Text credibility evaluation method based on nerve symbol causal reasoning

Abstract

The invention relates to a text credibility assessment method based on neuro-symbol causal reasoning. The method comprises the following steps: acquiring a to-be-assessed text, and preprocessing the to-be-assessed text; synchronously extracting an entity relationship triple from the preprocessed text based on a bidirectional gated graph neural network, calculating the confidence coefficient of the triple, and generating structured data with a confidence coefficient weight; based on the structured data, utilizing a causal perception graph embedding model, self-supervising and learning an explicit / implicit causal relationship between entities, and generating a dynamic probabilistic causal graph; automatically searching a multi-hop causal path by using a graph reinforcement learning framework, and converting a high-confidence path into an interpretable logic rule; and a hierarchical Bayesian graph network is utilized to fuse the symbol rule confidence feature and the neural semantic feature, a joint credibility evaluation value is generated, and evaluation of the text credibility is realized. Compared with the prior art, the method has the advantages of high evaluation result reliability, high decision process transparency, wide scene adaptability and the like.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making, and in particular to a text credibility assessment method based on neural symbolic causal reasoning. Background Technology

[0002] In intelligent decision-making, there are numerous tasks that utilize machine learning for text reasoning. However, while these machine learning-based methods perform exceptionally well in certain types of text reasoning tasks, the "black box" nature of deep learning leads to reliability issues in decision-making scenarios that are sensitive and require high transparency. By evaluating the confidence level of text reasoning results, we can accurately determine whether the results generated by the text model are reasonable, thereby better assisting relevant personnel in making decisions.

[0003] Currently, methods for assessing the credibility of textual reasoning can be mainly divided into rule-based methods and deep learning-based methods. Rule-based methods analyze the logical structure and linguistic expression of text content to construct a fixed set of evaluation criteria to judge the credibility of the text. This method is highly effective when dealing with structured and semantically explicit text, but its adaptability to a wide range of application scenarios is weak due to a lack of understanding of complex contexts and implicit meanings. In contrast, deep learning-based methods utilize neural network models to learn the deep semantic features of text and combine them with contextual information for reasoning, thereby improving the accuracy of textual credibility assessment. Although these methods perform well in reasoning tasks involving indirect expressions, the accuracy and stability of the model's inference results still face challenges when encountering sparse data or large sample bias.

[0004] Chinese invention patent CN120258453A discloses a method for constructing a decision-making model based on a big data environment, including: acquiring multi-source heterogeneous data; extracting explicit causal relationships from historical data using a causal discovery algorithm, and generating an interpretable knowledge graph by combining prior knowledge from equipment manuals; constructing a neural symbolic joint model, synchronously adjusting neural network parameters and rule confidence weights to achieve collaborative learning between data-driven and knowledge-driven approaches; and performing multi-dimensional logical verification on the trained model. This invention, based on real-time data streams and expert feedback, employs an incremental conflict detection and rule correction mechanism to dynamically optimize the model, constructing a closed-loop iterative decision support system. This effectively improves the accuracy and interpretability of the decision-making model, facilitating human understanding and acceptance, making the decision-making process more transparent, fair, and open, and resulting in more accurate and real-time decisions. However, it still suffers from performance fluctuations when data is sparse or sample biased, insufficient transparency in sensitive decision-making domains, and insufficient adaptability to scenarios involving complex text reasoning rich in implicit meaning.

[0005] In summary, there is currently a lack of a text credibility assessment method based on neural symbolic causal reasoning to solve or partially solve the aforementioned problems. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a text credibility assessment method based on neural symbolic causal reasoning, so as to solve or partially solve the problems of performance fluctuations when data is sparse or sample biased, insufficient decision transparency in sensitive decision domains, and insufficient adaptability to scenarios with complex text reasoning rich in implicit meaning.

[0007] The objective of this invention can be achieved through the following technical solutions: According to one aspect of the present invention, a method for evaluating text credibility based on neural symbolic causal reasoning is provided, specifically including: S1. Obtain the text to be evaluated and preprocess it, including entity normalization and encoding of unstructured information; S2. Based on a bidirectional gated graph neural network, entity relation triples are extracted synchronously from the preprocessed text, the confidence of the triples is calculated, and structured data with confidence weights is generated. S3. Based on the structured data, use the causal perception graph embedding model to self-supervised learn explicit / implicit causal relationships between entities and generate a dynamic probabilistic causal graph. S4. Based on the causal graph, use a graph reinforcement learning framework to automatically search for multi-hop causal paths, transform high-confidence paths into interpretable logical rules, and calculate the corresponding rule confidence. S5. Using the aforementioned rule confidence as the symbolic rule confidence feature, a hierarchical Bayesian graph network is used to fuse the symbolic rule confidence feature with the neurosemantic feature to generate a joint credibility evaluation value, thereby achieving the evaluation of text credibility.

[0008] As a preferred technical solution, the entity normalization is used to uniformly name entities in the text through pre-trained word vectors, and the unstructured information encoding is used to transform the unstructured text into a vectorized representation through a graph convolutional network.

[0009] As a preferred technical solution, the confidence level of the triples in S2 The formula, calculated using the gating mechanism, is as follows: In the formula, For activation function, For connection weights, This is a positive propagation state. In the backpropagation state, For dynamic parameters of the gating mechanism, This represents the product operation. This indicates the modulo operation.

[0010] As a preferred technical solution, the causal perception graph embedding model in S3 is jointly optimized by the causal direction discrimination loss function and the node similarity loss function, and the overall optimization function is... The formula is: In the formula, It is a node similarity loss function used to calculate the similarity measure between entities. It is a causal direction discrimination loss function that uses KL divergence to analyze the latent variables in the network. z To impose constraints, For the entity set Any two entities in, Representing a graph neural network, for Explicit relations in entity sets Represents a learnable relational latent mapping. For hyperparameters, KL Represents the KL divergence between two variables. To in a given entity pair In the case of latent variables z The distribution Latent variables z The prior distribution, This indicates the modulo operation.

[0011] As a preferred technical solution, the S4 graph reinforcement learning framework models causal path search as a Markov decision process, utilizes a policy network to select node transition actions, discovers multi-hop causal paths that satisfy the minimum confidence threshold δ, and transforms the paths into predicate logic rules stored in an extensible rule base. The action selection probability of the policy network is: In the formula, the policy network encodes the current node through a graph attention network. Based on neighborhood information, the next-hop node is dynamically selected. Represents time step t The state at any given moment includes the current node, the visited paths, and the maximum remaining hops. Representative from The node starts at time step t The selected action is the selectable adjacent node. This represents a learnable weight matrix. For activation function, For example, a graph attention network.

[0012] As a preferred technical solution, the reward function of the graph reinforcement learning framework in S4 for: In the formula, , , These are the coefficients for path confidence, causal strength, and path length, respectively. The reward function guides the strategy network to prioritize inference chains with high confidence, strong causal relationships, and simple paths.

[0013] As a preferred technical solution, the rule confidence level is calculated by combining the local confidence levels of all triples on the path and the path length, and the calculation formula is as follows: In the formula, γ is the causal intensity attenuation coefficient, and the rule confidence score serves as the symbolic rule confidence score feature. r .

[0014] As a preferred technical solution, the hierarchical Bayesian graph network in S5 models the probability distribution of entities through a variational autoencoder, and its formula is: In the formula, Indicates input data Mapping to latent variables The probability distribution, Represents a Gaussian distribution. Represents latent variables The One dimension, This represents the mean predicted by the encoder network. This represents the standard deviation predicted by the encoder network. To provide evidence for the next term, representing the optimization of goals. For a given The expected value under, To reconstruct the log probabilities, the model will include latent variables. Decode and restore to the original input ability, The set of input features to be fused specifically includes symbolic rule confidence feature vectors and neurosemantic feature vectors. for In the Information in dimensions KL Represents the KL divergence between two variables. Latent variables z The prior distribution, It is a variational posterior distribution.

[0015] As a preferred technical solution, the neural semantic feature vector is obtained by performing a pooling operation on the embedding representations of all entities in the dynamic probabilistic causal graph generated by S3. s The neural semantic feature vector s The hierarchical Bayesian graph network comprehensively encodes the local relational semantics and global causal structure of the text; it uses symbolic rule confidence features. r and neurosemantic features s As the observation input, the joint probability distribution of the entities is modeled through a variational autoencoder, and the hierarchical Bayesian graph network incorporates the symbolic rule confidence features. r With neurosemantic features s Common encoding as latent variables The probability distribution is obtained by sampling to obtain latent variables. A joint credibility assessment value is generated through variational inference.

[0016] As a preferred technical solution, based on the obtained joint credibility assessment value, and according to the preset credibility threshold and confidence interval, the credibility assessment result of the drug report text to be evaluated is obtained.

[0017] Compared with the prior art, the present invention has at least one of the following beneficial effects: (1) Improve the reliability and robustness of evaluation results: This invention utilizes the dynamic probabilistic causal graph generated by the causal perception graph embedding model, and combines the graph reinforcement learning framework to automatically search for multi-hop causal paths to transform high-confidence paths into interpretable logical rules. It integrates the interpretability of symbolic reasoning with the stability of causal logic, effectively overcoming the performance fluctuation problem of pure deep learning models when data is sparse or sample biased. At the same time, it is more flexible and in-depth than rule-based methods when dealing with complex contexts and implicit meanings, significantly improving the accuracy and stability of evaluation results.

[0018] (2) Enhance the transparency and interpretability of the decision-making process: This invention uses the self-supervised learning of explicit / implicit causal relationships between entities by embedding the causal perception graph model and generates a dynamic probabilistic causal graph, which provides a clear decision path and causal relationship for decision-making. Decision-makers can understand why the model assigns a high or low confidence level to a certain reasoning result, which meets the transparency requirements of sensitive decision-making fields and enhances users' trust in the intelligent decision-making system.

[0019] (3) Wide range of scenarios: By combining the symbolic rule confidence features extracted by logical rules and the neurosemantic features of deep learning, the present invention generates a joint text credibility assessment, which enables the method of the present invention to handle both structured and well-defined texts and to effectively deal with unstructured and complex text reasoning tasks rich in implicit meanings. It can adapt to a variety of application scenarios from routine analysis to high-end decision support, and solves the problem of weak adaptability of existing single methods. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0021] 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 should fall within the scope of protection of the present invention.

[0022] To address the problems existing in the aforementioned prior art, this embodiment provides a text credibility assessment method based on neural symbolic causal reasoning, such as... Figure 1 As shown, this embodiment uses a user report from an online drug forum as an example to illustrate the method of the present invention in detail, specifically including: Enter the text: "After taking brand A antihypertensive medication for three days, I experienced persistent dizziness and itchy skin. On the second day, my blood pressure dropped to 90 / 60 mmHg, and my liver function indicator ALT rose to three times the normal value." S1. The input sentence is preprocessed, and entities in the text are named uniformly by pre-trained word vectors. The unstructured text is transformed into a vectorized representation by a graph convolutional network.

[0023] The specific details of step S1 are as follows: S1.1. Use a pre-trained language model to perform part-of-speech tagging to find entities in the text.

[0024] S1.2. Use graph convolutional networks to encode the spatiotemporal features of the text.

[0025] S1.3. Output the standardized entity set and semantic vector.

[0026] S2. Based on a bidirectional gated graph neural network, entity relation triples are extracted synchronously from the preprocessed text, the confidence of the triples is calculated, and structured data with confidence weights is generated.

[0027] The specific details of step S2 are as follows: S2.1. Forward propagation captures local syntactic relations. Assume that forward propagation extracts ("taking", "Brand A antihypertensive drug"), ("occurring", "dizziness"), ("occurring", "skin itching"), and generates confidence scores.

[0028] S2.2. Backpropagation to verify global logical consistency and confirm the causal relationship between symptoms and medication time.

[0029] Calculate the confidence level of the triplet using the formula. : In the formula, For activation function, For connection weights, This is a positive propagation state. In the backpropagation state, For dynamic parameters of the gating mechanism, This represents the product operation. This indicates the modulo operation.

[0030] Output a set of triples with confidence scores. Each entity-relation triple contains two entities, the relationship between the two entities, and the confidence score of the triple.

[0031] S3. Based on structured data, utilize the causal perception graph embedding model to self-supervised learn explicit / implicit causal relationships between entities and generate dynamic probabilistic causal graphs.

[0032] The causal perception graph embedding model jointly optimizes the causal direction discrimination loss function and the node similarity loss function, resulting in an overall optimization function. The formula is: In the formula, It is a measure of similarity between entities. It uses KL divergence to analyze the latent variables in the network. z To impose constraints, For the entity set Any two entities in, Representing a graph neural network, for Explicit relations in entity sets Represents a learnable relational latent mapping. For hyperparameters, KL Represents the KL divergence between two variables. To in a given entity pair In the case of latent variables z The distribution Latent variables z The prior distribution, This indicates the modulo operation.

[0033] The specific details of step S3 are as follows: S3.1. Input the set of triples T from step S2.

[0034] S3.2. Construct a causal graph and generate a causal chain: "Taking brand A antihypertensive drug → Blood pressure decreases → Dizziness", with an intensity of min(0.98,0.88,0.85)×time decay factor=0.83.

[0035] Construct another causal chain: "Taking brand A antihypertensive drug → ALT increase → skin itching", with a strength of min(0.98,0.79,0.82)=0.77.

[0036] If the probability of Y occurring when X occurs is greater than the probability of Y occurring independently plus the threshold α, then add a causal edge X→Y, with strength equal to the difference in conditional probabilities. To determine the causal direction, analyze the semantics of the relational predicate. If there is a bidirectional relationship "(A, influence, B)@0.7, (B, countermeasure, A)@0.8", take the one with higher confidence as the dominant causal direction: B→A.

[0037] If both "A→B→C" and "A→B→D" exist, compare the total confidence of the paths and retain the path with the highest confidence.

[0038] If a cycle "A→B→C→A" is detected, the edge with the lowest strength is automatically pruned. This mechanism can effectively break the causal paradox, ensure the acyclicity of the graph and the logical consistency of the reasoning, thereby improving the accuracy and interpretability of subsequent credibility assessments.

[0039] Once the causal direction is determined, all relationships are unidirectional causal relationships.

[0040] S3.3. Output a drug adverse reaction map containing multiple causal pathways.

[0041] S4. Based on causal graphs, a graph reinforcement learning framework is used to automatically search for multi-hop causal paths and transform high-confidence paths into interpretable logical rules.

[0042] The specific details of step S4 are as follows: S4.1. Define the search space and search strategy, including the state (current node, visited path, maximum remaining hops) and actions (optional adjacent nodes).

[0043] S4.2. Initialize the search strategy, sort the adjacent edges of each node by confidence, and perform path memory management (prohibit repeated visits and record the product of historical path confidence).

[0044] The graph reinforcement learning framework models causal path search as a Markov decision process, uses a policy network to select node transition actions, discovers multi-hop causal paths that satisfy the minimum confidence threshold δ, and transforms these paths into predicate logic rules stored in an extensible rule base. The action selection probabilities of the policy network are: In the formula, the policy network encodes the current node through a graph attention network. Based on neighborhood information, the next-hop node is dynamically selected. Represents time step t The state at any given moment includes the current node, the visited paths, and the maximum remaining hops. Representative from The node starts at time step t The selected action is the selectable adjacent node. This represents a learnable weight matrix. For activation function, For example, a graph attention network.

[0045] Reward function of graph reinforcement learning framework for: In the formula, , , These are the coefficients for path confidence, causal strength, and path length, respectively. The reward function guides the policy network to prioritize inference chains with high confidence, strong causal relationships, and simple paths.

[0046] The rule confidence score is calculated by combining the local confidence scores of all triples along the path and the path length. The calculation formula is as follows: Where γ is the causal intensity attenuation coefficient, and this rule confidence score serves as a feature of the sign rule confidence score. r This is used for subsequent fusion evaluation.

[0047] A multi-hop path dynamic search is performed, starting from the initial node. Path exploration is completed when the termination conditions (reaching the target node, confidence level below a threshold, or reaching the maximum number of hops) are met. Starting from the node "taking brand A antihypertensive medication," two main paths are discovered: “Path 1: Taking medication → Lowering blood pressure → Dizziness (Path confidence: 0.83)”; “Pathway 2: Taking medication → ALT elevation → skin itching (pathway confidence: 0.77)”

[0048] S5. Based on logical rules, extract symbolic rule confidence features, and use a hierarchical Bayesian graph network to fuse symbolic rule confidence features and neurosemantic features to generate a joint credibility evaluation value, thereby achieving the evaluation of text credibility.

[0049] The specific details of step S5 are as follows: S5.1. Perform feature weighting. First, extract symbolic rule features by directly using the rule confidence from step S4. Then, extract neural semantic features and calculate the similarity of key entity vectors.

[0050] S5.2. Perform hierarchical uncertainty modeling.

[0051] First, the confidence of the rules is dynamically adjusted through symbolic logic processing. Short paths (≤2 hops) are retained first, while long paths (≥4 hops) are attenuated additionally (e.g., 3-hop paths are attenuated by 20%). Direct conflicts in the current batch of rules are detected (e.g., both "detection started" and "detection canceled" conclusions exist simultaneously). The confidence of conflicting rules is reduced proportionally. The medical correlation between elevated ALT and skin itching is relatively weak, so the confidence of the conflicting rules is reduced by 10%.

[0052] Hierarchical Bayesian graphical networks model the probability distribution of entities using variational autoencoders, with the following formula: In the formula, Represents the confidence feature of symbolic rules r and neural semantic feature vectors s Mapping to latent variables The probability distribution, Represents a Gaussian distribution. Represents latent variables The One dimension, This represents the mean predicted by the encoder network. This represents the standard deviation predicted by the encoder network. To provide evidence for the next term, representing the optimization of goals. For a given The expected value under, To reconstruct the log probabilities, the model will include latent variables. Decoding and restoring to symbolic rule confidence features r and neural semantic feature vectors s ability, Confidence features for symbolic rules r and neural semantic feature vectors s In the Information in dimensions KL Represents the KL divergence between two variables. Latent variables z The prior distribution, It is a variational posterior distribution.

[0053] Neuro-semantic feature vector s The joint confidence assessment value is a full-graph level feature vector obtained by pooling the embedding representations of all entities in the probabilistic causal graph generated by S3. This vector comprehensively encodes the local relational semantics and global causal structure of the text. The specific method for generating the joint confidence assessment value is as follows: a hierarchical Bayesian graph network uses symbolic rule confidence features. r and neurosemantic features s As the observation input, the joint probability distribution of the entities is modeled using a variational autoencoder. Its variational distribution is as follows: The aim is to approximate the true posterior distribution. The optimization objective of this model is to maximize the lower bound of evidence (ELBO): decoding Simultaneously reconstruct symbolic features r With neural characteristics s In the inference phase, the distribution will be... Latent variables obtained from mid-sampling The input is fed into a prediction layer to generate the final joint credibility assessment value.

[0054] A neurosemantic layer is used for processing. If the vector of an entity differs significantly from other entities of the same type within the batch, it is marked as an "abnormal entity." The overall semantic coherence of the rule path (such as the average similarity of each node in the path) is calculated, and a warning is triggered if it is below a threshold. The semantic coherence scores of "low blood pressure" and "dizziness" are high (0.89), while the semantic coherence scores of "high ALT" and "itchy skin" are moderate (0.72). This evaluation, as a common-sense-based semantic check, can effectively filter out illogical false paths, improving the accuracy and robustness of the final evaluation from another dimension.

[0055] S5.3. Dynamic fusion and confidence intervals.

[0056] Weights are assigned based on the symbol rule weights and neural feature weights, and a weighted average of the symbol and neural results (symbols account for 70%, neural accounts for 30%) is calculated. The resulting comprehensive value is 0.78.

[0057] Based on the fact that the semantic coherence score of path 2 is "moderate", the extended confidence interval is 0.78 ± 0.18.

[0058] Results: The overall credibility assessment value of 0.78 is higher than the credibility threshold of 0.70 for adverse drug reaction reports, and the confidence interval (0.60-0.96) is mostly within the credibility range. The system determines that the text description is normal and constitutes a reasonable adverse drug reaction report, and recommends that it be reported to the drug regulatory authority for further verification. This assessment result provides an important reference for drug safety monitoring.

[0059] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A text credibility assessment method based on neural symbolic causal reasoning, characterized in that, Specifically, it includes: S1. Obtain the text to be evaluated and preprocess it, including entity normalization and encoding of unstructured information; S2. Based on a bidirectional gated graph neural network, entity relation triples are extracted synchronously from the preprocessed text, the confidence of the triples is calculated, and structured data with confidence weights is generated. S3. Based on the structured data, use the causal perception graph embedding model to self-supervised learn explicit / implicit causal relationships between entities and generate a dynamic probabilistic causal graph. S4. Based on the causal graph, use a graph reinforcement learning framework to automatically search for multi-hop causal paths, transform high-confidence paths into interpretable logical rules, and calculate the corresponding rule confidence. S5. Using the aforementioned rule confidence as the symbolic rule confidence feature, a hierarchical Bayesian graph network is used to fuse the symbolic rule confidence feature with the neurosemantic feature to generate a joint credibility evaluation value, thereby achieving the evaluation of text credibility.

2. The text credibility assessment method based on neural symbolic causal reasoning according to claim 1, characterized in that, The entity normalization uses pre-trained word vectors to uniformly name entities in the text, and the unstructured information encoding uses a graph convolutional network to transform the unstructured text into a vectorized representation.

3. The text credibility assessment method based on neural symbolic causal reasoning according to claim 1, characterized in that, The confidence level of the triplet in S2 The formula, calculated using the gating mechanism, is as follows: In the formula, For activation function, For connection weights, This is a positive propagation state. In the backpropagation state, For dynamic parameters of the gating mechanism, This represents the product operation. This indicates the modulo operation.

4. The text credibility assessment method based on neural symbolic causal reasoning according to claim 1, characterized in that, The causal perception graph embedding model in S3 is jointly optimized using a causal direction discrimination loss function and a node similarity loss function, resulting in an overall optimization function. The formula is: In the formula, It is a node similarity loss function used to calculate the similarity measure between entities. It is a causal direction discrimination loss function that uses KL divergence to analyze the latent variables in the network. z To impose constraints, For the entity set Any two entities in, Representing a graph neural network, for Explicit relations in entity sets Represents a learnable relational latent mapping. For hyperparameters, KL Represents the KL divergence between two variables. To in a given entity pair In the case of latent variables z The distribution Latent variables z The prior distribution, This indicates the modulo operation.

5. The text credibility assessment method based on neural symbolic causal reasoning according to claim 1, characterized in that, The S4 graph reinforcement learning framework models causal path search as a Markov decision process, utilizes a policy network to select node transition actions, discovers multi-hop causal paths that satisfy the minimum confidence threshold δ, and transforms these paths into predicate logic rules stored in an extensible rule base. The action selection probability of the policy network is: In the formula, the policy network encodes the current node through a graph attention network. Based on neighborhood information, the next-hop node is dynamically selected. Represents time step t The state at any given moment includes the current node, the visited paths, and the maximum remaining hops. Representative from The node starts at time step t The selected action is the selectable adjacent node. This represents a learnable weight matrix. For activation function, For example, a graph attention network.

6. The text credibility assessment method based on neural symbolic causal reasoning according to claim 1, characterized in that, The reward function of the graph reinforcement learning framework in S4 for: In the formula, , , These are the coefficients for path confidence, causal strength, and path length, respectively. The reward function guides the strategy network to prioritize inference chains with high confidence, strong causal relationships, and simple paths.

7. The text credibility assessment method based on neural symbol causal reasoning according to claim 1, characterized in that, The rule confidence score is calculated by combining the local confidence scores of all triples along the path and the path length. The calculation formula is as follows: In the formula, γ is the causal intensity attenuation coefficient, and the rule confidence score serves as the symbolic rule confidence score feature. r .

8. The text credibility assessment method based on neural symbolic causal reasoning according to claim 1, characterized in that, The hierarchical Bayesian graph network in S5 models the probability distribution of entities through a variational autoencoder, and its formula is: In the formula, Indicates input data Mapping to latent variables The probability distribution, Represents a Gaussian distribution. Represents latent variables The One dimension, This represents the mean predicted by the encoder network. This represents the standard deviation predicted by the encoder network. To provide evidence for the next term, representing the optimization of goals. For a given The expected value under, To reconstruct the log probabilities, the model will include latent variables. Decode and restore to the original input ability, The set of input features to be fused specifically includes the symbolic rule confidence feature vector. r With neuro-semantic feature vectors s , for In the Information in dimensions KL Represents the KL divergence between two variables. Latent variables z The prior distribution, It is a variational posterior distribution.

9. A text credibility assessment method based on neural symbolic causal reasoning according to claim 8, characterized in that, The neural semantic feature vector is obtained by performing pooling operations on the embedding representations of all entities in the dynamically probabilistic causal graph generated by S3. s The neural semantic feature vector s It comprehensively encodes the local relational semantics and global causal structure of the text; The hierarchical Bayesian graph network uses symbolic regularity confidence features. r and neurosemantic features s As the observation input, the joint probability distribution of the entities is modeled through a variational autoencoder, and the hierarchical Bayesian graph network incorporates the symbolic rule confidence features. r With neurosemantic features s Common encoding as latent variables The probability distribution is obtained by sampling to obtain latent variables. A joint credibility assessment value is generated through variational inference.

10. The text credibility assessment method based on neural symbolic causal reasoning according to claim 1, characterized in that, Based on the obtained joint credibility assessment value, and according to the preset credibility threshold and confidence interval, the credibility assessment result of the drug report text to be evaluated is obtained.

Images