize the task type and context, and obtain the user request object.

[0014] First, the system receives raw request data input by the user through the front-end interactive interface. This raw request data may include natural language consultation content, historical conversation fragments, and supplementary parameters related to the current business scenario. The raw request data is then preprocessed, including standardizing the data format and filtering out noise. Furthermore, the system organizes and categorizes the request according to the task type, target scenario, time conditions, and basic context information based on the business configuration, resulting in a user request object for subsequent modeling processing.

[0015] S20: Perform semantic recognition and encoding on the user request object to obtain a semantic representation object.

[0016] The system performs intent recognition and semantic analysis on the text content of user request objects, extracting elements such as business intent, key entities, and constraints. Based on this, it calls a pre-defined semantic encoding model to map the processed semantic information into a vectorized semantic representation object. Through this semantic representation object, the user request is converted from natural language form into an internal representation that can be uniformly invoked for subsequent reasoning and decision-making.

[0017] S30: Based on the semantic representation object, drive the preset language model to reason and calculate information entropy, and construct entropy monitoring data objects and reasoning control parameters respectively.

[0018] During the stepwise reasoning of the language model, probability information of candidate outputs at each step is collected. Based on this probability information, metrics such as information entropy related to the current reasoning process are calculated, thereby constructing an entropy monitoring data object containing uncertainty measurement data. Simultaneously, based on the uncertainty reflected in the entropy monitoring data object, reasoning control parameters are generated to control the reasoning depth, reasoning path, and reasoning mode, providing a control basis for subsequent multi-step reasoning stages.

[0019] Combination Figure 2 As shown, step S30 specifically includes the following steps: S3011: Input the semantic representation object into the preset language model for step-by-step reasoning to collect the probability distribution of each candidate output and obtain a probability distribution data set.

[0020] The semantic representation objects obtained in the previous stage are input into a pre-defined language model, and intermediate results are generated through stepwise reasoning. An autoregressive generation mechanism can be used, whereby intermediate results are generated at each reasoning step. The pre-defined language model predicts the possible candidate symbols or word sets for the next output based on the current context state and outputs the corresponding probability distribution. After each forward inference step, the output probability distribution data for that step is collected and stored, accumulating sequentially to form a probability distribution data set covering all inference steps. In financial insurance scenarios, the probability distribution can correspond to the generation probability of different conclusions or explanatory statements such as claim rejection, partial claim, full claim, and supplementary investigation; in healthcare scenarios, the probability distribution can correspond to the generation probability of different medical recommendations such as conservative treatment, enhanced follow-up, adjustment of drug dosage, and referral to a specialist. By recording these progressively output probability distributions, basic data is provided for subsequent uncertainty measurement.

[0021] S3012: The information entropy is obtained by summing the product of the probability value and its logarithm in the probability distribution dataset and taking the opposite number.

[0022] For the Output probability distribution obtained by step inference The information entropy of this step is calculated by summing the product of the probability value and its logarithm and taking the opposite number. Simultaneously, a variance entropy is constructed from the entropy sequence in the continuous inference steps. Its mathematical definition can be expressed as: in, It represents the input conditions and context information of the current reasoning process, which typically includes the question text entered by the user in a financial insurance or medical scenario, historical dialogue records, and structured business parameters; Indicates the first The first possible output of the language model in step-by-step inference One candidate word element or symbol; Indicates that given input Given the current context, the language model in the first... Step output The conditional probability; Indicates the first The information entropy of a step in inference is used to measure the magnitude of uncertainty in the model during that step. A larger value indicates that the candidate outputs are relatively dispersed and the model is unable to form a clear judgment. A smaller value indicates that the model is more confident in the current output; Indicates from the first Step to the first Steps in succession The information entropy sequence of each inference step is used to describe the trajectory of uncertainty changes within the most recent inference window; The variance of the above entropy sequence, i.e., variance entropy, is used to measure the recent... The degree of fluctuation in entropy value during each step reflects the stability of the model in the continuous inference process; The preset window length parameter can be configured according to the inference length requirements of different business scenarios. Through the above calculation, it is possible to identify highly volatile states in financial insurance scenarios, such as "the conclusion of the same claim case fluctuates repeatedly in multiple inference steps," and unstable inference processes in medical scenarios, such as "a certain treatment strategy fails to converge for a long time."

[0023] S3013: Statistically analyze the step entropy values ​​in the information entropy to calculate the variance, and obtain the variance entropy that characterizes the fluctuation of the entropy value.

[0024] Using a sliding window or cumulative segmentation method, at each time point Select the most recent Entropy value per step The corresponding standard deviation is obtained by using standard deviation calculation. .when When it is at a low level, it indicates that the uncertainty of the model is relatively stable within that time window, meaning that the model maintains a consistent attitude towards the current decision direction regardless of whether it is continuously high or low entropy; when A significant increase indicates substantial fluctuations in entropy between adjacent inference steps, with the model frequently switching between different candidate conclusions. This phenomenon is particularly common in complex financial clause interpretations, boundary-based claims cases, or handling difficult diseases with multiple treatment pathways. This embodiment focuses on simultaneously... absolute size and The fluctuating characteristics enable the system to identify not only overall uncertain situations but also unstable situations, providing a more nuanced measurement basis for subsequent reasoning and adjustment.

[0025] S3014: Combine and format the information entropy and variance entropy to construct an entropy monitoring data object and map it to obtain inference control parameters.

[0026] Will and Entropy feature vectors are constructed and normalized, smoothed, or segmented as needed to obtain entropy monitoring data objects that combine numerical features and state labels. These entropy monitoring data objects can record information such as the time sequence number, business session identifier, entropy value range, and fluctuation level for each step of reasoning. This enables the differentiation between "highly uncertain but trend-stable long-term investment risk assessment tasks" and "short-term claims dispute interpretation tasks with drastically fluctuating conclusions" in financial insurance scenarios, and between "complex cases with consistently high uncertainty regarding multiple treatment options" and "general cases with transient instability at certain key steps" in medical scenarios. After constructing the entropy monitoring data object, it is then mapped according to preset mapping rules or models. and The mapping is used to several inference control parameters, such as whether to extend the current inference depth, whether to trigger supplementary information queries, and whether to suggest calling a higher-precision medical knowledge base or actuarial model, so as to adaptively adjust the inference mode, sampling strategy and external knowledge call strategy of the language model in subsequent steps, thereby achieving fine-grained control of the inference process in high-risk business scenarios such as finance and healthcare.

[0027] Combination Figure 3 As shown, step S30 further includes the following steps: S3021: Read the information entropy and variance entropy from the entropy monitoring data object respectively, and update the entropy monitoring index set data.

[0028] The information entropy and variance entropy corresponding to each inference step are read from the entropy monitoring data object constructed in the previous steps. Information entropy reflects the degree of uncertainty in the output probability distribution of the current inference step, while variance entropy characterizes the fluctuation of entropy values ​​within a certain time window. After reading the above data, the information entropy sequence and variance entropy sequence of each time step in this round of the session are uniformly organized and updated into an entropy monitoring indicator set data. This entropy monitoring indicator set data can be indexed by timestamp or inference step number, and includes information such as scene identifier and task type. This allows for differentiation between different tasks such as policy consultation, claims assessment, and risk pricing in financial scenarios, and between different application scenarios such as initial consultation, follow-up visits, and medication adjustments in medical scenarios, thereby providing a complete state basis for subsequent strategy judgments.

[0029] S3022: Compare the entropy monitoring index set data with multi-level preset thresholds to obtain the entropy determination result that characterizes the magnitude and fluctuation state of entropy.

[0030] Several threshold parameters can be set for information entropy and variance entropy respectively. For example, a high uncertainty threshold can be set for information entropy. With low uncertainty threshold Set a stability threshold for variance entropy. Furthermore, a mapping relationship is established between entropy indices and control actions, which can be expressed as follows: when and Time: High uncertainty but stable, which means deepening reasoning; when and When uncertainty is low and stability is achieved, inference is terminated. when Time: Uncertainty fluctuates greatly, i.e., branch expansion (generating multi-path reasoning to explore multiple possible solutions); in, Representation language model in the first t The information entropy corresponding to each step of reasoning is used to measure the uncertainty of the output result of that step. When the value is large, it indicates that for the current input, the probability distribution of the model among multiple candidate outputs is relatively dispersed, and a clear judgment has not yet been formed; Indicates the first t The variance entropy, calculated within a preset window length and with the step as the endpoint, is used to describe the fluctuation of information entropy across multiple recent inference steps. A larger value indicates that the model is in a "wavering" state during continuous reasoning; The high uncertainty threshold of information entropy is usually obtained by statistical analysis of samples that require further reasoning in historical financial or medical tasks, and is used to distinguish between "significantly high uncertainty" and "medium or low uncertainty" states. This is the low uncertainty threshold for information entropy, used to identify situations where the model is quite confident about the current conclusion, and generally satisfies... > Relationship; The stability threshold of variance entropy, when < When the entropy value changes little recently and the reasoning process is generally stable, > This indicates that the entropy value has fluctuated significantly recently and the direction of reasoning is still unstable; Indicates the first tThe control action type obtained based on the entropy determination result during step-by-step reasoning includes deepening reasoning, terminating reasoning, and branch expansion, which correspond to continuing to deepen reasoning, prematurely ending reasoning, and generating multiple reasoning paths for parallel exploration, respectively. Through the above comparison and mapping, cases such as "significant uncertainty about claim liability but generally consistent judgment direction" or "frequent switching between multiple conclusions regarding the interpretation of the same policy" can be identified in financial scenarios; and cases such as "lack of certainty about the diagnostic conclusion but consistent overall judgment direction" or "repeated wavering between multiple diagnostic options in several steps of reasoning" can be identified in medical scenarios, thereby obtaining entropy determination results that can be used to guide the adjustment of reasoning depth.

[0031] S3023: Generate deepening, termination, and branch expansion control instructions based on the entropy determination result, as a set of inference depth adjustment controls.

[0032] When the entropy determination result shows that the current step satisfies > and < When the system determines that the model has high uncertainty about the conclusion but relatively stable reasoning direction in the current scenario, it generates a deeper reasoning control instruction, instructing subsequent steps to increase the number of reasoning steps in the same decision-making direction, allowing the introduction of more intermediate reasoning chains, or supplementary evidence. In financial claims scenarios, this can be used to require the model to further refine the applicable clauses or call more historical claims cases. In medical diagnosis scenarios, it can require the model to conduct a more granular comprehensive analysis of past examination results and guideline evidence. When the entropy determination result shows that the current step satisfies... < and < When the system determines that the model has low uncertainty about the current conclusion and has remained stable over recent inference steps, it generates a termination inference control command. This command indicates that the system can proceed to the result processing and output stage without adding more inference steps, thus avoiding further consumption of computing resources when sufficient confidence is already achieved. When the entropy determination result indicates that the current step satisfies... > When the system determines that the recent inference process involves significant uncertainty and fluctuations, with multiple possible conclusions yet to be fully explored, it generates a branch extension control instruction. This instruction instructs the replication of multiple inference branches from the current intermediate state, each starting from different assumptions or constraints, thus generating multiple parallel inference paths. In the financial field, such branches could correspond to premium pricing paths under different risk preference assumptions, or claims conclusion paths under different liability allocation assumptions. In the medical field, they could correspond to examination combinations and treatment recommendation paths under different initial diagnostic assumptions. All of these control instructions are uniformly managed and encoded after generation, resulting in a set of inference depth adjustment controls that can be directly invoked in subsequent steps.

[0033] S3024: Adjust the inference mode of the preset language model according to the inference depth adjustment control set to update the inference control parameters.

[0034] When the set contains an instruction to deepen inference, the maximum number of inference steps in the current session can be increased accordingly, the sampling temperature can be appropriately reduced, or the beam search width can be adjusted, allowing the model to perform more detailed and conservative multi-step inferences in the same decision direction. When the set contains an instruction to terminate inference, the system will stop generating new intermediate inference content after the current step and instead switch to the result integration and structured output process. This allows for the rapid generation of interpretable claims conclusions or risk assessment reports in financial scenarios and clear treatment suggestions or follow-up plans in medical scenarios. When the set contains a branch expansion control instruction, the system will copy multiple language model context instances based on the current state and assign different sampling strategies, constraints, or knowledge retrieval focuses to each instance, allowing them to independently carry out inference on parallel paths. The multi-path outputs are then handed over to the subsequent data fusion module for weighted integration or expert rule filtering. Through the above adjustments to the inference mode, the inference control parameters can be updated in real time according to the entropy monitoring status in each iteration, enabling the system to achieve fine-grained adaptive control of inference depth and inference paths in high-risk fields such as finance and healthcare, effectively balancing inference accuracy, computational cost, and interpretive transparency.

[0035] Combination Figure 4 As shown, step S30 further includes the following steps: S3031: Read the information entropy and variance entropy respectively, and combine the information entropy and variance entropy to obtain the entropy feature vector.

[0036] After completing the construction of the aforementioned entropy monitoring data object, for the current [number] [item]... t Each inference step reads the corresponding information entropy and variance entropy from the entropy monitoring data object, and then... With variance entropy By splicing or stacking the data in a preset order, an entropy feature vector is obtained, which characterizes the uncertainty and volatility of the current reasoning. In the financial insurance scenario, this entropy feature vector can be used to describe the degree of hesitation in the current step regarding the conclusion, as well as the degree of oscillation between different payout conclusions in recent multi-step reasoning, when performing reasoning tasks such as claims liability allocation, premium adjustment, or long-term insurance configuration. In the healthcare scenario, this entropy feature vector can be used to characterize the system's grasp and stability of a certain treatment path when performing auxiliary diagnosis of difficult diseases, medication risk assessment, or follow-up plan selection, providing a unified numerical feature for adjusting subsequent reasoning actions.

[0037] S3032: Perform linear operations using the entropy feature vector and the preset weight matrix, and calculate the action probability distribution using a normalized exponential function.

[0038] Using the entropy feature vector as input to the control strategy network, a linear operation is performed on it using a preset weight matrix, and the probability distribution for different control actions is calculated using a normalized exponential function, which can be specifically expressed as: in, Indicates the first t The information entropy calculated from the probability distribution output by the language model in each inference step is used to measure the uncertainty of the inference result in the current step. Indicates the first t The variance entropy, obtained by statistically analyzing a set of consecutive step entropy values ​​within a preset time window, is used to measure the degree of uncertainty fluctuation in the most recent inference steps. Represented by information entropy and variance entropy The column vector formed and transposed in a fixed order is the entropy eigenvector; W This represents a preset weight matrix whose dimension matches the entropy feature vector and the number of candidate control actions. It is used to map the entropy feature vector to the action score space. This weight matrix can be obtained through offline training on a large number of financial claims samples or medical treatment samples. This represents the normalization exponential function, used to exponentially transform the scores of each candidate action and normalize them by the sum, so that the output becomes a valid probability distribution. Indicates the first t Each inference step offers selectable control action types, such as continuing the current depth of inference, expanding multiple inference branches, or terminating the current inference. This indicates choosing an action given the current entropy feature vector. The probability of [the outcome]. Through the above linear mapping and normalization process, the system can automatically form the preference strength for control actions such as continuing, expanding, or stopping based on different uncertainty states.

[0039] S3033: Random sampling is performed based on the action probability distribution to obtain adaptive control action markers.

[0040] Random sampling is performed based on the action probability distribution to obtain the adaptive control action label corresponding to the current inference step. Unlike simple threshold logic, the random sampling mechanism allows the system to retain a certain degree of exploratory nature while ensuring the overall strategy direction: when the probability of a certain action is significantly higher than that of other actions, that action is more likely to be selected during the sampling process, thus reflecting the utilization of mainstream decision-making patterns; when the probabilities of multiple actions are relatively close, there is still a chance to sample actions that are not the highest probability, thereby exploring claims paths under different liability assumptions in financial scenarios, and exploring different diagnostic assumptions or treatment paths in medical scenarios, avoiding the omission of potentially important situations due to premature convergence to a single conclusion. Finally, the sampled action label is encoded as the adaptive control action label for the current step, used to drive the adjustment of subsequent inference patterns.

[0041] S3034: Select inference control instructions from the preset instruction set according to the adaptive control action marker, and update the inference mode of the preset language model in combination with the inference control parameters.

[0042] When the action marker indicates that the current inference should continue, the maximum allowed number of inference steps can be increased, the sampling temperature can be appropriately reduced, or the bundle search width can be increased without changing the number of branches. This allows the language model to perform deeper and more refined multi-step deductions in the existing decision direction, so as to fully explore implicit information in financial claims disputes or complex disease analysis. When the action marker indicates that the branch should expand, multiple language model context instances will be copied based on the current intermediate state, and different conditional constraints or retrieval focuses will be configured for different instances, thereby generating multiple inference branches in parallel. For example, in the financial field, different risk preferences, different regulatory interpretations, or different historical case matching scenarios will be simulated respectively, and in the medical field, different preliminary diagnostic hypotheses or different combinations of medication strategies will be simulated respectively, so as to enrich the alternative decision space. When the action marker indicates that the inference should terminate, no new intermediate inference steps will be added after the current step, and the subsequent result processing and structured output process will be triggered, so as to converge in a timely manner when the model is highly confident in the conclusion and avoid unnecessary computational overhead. Through the above methods, the inference control instructions and inference control parameters work together to dynamically adjust the inference depth, number of branches, sampling strategy, and external knowledge retrieval strategy of the language model. This enables the present invention to achieve adaptive control of the inference process in application fields such as finance and medicine, which require a balance between accuracy, robustness, and resource efficiency.

[0043] S40: Perform multi-step reasoning based on the entropy monitoring data object and reasoning control parameters, and fuse the outputs of each reasoning path to obtain a set of candidate reasoning results.

[0044] During the stepwise generation of the language model, the number of inference steps, the expansion method, and the strategy for retaining intermediate results are dynamically adjusted according to the inference control parameters. For cases requiring extended inference, further inference expansion is performed, and for cases requiring focus, irrelevant branches are reduced. For different inference paths formed during the inference process, the system organizes and merges the outputs of each path, combining semantically representative and consistent conclusions from multiple paths to obtain a set of candidate inference results, which are then used for subsequent joint inference with external knowledge.

[0045] Combination Figure 5 As shown, step S40 specifically includes the following steps: S401: Based on the entropy monitoring data object and the inference control parameters, when the variance entropy in the entropy monitoring data object is high, the current inference state is copied into multiple paths and inferred separately to obtain a multi-path inference result set.

[0046] The current inference state is determined based on entropy monitoring data objects and inference control parameters. When a high level of variance entropy is detected within the current time window, it indicates that the information entropy fluctuates significantly in recent inference steps, the model frequently oscillates between different conclusions, and there is a clear multi-solution space. At this point, instead of continuing along a single path, the current inference hidden state, the generated intermediate conclusions, and the retrieved external knowledge are copied into multiple copies to form multiple initial inference paths. Each path is denoted as […]. Different assumptions or sampling strategies are introduced along each path to drive the pre-defined language model to reason independently along its respective path, resulting in a set of multi-path reasoning results. For example, in the financial insurance scenario, different paths can assume different combinations of accident liability division, interpretation of past medical history, applicable clauses, or interpretation of regulatory rules; in the medical scenario, different paths can correspond to different diagnostic assumptions such as infectious diseases and autoimmune diseases, and generate corresponding examination suggestions and treatment plans along their respective paths.

[0047] S402: The information entropy and variance entropy of each path in the multi-path inference result set are linearly combined according to preset coefficients and calculated by an exponential function to obtain the corresponding path confidence coefficient set.

[0048] The confidence coefficient of each path in the multi-path inference result set is calculated independently. To do this, the information entropy sequence and variance entropy sequence of each step in the inference process of each path are re-statistically analyzed within that path, and the comprehensive information entropy of that path is obtained. and comprehensive variance entropy subscript i Used to identify the first i There are several reasoning paths. A pre-defined combination of linear combinations and exponential functions is used to... and Mapped to path confidence coefficient Based on this, the conclusions output by each path are weighted and fused, and the mathematical expression can be expressed as: in, Indicates the first i The comprehensive information entropy of a reasoning path over several reasoning steps is used to characterize the uncertainty of the conclusion corresponding to the path. The larger the value, the more hesitation and disagreement there are in the internal reasoning process of the path. Indicates the first i The statistical results of the comprehensive variance entropy of the reasoning path within a preset time window are used to measure the degree of fluctuation of uncertainty within the path. The larger the value, the greater the change in the path's attitude towards the conclusion at different times and the poorer the stability. and The preset non-negative weight coefficients are used to adjust the relative importance of information entropy and variance entropy in path confidence calculation. In financial scenarios, they can be learned offline based on historical claims decisions or risk rating samples, and in medical scenarios, they can be configured in combination with clinical guidelines and expert experience. is the base of the natural logarithm; For the first i The confidence coefficient of a reasoning path; the larger the value, the more reliable the path is in terms of uncertainty and stability. For the first i The conclusion of the reasoning path can be expressed as financial quantitative indicators such as the probability of claim approval, risk level score, and rate adjustment range, or as medical quantitative indicators such as the confidence level of different diagnostic categories, the priority of treatment plans, or the prognostic risk score. The weighted and fused comprehensive reasoning result across all paths is used as the core candidate conclusion for the final output of the system; (n) represents the number of reasoning paths generated at the moment. This represents the sum of the confidence coefficients for all paths, used to normalize the confidence of each path. For the first i The weighting factors of each path reflect its relative contribution to the final integrated result. Through this method, paths with lower information entropy and smaller variance entropy will receive higher weights, while paths with high uncertainty or drastic fluctuations will have relatively lower weights, thus effectively suppressing the influence of noisy paths on the final decision.

[0049] S403: Normalize the set of confidence coefficients of the paths to obtain the weight factors of each path, and use the weight factors to perform a weighted summation of each path to obtain a weighted fusion inference set.

[0050] The confidence coefficients of all paths are summed. The confidence coefficient of each path is then divided by this sum to obtain the corresponding weighting factor, ensuring that the sum of all weights is equal. These weighting factors are then used to perform a weighted summation or weighted aggregation of the inference conclusions of each path, resulting in a weighted fusion inference set. For tasks with continuous numerical outputs, such as risk scoring, payout probability estimation, or capital occupancy level calculation in the financial field, the above weighted summation can be directly used to obtain a single comprehensive numerical result. For tasks with multi-category labels or structured suggestions, such as a multi-diagnostic candidate set or multi-treatment suggestions in the medical field, the categories or solutions can be encoded into a unified internal representation at the path level, and then aggregated based on the weighting factors to obtain a ranked candidate list and corresponding confidence distribution.

[0051] S404: The weighted fusion reasoning set is used as the target candidate conclusion, and the multi-path reasoning result set is used as the alternative conclusion, and the candidate reasoning result set is obtained by summarizing them.

[0052] On the one hand, the comprehensive results in the weighted fusion reasoning set can be directly provided to subsequent knowledge enhancement modules or decision output modules to generate user-oriented final responses or structured decision suggestions. On the other hand, the multi-path reasoning result set retains the independent conclusions of each path and their corresponding key information, providing traceable explanatory basis for scenarios such as financial auditing, claims review, or medical consultation. For example, in insurance claims scenarios, the system can show reviewers the weighted fusion results corresponding to the main conclusion and several alternative path conclusions with higher weights to explain why a certain liability allocation or compensation amount was chosen. In medical scenarios, the system can show clinicians the comprehensive diagnostic conclusion and several representative path diagnostic schemes and their weights, facilitating doctors to make a final judgment based on the patient's actual situation. Through the above-mentioned multi-path exploration and weighted fusion mechanism, this embodiment can fully utilize the information of each feasible reasoning path in tasks with high uncertainty and multiple solution spaces, reducing the risk of erroneous decisions due to local path deviations while maintaining the overall robustness of reasoning.

[0053] S50: Retrieve the knowledge graph through the candidate reasoning result set and generate a knowledge-enhanced context, input the knowledge-enhanced context into the preset language model for reasoning, and obtain a knowledge-enhanced reasoning result set.

[0054] Semantic parsing is performed on the candidate inference result set to extract semantic nodes or key semantic fragments that represent the current decision-making scenario. Then, using these semantic elements as retrieval clues, related entities, attributes, and relationship information are searched in the knowledge graph of the financial and insurance domain to obtain structured knowledge content relevant to the current task. This structured knowledge is encoded and organized to generate a knowledge-enhanced context representing external domain knowledge. This knowledge-enhanced context, along with the current inference state, is input into a preset language model, and inference is executed again to obtain a knowledge-enhanced inference result set that incorporates external knowledge.

[0055] Combination Figure 6 As shown, step S50 specifically includes the following steps: S501: Perform semantic parsing on the candidate reasoning result set to extract semantic nodes and obtain the retrieval target semantic node set.

[0056] During the semantic parsing process, key semantic nodes that reflect the current decision focus are extracted from the candidate reasoning results. For example, in the financial insurance scenario, semantic units such as critical illness insurance product names, waiting period clauses, descriptions of past medical history, and reimbursement ratios are extracted. In the medical scenario, semantic units such as suspected disease names, main symptoms, laboratory test indicators, drug names and dosages are extracted. These semantic units are then uniformly mapped to an internal set of retrieval target semantic nodes to drive the subsequent knowledge graph retrieval process.

[0057] S502: Calculate the similarity between the set of semantic nodes of the search target and the preset knowledge graph, and filter out relation triples.

[0058] Similarity matching is performed between the retrieved target semantic node set and the entity set in the pre-defined knowledge graph to filter out relation triples that are highly relevant to the current semantic node. Therefore, in the reasoning... t The first step is to represent the current semantic node as a vector. Then, search for all entity vectors in the insurance knowledge graph or the medical knowledge graph. And calculate the similarity between the two, when the similarity is greater than a threshold. If the entity is considered relevant to the current semantic node, then the triples involving that entity are extracted. Add to the current search results set to obtain the result related to the first... t The set of relational triples related to step-by-step reasoning The above process can be formally represented as: in, Indicates the first tThe semantic vector obtained from the candidate reasoning results or the current semantic node encoding during step reasoning is used to represent the semantic requirements that need to be retrieved at the current step. Represents the first in a knowledge graph i Each entity vector can be a specific insurance product, clause number, or risk label in an insurance scenario, or a disease, symptom, drug, or examination item in a medical scenario. Representation and entity The types of relationships involved, such as containment, constraint, dependency, triggering, etc. Representation and entity Through relationships The other entity connected to it, such as insurance products, restrictions, upper age limit for insured, diseases, recommended examinations, specific imaging projects, etc.; This represents a function used to measure the similarity between two vectors; it can be cosine similarity or other distance metrics. This is a similarity threshold used to filter out entities that have a low relevance to the current semantic node; In the first t The set of relation triples retrieved from the knowledge graph during step-by-step reasoning; This indicates that the pre-defined language model, without introducing external knowledge, is in its first stage. t The hidden state vector of the step is used to characterize the internal semantic representation formed by the model in that step; Indicates according to The knowledge context embedding vector obtained by encoding each relation triple is used to represent the structured domain knowledge that should be considered in the current step of reasoning; This indicates that the model's hidden state vectors With knowledge context embedding vector The new vector obtained by concatenating vectors according to their dimensions; The knowledge-enhanced hidden state vector obtained after introducing external knowledge will be used as a new input in subsequent inference.

[0059] S503: Encode the relation triples to generate a knowledge context embedding vector that represents external structured knowledge.

[0060] For the obtained set of relation triples A unified encoding process is performed to generate knowledge context embedding vectors that represent external structured knowledge. During encoding, the head entities in the triples can be encoded separately. ,relation Tail-end entity Embedding mapping is performed, and the data is aggregated into one or more high-dimensional vectors through attention weighting, concatenation, or pooling. This compresses rule information such as product details, terms, constraints, diseases, medications, contraindications, symptoms, examinations, and recommendation levels into a continuous vector space. For financial insurance tasks, The document highlights key information such as the boundaries of coverage, triggering conditions for claims, premium calculation logic, and regulatory compliance constraints; for medical tasks, The key information includes the correspondence between diseases and symptoms, recommended examination pathways, indications and contraindications for medication, and guideline-level evidence, so that these structured constraints can be explicitly perceived in subsequent reasoning processes.

[0061] S504: Concatenate the knowledge context embedding vector with the current inference hidden state vector, and input it into the preset language model for inference to obtain a set of knowledge-enhanced inference results.

[0062] Embedding knowledge context into vectors With the current inference hidden state vector The vectors are concatenated to form a knowledge-enhanced hidden state vector. and will As a new latent state input to the subsequent reasoning unit of the pre-defined language model, this mechanism simultaneously considers the model's native semantic reasoning capabilities and external knowledge constraints at each generation step. Based on this mechanism, in financial insurance scenarios, when the model attempts to generate a certain compensation conclusion or premium rate suggestion, it automatically refers to relevant triples in the knowledge graph regarding product liability scope, deductible conditions, waiting period clauses, and regulatory policies, avoiding the generation of conclusions that conflict with the terms or regulatory requirements. In medical scenarios, when the model attempts to provide diagnostic or treatment suggestions, it automatically refers to the structured relationships between diseases, symptoms, examinations, and medications in the knowledge graph, avoiding the proposal of solutions that contradict clinical guidelines or pose significant safety risks.

[0063] S60: Generate multiple decision information based on the knowledge-enhanced reasoning result set, and combine the multiple decision information to obtain an intelligent decision output object.

[0064] The knowledge-enhanced reasoning results are organized, and the judgments, reasoning, risk warnings, and feasible suggestions related to business objectives are structured and summarized to obtain multiple decision information objects. Based on this, according to a preset decision output format and business specifications, these multiple decision information objects are combined and encapsulated to construct intelligent decision output objects, including text responses, structured tags, and strategy instructions. Finally, these intelligent decision output objects are returned to the front end or business system to support intelligent question answering, assisted decision-making, or process automation in financial and healthcare scenarios.

[0065] Combination Figure 7 As shown, step S60 specifically includes the following steps: S601: Organize the reasoning data of the knowledge-enhanced reasoning result set to generate a reasoning state data set for decision calculation.

[0066] The aforementioned knowledge-enhanced reasoning results set is uniformly organized, and the hidden state vectors, information entropy, variance entropy, selected control actions, and corresponding candidate conclusions at each reasoning step are organized in chronological order into a reasoning state data set, resulting in a state sequence covering the entire conversation process. For financial insurance scenarios, this reasoning state data set can include intermediate judgments and their confidence levels regarding claims conclusions, rate adjustments, and risk levels at each step; for medical scenarios, it can include step-by-step reasoning traces of different diagnostic hypotheses, examination recommendations, and medication regimens, providing a complete data foundation for subsequent confidence assessment and reinforcement learning.

[0067] S602: Take the opposite of the information entropy of each step in the inference state data set and input it into the exponential function to calculate the confidence sequence.

[0068] For the t The system denotes the information entropy of this step as follows: The confidence level is then calculated using an exponential function. Simultaneously, the system maintains a reasoning experience cache pool to record the entropy state, action selection, and result reward for each step, which can be mathematically represented as: in, Indicates the first t The information entropy corresponding to each step of inference is used to measure the uncertainty of the output probability distribution of that step; the larger the value, the more uncertain the model is. Indicates the relationship with the first t The overall confidence level (i.e., the confidence level sequence) corresponding to the step conclusion is expressed as an exponential mapping with respect to the inverse of the entropy. When smaller A value close to 1 indicates that the model is relatively confident in the current conclusion. When larger A value close to 0 reflects a high degree of uncertainty. M This represents the reasoning experience cache pool, where each element... In the first t The extended form of the empirical triplet recorded during step-by-step reasoning, where For the first t The variance entropy corresponding to each step is used to measure the degree of fluctuation of the entropy value in the most recent steps; In the first tThe control action selected by the step is based on the entropy feature vector and the control strategy network, such as continuing to deepen the reasoning, branch expansion, early termination, etc. In order to be with the first t The return value related to the overall reasoning result of the step can be defined in the financial scenario based on indicators such as whether the final claim is correct, whether the risk assessment is consistent with historical defaults, and whether the rate adjustment improves profitability and compliance. In the medical scenario, it can be defined based on indicators such as whether the doctor adopts the advice, whether the diagnosis is accurate, whether the treatment plan complies with guidelines, and the patient follow-up outcome. These values ​​are used to reflect the comprehensive effect of the step.

[0069] S603: Perform continuous step size detection on the confidence sequence, and generate target decision information when the confidence exceeds the threshold and the decisions are consistent within a preset number of steps.

[0070] Preset reliability threshold and a continuous step parameter m When continuously within a certain time period m Step to satisfy And this m If the decision conclusions of each step are semantically consistent (e.g., multiple consecutive steps all point to claim approval and a payout ratio within a fixed range, a primary diagnosis of a specific disease, and no further fundamental changes in the treatment strategy), then the current reasoning is considered to have reached a sufficiently stable and confident state. Corresponding target decision information is then generated, and further reasoning ceases, thus avoiding the consumption of additional computing power and time in the high-confidence stage. In the financial field, this means a stable judgment has been formed regarding a claim case, credit limit, or premium adjustment; in the medical field, it means that a clinically relevant recommendation has been formed for a diagnosis and corresponding treatment strategy.

[0071] S604: Use the target decision information to construct the entropy action result and write it into the reasoning experience cache pool.

[0072] Each piece of target decision information not only records the final decision conclusion itself, but also links it to the key reasoning steps in its formation process. Data enables the experience cache pool to simultaneously reflect the complete technical solution that, under what level of uncertainty and volatility, which control actions were chosen and what business returns were ultimately achieved. For financial applications, these entropy actions can manifest, for example, "in cases of moderate uncertainty but low entropy fluctuation, choosing to moderately deepen the reasoning and supplement the clause retrieval, ultimately obtaining a claims conclusion consistent with manual review"; for medical applications, it can manifest, "in stages of high uncertainty and large volatility, exploring multiple diagnostic paths through branch expansion, ultimately obtaining a correct diagnosis confirmed by experts on a certain path."

[0073] S605: Invoke the reinforcement learning algorithm based on the inference experience cache pool to generate multiple decision information.

[0074] Reinforcement learning methods, such as proximal policy optimization algorithms, can be used to optimize entropy states. And knowledge enhancement allows the hidden state to be used as the environmental state to control actions. As a strategy output, in exchange for reward To optimize the objective, the mapping relationship is... Iterative training allows the system to gradually develop reasoning styles adapted to different scenarios through long-term experience: in high-risk, high-sensitivity medical diagnosis or high-amount claims scenarios, it tends towards cautious reasoning, appropriately increasing multi-path exploration and knowledge retrieval; in low-risk, clearly defined standardized queries or small-amount claims scenarios, it tends towards rapid convergence and early termination to improve overall processing efficiency. The action sequences output by the trained policy network in new tasks yield multiple sets of candidate decision information, enriching the final output.

[0075] S606: Combine and encapsulate multiple decision information items to obtain an intelligent decision output object.

[0076] The intelligent decision-making output can include structured main decision conclusions (such as whether to pay compensation, the range of compensation amount, risk level or main diagnostic results), auxiliary decision-making suggestions (such as supplementary underwriting suggestions, risk warnings, suggestions for follow-up inspections and visits, etc.), and key explanatory information related to the decision-making process (such as the core clauses or guidelines that trigger the conclusion, important evidence chains and confidence levels, etc.).

[0077] In one application scenario, the entropy-guided intelligent decision-making method of this application is applied to the health insurance claims review scenario. A user inputs the inquiry "I had surgery last week under my health insurance policy, how much compensation can I expect?" through the claims service terminal. First, the perception module performs intent recognition and task classification on the input text, parsing the current task as a "claims calculation" type task, and organizing relevant policy information, medical records, and the user's historical interaction context into a standardized user request object. The re-inference control module initiates a multi-step inference process. In step 1, it identifies the user's corresponding surgical type and performs semantic understanding by combining it with previous medical records. At this point, the uncertainty of the model's output result, measured by entropy, is approximately 0.48. In step 2, it invokes the insurance knowledge graph and clause retrieval service to automatically associate the liability clauses and exclusion clauses in the user's policy. After knowledge-enhanced inference, the entropy value decreases to 0.22 and the fluctuation tends to stabilize. In step 3, it calculates the compensation ratio based on the conclusions of the first two steps and detects the entropy value fluctuation variance over several consecutive steps. When the entropy fluctuation variance is determined to be less than 0.05 and the conclusions remain consistent, a "terminate inference" command is triggered, ending this round of multi-step inference. Finally, a structured claims conclusion is generated, such as: "According to Clause 12, the compensation ratio is 80%, and the expected compensation amount is ×× yuan." Simultaneously, it automatically generates an inference explanation chain and a confidence report, including an entropy change curve, a key clause reference chain, and knowledge graph hit records, for claims review personnel to verify and audit.

[0078] In terms of system implementation and deployment structure in this embodiment, it can be deployed in an insurance company's data center or an enterprise's private cloud environment, adopting a modular microservice architecture. The core inference engine is customized and fine-tuned based on an open-source large language model (such as Qwen, LLaMA2, or GPT-like architecture) to adapt to the semantic features of health insurance claims corpus and terms; the entropy monitoring and strategy module runs as an independent service process, responsible for receiving the probability distribution of each step of inference output in real time, calculating information entropy and variance entropy, and generating corresponding control actions; the knowledge graph service provides SPARQL query interface and embedded retrieval capabilities, used to quickly retrieve structured knowledge such as policy terms, disease codes, and medical expense catalogs during the claims inference process; the log and interpretability module is used to generate decision chains and knowledge reference paths, obtaining a complete and traceable record of each claims review inference; the training and optimization module periodically samples from the inference log, updates the entropy threshold, strategy network parameters, and related hyperparameters, and realizes continuous optimization of the inference control strategy. Data interaction between microservices is achieved through RESTful or gRPC interfaces, supplemented by message queues (such as Kafka) for asynchronous communication, thereby supporting high-concurrency claims review request processing while ensuring low latency, and meeting the real-time and reliability requirements of insurance institutions in the production environment.

[0079] As can be seen, the entropy-guided intelligent decision-making method provided in this application breaks through the technical bottlenecks of fixed inference depth and unquantifiable uncertainty in existing financial intelligent systems. By simultaneously introducing information entropy and variance entropy to dynamically measure and guide the model inference process, the inference depth can be adaptively adjusted according to real-time uncertainty. In complex financial decision-making tasks, this effectively avoids over-inference and premature termination, significantly reducing inference costs and improving inference accuracy. The method simultaneously generates entropy change curves and knowledge reference links during the inference process, recording and outputting the model evolution trajectory and external knowledge sources in a structured form. This enhances the interpretability and audit traceability of the decision-making process, meeting the compliance and verifiability requirements of scenarios such as insurance claims review and risk assessment. Furthermore, it addresses the challenges in high variance entropy scenarios. By automatically triggering multi-path exploration and weighted fusion of conclusions from each path, this application effectively avoids single-path reasoning from getting trapped in local optima, improving robustness and stability for complex multivariate tasks. Relying on a reasoning experience cache pool and reinforcement learning mechanism, the system can automatically optimize entropy thresholds and control strategy functions based on historical tasks during long-term business operations, gradually forming a self-evolving reasoning style that matches the characteristics of financial business. At the same time, this method maintains universality in model interface and strategy design. In addition to being applicable to insurance claims review, risk assessment, and intelligent consultation, it can also be smoothly migrated to various financial decision-making scenarios such as bank credit granting, portfolio evaluation, and credit risk prediction, providing a unified technical foundation for building a high-precision, interpretable, and sustainably optimized financial intelligent decision-making system that spans businesses and institutions.

[0080] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0081] In one embodiment, an entropy-guided intelligent decision-making device is provided, which corresponds one-to-one with the entropy-guided intelligent decision-making method described in the above embodiments. For example... Figure 8 As shown, the entropy-guided intelligent decision-making device includes: a data acquisition module 100, a semantic recognition module 200, a data reasoning module 300, a data fusion module 400, a knowledge enhancement module 500, and a decision output module 600.

[0082] Detailed descriptions of each functional module are as follows: The data acquisition module 100 is used to preprocess the raw request data input by the user in order to organize the task type and context and obtain the user request object; The semantic recognition module 200 is used to perform semantic recognition and encoding on the user request object to obtain a semantic representation object; The data reasoning module 300 is used to drive a preset language model to reason and calculate information entropy based on the semantic representation object, and to construct entropy monitoring data objects and reasoning control parameters respectively. The data fusion module 400 is used to perform multi-step reasoning based on the entropy monitoring data object and reasoning control parameters, and to fuse the outputs of each reasoning path to obtain a set of candidate reasoning results. The knowledge enhancement module 500 is used to retrieve a knowledge graph through the candidate reasoning result set and generate a knowledge enhancement context, and input the knowledge enhancement context into the preset language model reasoning to obtain a knowledge enhancement reasoning result set. The decision output module 600 is used to generate decision information based on the knowledge-enhanced reasoning result set, and combine the decision information to obtain an intelligent decision output object.

[0083] In one embodiment, the data inference module 300 is specifically used for: The semantic representation object is input into the preset language model for step-by-step reasoning to collect the probability distribution of candidate outputs at each step, thereby obtaining a probability distribution data set; The information entropy is calculated by summing the product of the probability value and its logarithm in each step of the probability distribution dataset and taking the opposite number. The step entropy values ​​in the information entropy are statistically analyzed to calculate the variance, thus obtaining the variance entropy that characterizes the fluctuation of the entropy values; The information entropy and variance entropy are combined and formatted to construct an entropy monitoring data object, which is then mapped to obtain inference control parameters.

[0084] In one embodiment, the data inference module 300 is further specifically used for: Read the information entropy and variance entropy from the entropy monitoring data object respectively, and update the entropy monitoring indicator set data. The entropy monitoring index set data is compared with multi-level preset thresholds to obtain the entropy determination result that characterizes the magnitude and fluctuation state of entropy; Based on the entropy determination result, deepen, terminate, and branch expansion control instructions are generated as a set of inference depth adjustment controls; The inference mode of the preset language model is adjusted according to the inference depth adjustment control set to update the inference control parameters.

[0085] In one embodiment, the data inference module 300 is further specifically used for: Read the information entropy and variance entropy respectively, and combine the information entropy and variance entropy to obtain the entropy feature vector; The action probability distribution is obtained by performing linear operations on the entropy feature vector and the preset weight matrix, and then calculating the result using a normalized exponential function. Random sampling is performed based on the action probability distribution to obtain adaptive control action tags; Based on the adaptive control action marker, inference control instructions are selected from the preset instruction set, and the inference mode of the preset language model is updated in conjunction with the inference control parameters.

[0086] In one embodiment, the data fusion module 400 is specifically used for: Based on the entropy monitoring data object and inference control parameters, when the variance entropy in the entropy monitoring data object is high, the current inference state is copied into multiple paths and inferred separately to obtain a multi-path inference result set. The information entropy and variance entropy of each path in the multi-path inference result set are linearly combined according to preset coefficients and calculated by an exponential function to obtain the corresponding path confidence coefficient set. The set of confidence coefficients of the paths is normalized to obtain the weight factor of each path, and the weight factor is used to perform a weighted summation of each path to obtain a weighted fusion inference set. The weighted fusion reasoning set is used as the target candidate conclusion, and the multi-path reasoning result set is used as the alternative conclusion, and the candidate reasoning result set is obtained by summarizing them.

[0087] In one embodiment, the knowledge enhancement module 500 is specifically used for: Semantic parsing is performed on the candidate reasoning result set to extract semantic nodes, thereby obtaining the retrieval target semantic node set; The similarity between the target semantic node set and the preset knowledge graph is calculated to obtain relation triples; The relation triples are encoded to generate knowledge context embedding vectors that represent external structured knowledge; The knowledge context embedding vector is concatenated with the current inference hidden state vector and input into the preset language model for inference, and then integrated to obtain a set of knowledge-enhanced inference results.

[0088] In one embodiment, the decision output module 600 is specifically used for: The knowledge-enhanced reasoning result set is processed to generate a reasoning state data set for decision-making calculation; The confidence sequence is obtained by taking the inverse of the information entropy of each step in the inference state data set and inputting it into an exponential function. The confidence sequence is continuously tested for step size. When the confidence exceeds the threshold and the decisions are consistent within a preset number of steps, target decision information is generated. The target decision information is used to construct the entropy action result and written into the reasoning experience cache pool; The reinforcement learning algorithm is invoked based on the inference experience cache pool to generate multiple decision information; The multiple decision information items are combined and encapsulated to obtain an intelligent decision output object.

[0089] For specific limitations regarding the entropy-guided intelligent decision-making device, please refer to the limitations of the entropy-guided intelligent decision-making method above, which will not be repeated here. Each module in the aforementioned entropy-guided intelligent decision-making device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0090] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side entropy-guided intelligent decision-making method.

[0091] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 10 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements client-side functions or steps of an entropy-guided intelligent decision-making method.

[0092] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed, can perform the steps provided in the above embodiments.

[0093] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0094] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0095] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0096] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method of entropy-guided intelligent decision making, characterized in that, include: The raw request data input by the user is preprocessed to organize the task type and context, and obtain the user request object; The user request object is semantically recognized and encoded to obtain a semantic representation object; Based on the semantic representation object, a preset language model is driven to reason and calculate information entropy, and entropy monitoring data objects and reasoning control parameters are constructed respectively. Multi-step reasoning is performed based on the entropy monitoring data object and reasoning control parameters, and the outputs of each reasoning path are fused to obtain a set of candidate reasoning results; The knowledge graph is retrieved from the candidate reasoning result set and a knowledge-enhanced context is generated. The knowledge-enhanced context is then input into the preset language model for reasoning to obtain a knowledge-enhanced reasoning result set. Multiple decision information is generated based on the knowledge-enhanced reasoning result set, and the multiple decision information is combined to obtain an intelligent decision output object.

2. The entropy guide based intelligent decision making method of claim 1, wherein, The process of driving a preset language model for inference and calculating information entropy based on the semantic representation object involves constructing entropy monitoring data objects and inference control parameters, including: The semantic representation object is input into the preset language model for step-by-step reasoning to collect the probability distribution of candidate outputs at each step, thereby obtaining a probability distribution data set; The information entropy is calculated by summing the product of the probability value and its logarithm in each step of the probability distribution dataset and taking the opposite number. The step entropy values ​​in the information entropy are statistically analyzed to calculate the variance, thus obtaining the variance entropy that characterizes the fluctuation of the entropy values; The information entropy and variance entropy are combined and formatted to construct an entropy monitoring data object, which is then mapped to obtain inference control parameters.

3. The entropy guide based intelligent decision making method of claim 2, wherein, The step of driving a preset language model for inference based on the semantic representation object and calculating information entropy, and constructing entropy monitoring data objects and inference control parameters respectively, also includes: Read the information entropy and variance entropy from the entropy monitoring data object respectively, and update the entropy monitoring indicator set data. The entropy monitoring index set data is compared with multi-level preset thresholds to obtain the entropy determination result that characterizes the magnitude and fluctuation state of entropy; Based on the entropy determination result, deepen, terminate, and branch expansion control instructions are generated as a set of inference depth adjustment controls; The inference mode of the preset language model is adjusted according to the inference depth adjustment control set to update the inference control parameters.

4. The entropy guide based intelligent decision making method of claim 2, wherein, The step of driving a preset language model for inference based on the semantic representation object and calculating information entropy, and constructing entropy monitoring data objects and inference control parameters respectively, also includes: Read the information entropy and variance entropy respectively, and combine the information entropy and variance entropy to obtain the entropy feature vector; The action probability distribution is obtained by performing linear operations on the entropy feature vector and the preset weight matrix, and then calculating the result using a normalized exponential function. Random sampling is performed based on the action probability distribution to obtain adaptive control action tags; Based on the adaptive control action marker, inference control instructions are selected from the preset instruction set, and the inference mode of the preset language model is updated in conjunction with the inference control parameters.

5. The entropy guide based intelligent decision making method of claim 1, wherein, The process involves multi-step inference based on the entropy monitoring data object and inference control parameters, and the fusion of the outputs of each inference path to obtain a set of candidate inference results, including: Based on the entropy monitoring data object and inference control parameters, when the variance entropy in the entropy monitoring data object is high, the current inference state is copied into multiple paths and inferred separately to obtain a multi-path inference result set. The information entropy and variance entropy of each path in the multi-path inference result set are linearly combined according to preset coefficients and calculated by an exponential function to obtain the corresponding path confidence coefficient set. The set of confidence coefficients of the paths is normalized to obtain the weight factor of each path, and the weight factor is used to perform a weighted summation of each path to obtain a weighted fusion inference set. The weighted fusion reasoning set is used as the target candidate conclusion, and the multi-path reasoning result set is used as the alternative conclusion, and the candidate reasoning result set is obtained by summarizing them.

6. The intelligent decision-making method based on entropy guidance according to claim 1, characterized in that, The process involves retrieving a knowledge graph from the candidate inference result set and generating a knowledge-enhanced context, then inputting the knowledge-enhanced context into the preset language model for inference to obtain a knowledge-enhanced inference result set, including: Semantic parsing is performed on the candidate reasoning result set to extract semantic nodes, thereby obtaining the retrieval target semantic node set; The similarity between the target semantic node set and the preset knowledge graph is calculated to obtain relation triples; The relation triples are encoded to generate knowledge context embedding vectors that represent external structured knowledge; The knowledge context embedding vector is concatenated with the current inference hidden state vector and input into the preset language model for inference, and then integrated to obtain a set of knowledge-enhanced inference results.

7. The intelligent decision-making method based on entropy guidance according to claim 1, characterized in that, The process of generating multiple decision information items based on the knowledge-enhanced reasoning result set, and combining the multiple decision information items to obtain an intelligent decision output object, includes: The knowledge-enhanced reasoning result set is processed to generate a reasoning state data set for decision-making calculation; The confidence sequence is obtained by taking the inverse of the information entropy of each step in the inference state data set and inputting it into an exponential function. The confidence sequence is continuously tested for step size. When the confidence exceeds the threshold and the decisions are consistent within a preset number of steps, target decision information is generated. The target decision information is used to construct the entropy action result and written into the reasoning experience cache pool; The reinforcement learning algorithm is invoked based on the inference experience cache pool to generate multiple decision information; The multiple decision information items are combined and encapsulated to obtain an intelligent decision output object.

8. An intelligent decision-making device based on entropy guidance, characterized in that, include: The data acquisition module is used to preprocess the raw request data input by the user in order to organize the task type and context and obtain the user request object; The semantic recognition module is used to perform semantic recognition on the user request object and encode it to obtain a semantic representation object; The data reasoning module is used to drive a preset language model to reason and calculate information entropy based on the semantic representation object, and to construct entropy monitoring data objects and reasoning control parameters respectively. The data fusion module is used to perform multi-step inference based on the entropy monitoring data object and inference control parameters, and to fuse the outputs of each inference path to obtain a set of candidate inference results. The knowledge enhancement module is used to retrieve a knowledge graph through the candidate reasoning result set and generate a knowledge enhancement context, and input the knowledge enhancement context into the preset language model for reasoning to obtain a knowledge enhancement reasoning result set; The decision output module is used to generate decision information based on the knowledge-enhanced reasoning result set, and combine the decision information to obtain an intelligent decision output object.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the entropy-guided intelligent decision-making method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the entropy-guided intelligent decision-making method as described in any one of claims 1 to 7.

Images