Intelligent reasoning method and device based on neural symbols and medium

Abstract

The embodiment of the invention discloses an intelligent reasoning method and device based on neural symbols and a medium, and relates to the technical field of artificial intelligence, the method comprises the steps that user input information corresponding to a user and a pre-constructed mixed domain knowledge base are acquired, and the mixed domain knowledge base comprises a first-order predicate logic rule base and a closed Markov logic network; analyzing the user input information, determining an input feature vector, performing neural network reasoning on the input feature vector through a pre-constructed neural reasoning module, and determining corresponding pseudo tag information; and determining a candidate rule through a pre-constructed symbol reasoning module according to the pseudo-tag information and a first-order predicate logic rule base, so as to carry out collaborative reasoning on the pseudo-tag information and the candidate rule by utilizing a closed Markov logic network, and determine corresponding reasoning knowledge information.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and in particular, to a neural-symbolic based intelligent reasoning method, device, and medium. Background Art

[0002] With the rapid development of artificial intelligence technology, general large models (such as GPT series) and vertical domain large models (such as specialized models for medical and financial fields) have made remarkable progress in terms of parameter scale and industry data mining. However, general large models lack the support of professional domain knowledge bases and front-line expert experience, and it is difficult to generate systematic and operable professional solutions; while vertical domain large models, although deeply involved in specific industries, their architectures are limited to a single domain and cannot be generalized to complex cross-industry scenarios, such as simultaneously meeting the needs of medical diagnosis and financial risk control. In existing methods, there is a heavy reliance on the black-box reasoning of pure neural networks or the rigid rules of symbolic systems, resulting in the difficulty of achieving both interpretability and generalization ability. For example, neural networks cannot explain the decision-making logic, while traditional expert systems (such as rule-based engines) are difficult to handle ambiguity and uncertainty. Therefore, in existing reasoning methods, due to the single knowledge representation and fragmented reasoning methods, there is a problem that interpretability and generalization ability are difficult to be compatible, and it cannot meet the complex scenarios of multi-industry reasoning. Summary of the Invention

[0003] One or more embodiments of this specification provide a neural-symbolic based intelligent reasoning method, device, and medium, which are used to solve the following technical problems: In existing reasoning methods, due to the single knowledge representation and fragmented reasoning methods, there is a problem that interpretability and generalization ability are difficult to be compatible, and it cannot meet the complex scenarios of multi-industry reasoning.

[0004] One or more embodiments of this specification adopt the following technical solutions:

[0005] One or more embodiments of this specification provide a neural-symbolic based intelligent reasoning method, the method includes: obtaining user input information corresponding to a user and a pre-constructed hybrid domain knowledge base, wherein the hybrid domain knowledge base includes a first-order predicate logic rule base and a closed Markov logic network; parsing the user input information to determine an input feature vector, so as to perform neural network reasoning on the input feature vector through a pre-constructed neural reasoning module to determine corresponding pseudo-label information; through a pre-constructed symbolic reasoning module, according to the pseudo-label information and the first-order predicate logic rule base, determine candidate rules, and use the closed Markov logic network to perform collaborative reasoning on the pseudo-label information and the candidate rules to determine corresponding reasoning knowledge information.

[0006] One or more embodiments of this specification provide a neural-symbolic based intelligent reasoning device, including:

[0007] At least one processor; and,

[0008] A memory communicatively connected to the at least one processor; wherein,

[0009] The memory stores instructions executable by the at least one processor, and when executed by the at least one processor, enables the at least one processor to execute the above-mentioned method.

[0010] A non-volatile computer storage medium provided by one or more embodiments of this specification, storing computer-executable instructions, where the computer-executable instructions are configured to: execute the above-mentioned method.

[0011] The above-mentioned at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects: Through the technical solution of the embodiments of this specification, the hybrid domain knowledge base includes a first-order predicate logic rule base and a closed Markov logic network, combining the precise rule representation of symbolicism with the Markov logic network based on the probabilistic graphical model, which can not only handle deterministic knowledge but also handle knowledge with fuzziness and uncertainty, making up for the deficiencies of traditional single knowledge representation methods, making knowledge representation more comprehensive and flexible, and meeting the needs of different types of knowledge representation in complex scenarios of multiple industries; first, the neural inference module performs neural network inference on the input feature vector to obtain pseudo-label information, then the symbolic inference module combines the pseudo-label information and the first-order predicate logic rule base to determine candidate rules, and finally, collaborative inference is performed through the closed Markov logic network, combining the powerful feature extraction and learning ability of the neural network and the interpretability and logical reasoning ability of the symbolic system, which can not only handle complex non-linear data but also provide an interpretable reasoning process, and to a certain extent solves the problem that it is difficult to have both interpretability and generalization ability; the neural inference module can learn potential features and patterns from a large amount of data, providing a basis for subsequent inference, the symbolic inference module performs inference based on the rule base, ensuring the logic and accuracy of the inference, and the collaborative inference of the closed Markov logic network further comprehensively considers the uncertainty of various factors and optimizes the results through probabilistic inference. The collaborative working mode of multiple modules can make full use of the advantages of different modules, improve the accuracy and generalization ability of the inference, enabling it to better adapt to complex scenarios across industries, rather than being limited to a single domain like a vertical domain large model and unable to generalize to other domains; different from the black-box inference of pure neural networks, the symbolic inference module performs inference based on a clear first-order predicate logic rule base, and can display the rules and logical relationships used during the inference process, making the inference result interpretable, and users can understand how the system draws conclusions based on the rules and input information, which is crucial for scenarios that require a reliable decision-making basis (such as medical diagnosis, financial risk control, etc.), and helps to improve users' trust in the inference results. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] To more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments recorded in this specification. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings. In the drawings:

[0013] Figure 1 It is a schematic flowchart of an intelligent reasoning method based on neural symbols provided by an embodiment of this specification;

[0014] Figure 2 It is a schematic structural diagram of an intelligent reasoning device based on neural symbols provided by an embodiment of this specification. Specific implementation manners

[0015] In order to enable those skilled in the art of this technology to better understand the technical solutions in this specification, the following will clearly and completely describe the technical solutions in the embodiments of this specification in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all the embodiments. Based on the embodiments of this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this specification.

[0016] An embodiment of this specification provides an intelligent reasoning method based on neural symbols. It should be noted that the execution subject in the embodiments of this specification can be a server or any device with data processing capabilities. Figure 1 It is a schematic flowchart of an intelligent reasoning method based on neural symbols provided by an embodiment of this specification, as Figure 1 shown, mainly including the following steps:

[0017] Step S101, obtain the user input information corresponding to the user and the pre-constructed hybrid domain knowledge base.

[0018] Among them, the hybrid domain knowledge base includes a first-order predicate logic rule base and a closed Markov logic network;

[0019] This method further includes: obtaining industry professional data in a specified industry, parsing the industry professional data, extracting atomic logical expressions, generating a first-order predicate logic rule base; representing the first-order predicate logic rule base through a Markov logic network to generate the closed Markov logic network.

[0020] The first-order predicate logic rule base is represented by a Markov logic network to generate the closed Markov logic network, specifically including: obtaining the first-order predicate logic rules in the first-order predicate logic rule base, replacing the variables in the first-order predicate logic rules with constants to construct closed atoms; constructing the nodes of the Markov logic network with the closed atoms, and converting the weights of the first-order predicate logic rules into edge weights to form a joint probability distribution, so as to generate the closed Markov logic network.

[0021] In one embodiment of this specification, professional data is obtained from target industries (such as healthcare and finance). For example, in the healthcare field, diagnostic records and symptom-disease association data are extracted from the electronic medical record library; in the financial field, financial indicators and risk rules are obtained from corporate annual reports and risk event libraries. Unstructured data (such as text and tables) is cleaned and annotated, and key entities (such as "symptom names", "disease types", "asset-liability ratios") and relationships (such as "symptom A is related to disease B", "indicator C exceeding the threshold triggers a risk") are extracted. Industry knowledge is transformed into first-order predicate logic expressions. According to expert experience or historical data statistics, an initial weight is assigned to each rule. For example, the weight of the rule "fever + cough → pneumonia" in the medical rule is 0.9, and the weight of the rule "current ratio < 1.5 → liquidity risk" in the financial rule is 0.85.

[0022] Replace the variables in the first-order predicate logic rules with specific instances. For example, the symptom records of patient "Zhang San" are "fever" and "cough". The rule base rules include rule R1: has_symptom(patient, fever) ∧ has_symptom(patient, cough) suspected_diagnosis(patient, pneumonia). Traverse all rules and instances to generate a set of closed atoms. For example, in a medical scenario, there may be closed atoms such as has_symptom(Zhang San, fever), has_symptom(Zhang San, cough), suspected_diagnosis(Zhang San, pneumonia), etc. Each closed atom corresponds to a binary node in the network, and the node value is 1 (true) or 0 (false). Convert the weights of the first-order predicate logic rules into the connection strength of the edges. For example, the rule weight of 0.9 corresponds to the potential function value of the edge as e^0.9. Based on the closed atom nodes and their connection relationships, construct a joint probability distribution formula where Z(ω) is the partition function, which is the sum of all closed atoms. A = {A1, A2,...} is all the closed atoms in the knowledge base. A r is a set of closed atoms composed of the closed atoms in a first-order predicate logic. φ1 is a potential function representing the number of times the first-order predicate logic rule is true. ω is the weights of all first-order predicate logic rules, and ω r is the weight of a predicate logic rule. The closed Markov logic network is constructed in the above manner. Obtain a hybrid domain knowledge base, and by replacing industry data and logical rules, it can be quickly adapted to multiple fields such as healthcare, finance, and law.

[0023] In step S102, the user input information is parsed to determine the input feature vector, and then the neural network inference is performed on the input feature vector through a pre-constructed neural inference module to determine the corresponding pseudo-label information.

[0024] Performing neural network inference on the input feature vector through a pre-constructed neural inference module to determine the corresponding pseudo-label information specifically includes: defining multiple types of neural network architectures in the neural inference module in advance, matching based on the input feature vector in the multiple types of neural network architectures to determine the target neural network architecture; performing inference on the input feature vector through the target neural network architecture, and outputting the pseudo-label and the corresponding label category confidence to determine the pseudo-label information.

[0025] In one embodiment of this specification, multiple types of neural network architectures are defined in the neural inference module in advance, including image processing architectures, sequence processing architectures, and structured data processing architectures. The image processing architecture uses a convolutional neural network (CNN), including a convolutional layer, a pooling layer, and a fully connected layer, and is used for image feature extraction (such as medical images, industrial inspection images); the sequence processing architecture uses a recurrent neural network (RNN) or Transformer, equipped with an attention mechanism, and is used for sequential data such as text and time series (such as financial text reports, device sensor time series signals); the structured data processing architecture uses a multi-layer perceptron (MLP), including a batch normalization layer and a Dropout layer, and is used for tabular data (such as enterprise financial indicators, customer attribute tables). The input dimension of the image features is [batch size, number of channels, height, width], which automatically matches the convolutional neural network architecture. The input dimension of the text / temporal features is [batch size, sequence length, word vector dimension], which matches the recurrent neural network or Transformer architecture. The input dimension of the structured features is [batch size, feature dimension], which matches the multi-layer perceptron architecture.

[0026] Perform standardization according to the input data type. For image data, perform normalization (such as scaling pixel values to [0,1]) and chunking. For text data, generate embedding representations through pre-trained word vectors (such as Word2Vec, BERT). For structured data, perform Z-Score standardization on numerical features and one-hot encoding on categorical features. Automatically select the corresponding architecture based on the input feature dimension. For example, if the input dimension is [256, 3, 224, 224] (image data), select the CNN architecture. Train a lightweight classification model (such as Random Forest) to predict the optimal architecture type according to the input feature statistics (such as mean, variance, sequence length). Input the input feature vector into the target architecture and generate the output probability distribution through calculations in each layer. Next, take CNN as an example to illustrate. Extract local features through the convolutional layer and output the disease classification probability through the fully connected layer. The output layer uses the Softmax function to generate the multi-class probability distribution or the Sigmoid function to generate the multi-label binary probability. Take the class corresponding to the highest probability in the probability distribution as the pseudo-label. For example, the medical image outputs [pneumonia: 0.92, cold: 0.05, others: 0.03], and the pseudo-label is "pneumonia". Take the highest probability value as the label confidence, and the confidence of the above pseudo-label is 0.92.

[0027] Set the confidence threshold according to the task requirements (such as medical diagnosis requires ≥0.9). If the confidence is lower than the threshold, trigger the manual review process and feedback the review result to the symbolic reasoning module to correct the rule weights. Regularly add the high-confidence pseudo-labels and their feature vectors feedback by users to the training set and fine-tune the target architecture parameters. Statistically calculate the inference accuracy and response time of each architecture type and dynamically eliminate inefficient architectures (such as RNN being replaced by Transformer).

[0028] Support unified reasoning of cross-modal data (image, text, numerical) through multi-type architecture matching. Automatically select the optimal architecture according to the input features to avoid wasting resources of a single architecture. For example, using MLP to process structured data is faster than CNN. Adding a new architecture type (such as a graph neural network) only requires expanding the architecture library without reconstructing the overall system.

[0029] In step S103, through the pre-constructed symbolic reasoning module, determine candidate rules according to the pseudo-label information and the first-order predicate logic rule library, and use the closed Markov logic network to perform collaborative reasoning on the pseudo-label information and the candidate rules to determine the corresponding inference knowledge information.

[0030] Determine candidate rules according to the pseudo-label information and the first-order predicate logic rule library, specifically including:

[0031] Parse the pseudo-label information into a first-order predicate logic expression to extract the core predicate and entity parameters; according to the core predicate, filter in the first-order predicate logic rule base to obtain at least one first-order logic rule whose rule head or rule body in the first-order predicate logic rule base contains the core predicate; based on the rule weights of the first-order logic rules, determine the candidate rule with the rule having the maximum weight.

[0032] Use the closed Markov logic network to perform collaborative reasoning on the pseudo-label information and the candidate rule to determine the corresponding inference knowledge information, specifically including: replacing the variables in the candidate rule with the entity parameters of the pseudo-label in the pseudo-label information to generate a closed atom corresponding to the candidate rule to determine the nodes of the closed Markov logic network; calculating the joint probability of the co-establishment of the pseudo-label and the candidate rule based on the probability distribution formula of the closed Markov logic network; when the joint probability meets the preset conditions, determine the corresponding inference knowledge information with the pseudo-label, where the inference knowledge information includes target closed atom assignment information, the joint probability, and an evidence chain, and the evidence chain includes a rule set and corresponding rule weights.

[0033] In an embodiment of the present specification, parse the pseudo-label into a first-order predicate logic expression. Assume that the pseudo-label output by the neural inference module is "Patient Zhang San is suspected of having pneumonia (confidence 0.92)" or "Enterprise A has liquidity risk (confidence 0.85)". Convert the medical pseudo-label to suspected diagnosis (Zhang San, pneumonia), where "Zhang San" is the patient entity and "pneumonia" is the diagnosis result. Convert the financial pseudo-label to risk warning (Enterprise A, liquidity risk), where "Enterprise A" is the entity and "liquidity risk" is the warning type. Extract the predicates (such as "suspected diagnosis", "risk warning") and entity parameters (such as "Zhang San", "Enterprise A") from the expression. The rule base is stored in the form of "rule number, logical rule, weight". For example: Rule R1: has symptom(patient, fever) ∧ has symptom(patient, cough) suspected diagnosis(patient, pneumonia), weight 0.9. Rule R2: current ratio(enterprise, X) ∧ X < 1.5 Risk warning (enterprise, liquidity risk), weight 0.85. Traverse the rule base and filter out the rules whose rule heads or rule bodies contain the core predicates. If the pseudo-label predicate is "suspected diagnosis", match rule R1; if it is "risk warning", match rule R2. Sort the rules in descending order of weight and select the rule with the highest weight as the candidate rule. For example, if R1 (weight 0.9) and R3 (weight 0.8) are matched, then R1 is selected as the candidate rule. Replace the variables in the candidate rule with specific entity parameters. Replace the variable "patient" in rule R1 with "Zhang San" to generate closed atoms: has_symptom(Zhang San, fever), has_symptom(Zhang San, cough), suspected_diagnosis(Zhang San, pneumonia). Each closed atom is used as a network node, and the node value is 1 (true) or 0 (false); the rule weight is converted into an edge weight.

[0034] Calculate the joint probability of the pseudo-label and the candidate rule being jointly established based on the joint probability formula of the closed Markov logic network. For example, if all the closed atoms of the candidate rule R1 are true, then its contribution value is 0.9×1 = 0.9, and the total probability is calculated in combination with other rules. If the joint probability exceeds the threshold (e.g., ≥0.8), adopt the pseudo-label as the final conclusion. Record the rules that support the conclusion and their weights. For example: Diagnostic conclusion: pneumonia (probability 0.88); Evidence chain: rule R1 (weight 0.9), rule R4 (weight 0.7).

[0035] Through the above technical solutions, by matching the pseudo-label with the first-order predicate logic rule base, all inference conclusions are accompanied by clear logical evidence chains, avoiding the problem of the unexplainability of the "black box inference" of pure neural networks; the neural inference module provides generalization ability (processing unstructured data such as images and texts), and the symbolic inference module injects domain knowledge (such as medical rules, financial risk control indicators) to solve the problem of the lack of professionalism of general large models; dynamically adjust the rule weights according to user feedback (such as reducing the weight in case of misjudgment) to avoid the rigidity of knowledge in traditional expert systems and adapt to the iteration of industry knowledge; the closed Markov logic network quantifies the conclusion confidence through joint probability calculation, supporting fuzzy logic and uncertainty scenarios; replacing the industry rule base can quickly switch application scenarios without reconstructing the overall architecture, reducing the multi-domain deployment cost.

[0036] After calculating the joint probability of the pseudo-label and the candidate rule being jointly established, the method further includes: when the joint probability does not meet the preset conditions, in the closed Markov logic network, use the assignment result with the highest probability to replace the pseudo-label to determine the corresponding inference knowledge; according to the probability corresponding to the assignment result and the probability corresponding to the pseudo-label, determine the probability difference, and use this probability difference as a conflict signal to inject it into the neural inference module through backpropagation to update the network parameters of the neural inference module.

[0037] In one embodiment of this specification, a preset probability threshold, such as 0.8 in a medical diagnosis scenario, is set. If the joint probability (e.g., 0.75) of the pseudo-label and the candidate rule holding jointly is lower than the threshold, it is determined as "not meeting the preset conditions". All possible combinations of closed atomic assignments are traversed in the closed Markov logic network, the joint probability of each combination is calculated, and the combination with the highest probability is selected as the alternative result. For example, the pseudo-label is "pneumonia (probability 0.75)", but there is another assignment combination in the closed network, "bronchitis (probability 0.83)", then the alternative result is "bronchitis". The result with the highest probability assignment is used as the final conclusion, and its probability and evidence chain are recorded. For example, the revised conclusion is bronchitis (probability 0.83); the evidence chain is rule R5 (weight 0.85), rule R6 (weight 0.78).

[0038] Calculate the difference between the pseudo-label probability and the alternative result probability, such as |0.75 - 0.83| = 0.08, as the conflict intensity index. Convert the conflict signal into a loss value of the neural inference module, and the loss value = conflict intensity ×

[0039] (alternative result probability - pseudo-label probability). For example, the conflict intensity is 0.08, and the loss value is 0.08 × (0.83 - 0.75) = 0.0064. Propagate the loss value back to the neural inference module through the gradient descent algorithm to adjust the network weights. For the target neural network architecture (such as CNN or MLP), update the convolutional kernel parameters or the weights of the fully connected layer according to the proportion of the loss value. For example, if the original pseudo-label is misjudged due to image feature extraction deviation, the recognition ability of relevant features is enhanced after the parameter update.

[0040] Push the revised conclusion (such as "bronchitis") to the user side for verification. If the user confirms it is correct, synchronously update the weights of relevant rules in the rule base (such as increasing the weight of R5 to 0.88). Statistically count the conflict frequency of the same pseudo-label type. If there are multiple conflicts in a short period (such as "pneumonia" being replaced continuously 3 times in medical diagnosis), trigger the reinforcement training of the neural inference module. Extract the high-conflict samples of this category from the historical data and perform targeted fine-tuning on the neural module.

[0041] Through probability verification and assignment replacement, avoid incorrect conclusions caused by low-confidence pseudo-labels, improve the reliability of system output. The backpropagation driven by the conflict signal realizes a "reasoning - feedback - learning" closed loop, enabling the neural module to gradually adapt to complex scenarios. The revised conclusion still comes with a complete rule and probability description to ensure the traceability of the decision-making process.

[0042] After determining the corresponding inference knowledge information, the method further includes: receiving the user's adoption feedback data and actual execution result data for the inference knowledge information; when the adoption feedback data is an adoption feedback, adjusting the logic rule weight corresponding to the inference knowledge information in the first-order predicate logic rule base according to the actual execution result data, so as to dynamically update the first-order predicate logic rule base.

[0043] In an embodiment of the present specification, the user submits an adoption feedback on the inference conclusion through the system interface, such as clicking "confirm correct" or "mark misjudgment", and uploads the actual execution result, such as the pathological report after medical diagnosis, the risk audit result after financial warning. For example, in a medical scenario, the user marks "the pneumonia diagnosis is correct" and uploads a report showing that the lesions have disappeared in the CT image of the patient after treatment. In a financial scenario, the user marks "the liquidity risk warning is effective" and uploads the audit document of the subsequent capital chain break of the enterprise.

[0044] Locate the set of logical rules (such as rules R1, R4) that support the conclusion in the rule base according to the feedback conclusion (such as "pneumonia") and the associated entity parameters (such as the patient "Zhang San"). If the feedback is "adopt", count the number of successful times of the rule in recent historical inferences; if it is "misjudgment", count the number of failure times. Dynamically calculate the rule weight according to the new weight calculation formula, new weight = original weight × decay coefficient + recent success rate × feedback coefficient, where the decay coefficient is 0.9 and the feedback coefficient is 0.1. Assuming that the recent success rate of rule R1 is 92%, the new weight is 0.9×0.9 + 0.92×0.1 = 0.81 + 0.092 = 0.902. Adjust the weight based on the prior probability of the actual execution result. If the prior correct rate of rule R1 is 90% and there are 10 new successful feedbacks, the posterior probability is (90 + 10) / (100 + 10) ≈ 90.9%, and the weight is updated to 0.909. If multiple rules in the same scenario derive conflicting conclusions (such as rule R1 supports "pneumonia" and rule R7 supports "bronchitis"), punish and reduce the weight of the wrong rule according to the feedback result. The user confirms that "pneumonia" is correct, then reduce the weight of rule R7. The updated weight is written into the rule base in real time, and the new weight is immediately used to calculate the joint probability in subsequent inferences. Extract high-conflict cases from historical data (such as "pneumonia → bronchitis" that has been corrected), and re-run the inference process to verify the rationality of the conclusion after weight adjustment.

[0045] Dynamically adjust the rule weight through feedback, so that the system gradually approaches the real scenario requirements, avoiding manual maintenance costs; manage conflicting rules differently, enhancing decision consistency in complex scenarios. The weight update is decoupled from the inference process, ensuring the stable operation of the system in high-concurrency scenarios.

[0046] Through the technical solution of the embodiments of this specification, the hybrid domain knowledge base includes a first-order predicate logic rule base and a closed Markov logic network, which combines the precise rule representation of symbolicism with the Markov logic network based on the probabilistic graphical model. It can handle both deterministic knowledge and knowledge with ambiguity and uncertainty, making up for the deficiencies of traditional single knowledge representation methods, making knowledge representation more comprehensive and flexible, and meeting the needs of different types of knowledge representation in complex scenarios of multiple industries; first, the neural inference module performs neural network inference on the input feature vector to obtain pseudo-label information, then the symbolic inference module combines the pseudo-label information and the first-order predicate logic rule base to determine candidate rules, and finally, the closed Markov logic network performs collaborative inference, combining the powerful feature extraction and learning ability of the neural network and the interpretability and logical reasoning ability of the symbolic system. It can not only process complex non-linear data but also provide an interpretable reasoning process, to a certain extent solving the problem that it is difficult to have both interpretability and generalization ability; the neural inference module can learn potential features and patterns from a large amount of data, providing a basis for subsequent inference. The symbolic inference module performs inference based on the rule base, ensuring the logic and accuracy of the inference. The collaborative inference of the closed Markov logic network further comprehensively considers the uncertainty of various factors and optimizes the results through probabilistic inference. The collaborative working mode of multiple modules can make full use of the advantages of different modules, improve the accuracy and generalization ability of the inference, and enable it to better adapt to complex scenarios across industries, unlike vertical domain large models that are limited to a single domain and cannot be generalized to other domains; different from the black-box inference of pure neural networks, the symbolic inference module performs inference based on a clear first-order predicate logic rule base, and can display the rules and logical relationships used during the inference process, making the inference result interpretable. Users can understand how the system draws conclusions based on the rules and input information, which is crucial for scenarios that require a reliable decision-making basis (such as medical diagnosis, financial risk control, etc.), and helps improve users' trust in the inference results.

[0047] The embodiments of this specification also provide a neural-symbolic based intelligent inference device, as Figure 2 shown. The device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above method.

[0048] The embodiments of this specification also provide a non-volatile computer storage medium storing computer-executable instructions, and the computer-executable instructions are set to: execute the above method.

[0049] Each embodiment in this specification is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other, and the key point of each embodiment is to illustrate the differences from other embodiments. In particular, for the embodiments of devices, equipment, and non-volatile computer storage media, since they are basically similar to the method embodiments, the description is relatively simple, and reference can be made to the corresponding parts of the method embodiments for relevant content.

[0050] The above describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be executed in a different order than in the embodiments and still achieve the desired results. Additionally, the processes depicted in the figures do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0051] The devices and media provided in the embodiments of this specification correspond one-to-one with the methods. Therefore, the devices and media also have beneficial technical effects similar to those of their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be elaborated here.

[0052] Those skilled in the art should understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, this specification can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0053] This specification is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices produce a device for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0054] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction means that implements the function specified in one or more of the blocks and / or processes. Figure 1 one or more processes and / or blocks Figure 1 specified in the block or blocks.

[0055] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable apparatus provide steps for implementing the function specified in one or more of the blocks and / or processes. Figure 1 one or more processes and / or blocks Figure 1 specified in the block or blocks.

[0056] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0057] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM) and / or non-volatile memory such as read only memory (ROM) or flash memory (flash RAM). Memory is an example of computer-readable media.

[0058] Computer-readable media includes both permanent and non-permanent, removable and non-removable media implemented by any method or technology for storage of information such as computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital versatile discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0059] It should also be noted that the term "comprising", "including" or any other variation thereof is intended to cover non-exclusive inclusion, such that a process, method, commodity or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "comprising an..." does not preclude the presence of additional identical elements in the process, method, commodity or device comprising said element.

[0060] The above description is only one or more embodiments of this specification and is not intended to limit this specification. For those skilled in the art, one or more embodiments of this specification can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of one or more embodiments of this specification shall be included within the scope of the claims of this specification.

Claims

1. An intelligent reasoning method based on neuro-symbolism, characterized in that, The method includes: Obtaining user input information corresponding to a user and a pre-constructed hybrid domain knowledge base, where the hybrid domain knowledge base includes a first-order predicate logic rule base and a closed Markov logic network; Parsing the user input information to determine an input feature vector, and performing neural network inference on the input feature vector through a pre-constructed neural inference module to determine corresponding pseudo-label information; Determining candidate rules through a pre-constructed symbolic inference module according to the pseudo-label information and the first-order predicate logic rule base, and performing collaborative inference on the pseudo-label information and the candidate rules by using the closed Markov logic network to determine corresponding inference knowledge information.

2. The intelligent reasoning method based on neural symbols according to claim 1, wherein The method further includes: Obtaining industry professional data within a specified industry, parsing the industry professional data, extracting atomic logical expressions, and generating a first-order predicate logic rule base; Representing the first-order predicate logic rule base through a Markov logic network to generate the closed Markov logic network.

3. The intelligent reasoning method based on neural symbols according to claim 2, characterized in that Representing the first-order predicate logic rule base through a Markov logic network to generate the closed Markov logic network, specifically including: Obtaining the first-order predicate logic rules in the first-order predicate logic rule base, replacing the variables in the first-order predicate logic rules with constants to construct closed atoms; Using the closed atoms to construct the nodes of the Markov logic network, converting the weights of the first-order predicate logic rules into edge weights to form a joint probability distribution, and generating the closed Markov logic network.

4. The intelligent reasoning method based on neural symbols according to claim 1, wherein Performing neural network inference on the input feature vector through a pre-constructed neural inference module to determine corresponding pseudo-label information, specifically including: Pre-defining multiple types of neural network architectures in the neural inference module, matching based on the input feature vector in the multiple types of neural network architectures to determine a target neural network architecture; Performing inference on the input feature vector through the target neural network architecture, outputting a pseudo-label and the corresponding label category confidence level to determine the pseudo-label information.

5. A neuro-symbolic based intelligent reasoning method according to claim 1, characterized in that, Determining candidate rules according to the pseudo-label information and the first-order predicate logic rule base, specifically including: Parsing the pseudo-label information into a first-order predicate logic expression to extract a core predicate and entity parameters; Filtering in the first-order predicate logic rule base according to the core predicate to obtain at least one first-order logic rule whose rule head or rule body in the first-order predicate logic rule base contains the core predicate; Based on the rule weights of the first-order logic rules, determining the candidate rule with the maximum weight.

6. The intelligent reasoning method based on neural symbols according to claim 1, characterized in that Performing collaborative inference on the pseudo-label information and the candidate rules by using the closed Markov logic network to determine corresponding inference knowledge information, specifically including: Replacing the variables in the candidate rules with the entity parameters of the pseudo-labels in the pseudo-label information to generate closed atoms corresponding to the candidate rules to determine the nodes of the closed Markov logic network; Calculating the joint probability of the co-establishment of the pseudo-label and the candidate rules based on the probability distribution formula of the closed Markov logic network; When the joint establishment probability meets the preset conditions, the corresponding inference knowledge information is determined based on the pseudo-label, where the inference knowledge information includes target closed atom assignment information, the joint establishment probability, and an evidence chain, and the evidence chain includes a rule set and corresponding rule weights.

7. The intelligent reasoning method based on neural symbols according to claim 6, wherein After calculating the joint establishment probability of the pseudo-label and the candidate rule being jointly established, the method further includes: When the joint establishment probability does not meet the preset conditions, in the closed Markov logic network, the assignment result with the highest probability is used to replace the pseudo-label to determine the corresponding inference knowledge; According to the probability corresponding to the assignment result and the probability corresponding to the pseudo-label, a probability difference is determined, and the probability difference is used as a conflict signal and injected into the neural inference module through backpropagation to update the network parameters of the neural inference module.

8. The intelligent reasoning method based on neural symbols according to claim 6, wherein, After determining the corresponding inference knowledge information, the method further includes: Receiving the adoption feedback data and actual execution result data of the user for the inference knowledge information; When the adoption feedback data is an adoption feedback, according to the actual execution result data, the logical rule weights corresponding to the inference knowledge information are adjusted in the first-order predicate logic rule base to dynamically update the first-order predicate logic rule base.

9. An intelligent inference device based on neuro-symbolics, characterized in that, The device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method according to any one of claims 1-8.

10. A non-volatile computer storage medium stores computer-executable instructions, characterized in that, The computer-executable instructions are set to: execute the method according to any one of claims 1-8.

Images