Audit method and system for output text

Abstract

The embodiment of the specification provides an auditing method and system of output text, the method comprises extracting a cause-effect chain from the output text of a text generation algorithm, the cause-effect chain comprising a cause entity, an effect entity and a relationship type; performing hierarchical progressive verification on the cause-effect chain based on multiple cause-effect layers to determine a confidence sequence, the confidence sequence comprising a confidence determined based on at least one of the multiple cause-effect layers, and the multiple cause-effect layers at least comprising an expression perception layer and a negative cause-effect layer; and determining the credibility of the output text based on the confidence sequence. The method can realize multi-layer cause-effect layer progressive verification, effectively verify objective deep cause-effect, and enhance the rigor and accuracy of the credibility evaluation of the output text.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and in particular to an auditing method and system for output text. Background Technology

[0002] Detecting causal relationships between variables is a crucial aspect of auditing the output text of AI models. Existing technologies include various methods for detecting causal relationships, primarily relying on statistical associations (such as the lower layers of Pearl's causal ladder) or domain-specific models (such as financial forecasting and agricultural optimization). These methods typically treat causal relationships as a single type (e.g., positive / negative impact), lacking a systematic classification of complex causal scenarios such as subjective expressions and preventative effects. Their credibility assessments are often based on single evidence sources or domain indicators, failing to construct a universal, hierarchically weighted quantitative system, making it difficult to conduct cross-domain, interpretable audits and verifications of causal claims from general AI outputs.

[0003] Therefore, an auditing method and system for output text is provided, which can realize multi-layered causal progressive verification, effectively verify objective deep causality, and enhance the rigor and accuracy of output text credibility assessment. Summary of the Invention

[0004] This invention provides an auditing method for output text, comprising: extracting causal chains from the output text of a text generation algorithm, wherein the causal chains include causal entities, effect entities, and relation types; performing hierarchical progressive verification on the causal chains based on multiple causal layers to determine a confidence sequence, wherein the confidence sequence includes confidence levels determined based on at least one of the multiple causal layers, wherein the multiple causal layers include at least an expression perception layer and a negative causal layer; and determining the credibility of the output text based on the confidence sequence.

[0005] This invention provides an auditing system for output text, comprising: an extraction module configured to extract causal chains from the output text of a text generation algorithm, the causal chains including causal entities, effect entities, and relation types; and a verification and determination module configured to perform hierarchical progressive verification of the causal chains based on multiple causal layers to determine a confidence sequence, the confidence sequence including confidence levels determined based on the multiple causal layers, the multiple causal layers including at least an expression perception layer and a negative causal layer; and a credibility determination module configured to determine the credibility of the output text based on the confidence sequence.

[0006] The beneficial effects of this invention include, but are not limited to: (1) proposing a systematic causal hierarchy classification framework, which, through multi-layer causal layer progressive verification, not only captures surface features but also realizes the verification of objective deep causality; (2) the mechanism gradually delves into the subjective language expression (subjective expression layer) to the preventive effect (negative causal layer), effectively eliminating false correlations, significantly enhancing the rigor and accuracy of the credibility assessment of the output text, and is conducive to judging the credibility of complex causal claim texts. Attached Figure Description

[0007] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0008] Figure 1 This is a block diagram of an auditing system for output text, as shown in some embodiments of this specification; Figure 2 This is an exemplary flowchart of an auditing method for output text according to some embodiments of this specification; Figure 3 This is an exemplary schematic diagram illustrating the hierarchical verification of a causal chain according to some embodiments of this specification; Figure 4 These are exemplary schematic diagrams illustrating the determination of subjective confidence levels according to some embodiments of this specification; Figure 5 This is an exemplary schematic diagram illustrating the determination of negative causality confidence based on some embodiments of this specification; Figure 6 This is an exemplary flowchart illustrating the determination of causal chains and path strength according to some embodiments of this specification; Figure 7 This is an exemplary schematic diagram of an initial knowledge graph according to some embodiments of this specification; Figure 8 This is an exemplary schematic diagram of a knowledge graph according to some embodiments of this specification. Detailed Implementation

[0009] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of this specification. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0010] It should be understood that the terms “system,” “device,” “unit,” and / or “module” used herein are one way to distinguish different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0011] As indicated in this specification, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0012] Flowcharts are used in this specification to illustrate the operations performed by the system according to embodiments of this specification. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0013] Figure 1 This is a block diagram of an auditing system for output text, as shown in some embodiments of this specification.

[0014] In some embodiments, such as Figure 1 As shown, the auditing system 100 for output text may include an extraction module 110, a verification and determination module 120, and a credibility determination module 130.

[0015] In some embodiments, the extraction module 110 is configured to extract causal chains from the output text of a text generation algorithm. In some embodiments, the causal chain includes a cause entity, an effect entity, and a relation type.

[0016] In some embodiments, the verification determination module 120 is configured to perform hierarchical progressive verification of the causal chain based on multiple causal layers to determine a confidence sequence. In some embodiments, the confidence sequence includes confidence levels determined based on multiple causal layers, which include at least an expression perception layer and a preventive causal layer.

[0017] In some embodiments, the confidence determination module 130 is configured to determine the confidence of the output text based on a confidence sequence.

[0018] In some embodiments, the extraction module 110, the verification and determination module 120, and the confidence determination module 130 may be integrated onto a processor. The processor may include one or more combinations of a microcontroller (MCU), an embedded processor, a graphics processing unit (GPU), etc.

[0019] For further explanation of the above content, please refer to Figure 2 Related explanations.

[0020] It should be noted that the above description of the auditing system 100 and its modules for output text is for convenience only and should not limit this specification to the scope of the embodiments described. It is understood that those skilled in the art, after understanding the principles of this system, may arbitrarily combine the various modules or construct subsystems connected to other modules without departing from these principles. In some embodiments, Figure 1 The extraction module 110, verification module 120, and credibility determination module 130 disclosed herein can be different modules within a single system, or a single module can implement the functions of two or more of the aforementioned modules. For example, the modules can share a single storage module, or each module can have its own separate storage module. Such variations are all within the scope of protection of this specification.

[0021] Figure 2 This is an exemplary flowchart of an auditing method for output text according to some embodiments of this specification. Figure 2 As shown, process 200 includes steps 210-230 as described below. In some embodiments, process 200 may be executed by a processor.

[0022] Step 210: Extract the causal chain from the output text of the text generation algorithm.

[0023] Text generation algorithms refer to computational models or systems capable of automatically generating natural language text. Text generation algorithms can be AI models or expert systems; for example, they include at least one of recurrent neural networks (RNNs) and large-scale natural language generation models based on the Transformer architecture. Expert systems can be built based on predefined rules, such as customer service robots, fault diagnosis robots, and credit approval systems.

[0024] Output text refers to the text content automatically processed and output by text generation algorithms. For example, output text includes dialogue responses, articles, or summary reports generated by large language models.

[0025] A causal chain refers to a logically related path extracted from textual information that includes a complete derivation process from cause to effect. In some embodiments, a causal chain includes a cause entity, an effect entity, and a relation type.

[0026] A causal entity refers to a concept, event, entity, or state that serves as the cause or starting point in a logical causal relationship. An effect entity refers to a concept, event, entity, or state that serves as the result or end point in a logical relationship. A relationship type refers to a classification used to describe the attributes or properties of the logical connection between cause and effect. For example, relationship types include positive causal relationships that indicate the occurrence of an event, and negative preventive causal relationships that indicate the prevention of an event. For instance, a causal chain includes a triplet logic extracted from text consisting of "taking medicine" (causal entity) leading to "symptom relief" (effect entity) and its relationship type (positive causal relationship).

[0027] In some embodiments, the processor can input the output text into a pre-trained relation extraction model, use the model to identify and classify entities in the output text, thereby obtaining causal entities, effect entities, and corresponding causal types.

[0028] In some embodiments, the processor can perform dependency parsing on the output text to identify the subject-verb-object structure; based on the subject-verb-object structure, it matches causal conjunctions and sentence pattern rule templates in a preset template to determine the relation type, and treats the phrases before and after the causal conjunctions as cause and effect entities, respectively. The preset template includes multiple causal conjunctions, sentence pattern rule templates, and corresponding relation types. The preset template can be pre-set.

[0029] In some embodiments, the processor can directly match cause entities, effect entities, and relation types in the output text based on regular expression matching methods.

[0030] In some embodiments, the processor can also extract causal chains through other feasible methods. For example, the processor can guide a large language model to directly output triples based on the output text through prompt engineering. Furthermore, the relevant extraction parameters can be adjusted according to the scenario configuration, and the data source can be either locally collected or remotely acquired.

[0031] Step 220: Based on multiple causal layers, perform hierarchical progressive verification of the causal chain to determine the confidence sequence.

[0032] A confidence sequence refers to a set of sequences composed of confidence levels assessed separately at multiple causal levels of different dimensions. In some embodiments, the confidence sequence includes confidence levels determined based on at least one of the multiple causal levels.

[0033] A causal layer refers to different dimensions or levels used to analyze, verify, and evaluate causal logic. In some embodiments, multiple causal layers include at least an expression perception layer and a negative causal layer.

[0034] The expression perception layer is a verification layer used to validate subjective expressions that demonstrate causal relationships in the output text. The confidence sequence includes subjective confidence. Subjective confidence refers to the degree of credibility of subjective expressions demonstrating causal relationships in the output text, as verified by the expression perception layer.

[0035] The negative causality layer is a verification level used to validate the existence of negative causal relationships in the output text. A negative causal relationship refers to a relationship where the cause entity has a negative impact on the effect entity, such as "infection rate decreased by 30% after vaccination" or "transmission rate decreased by 40% after wearing masks." The confidence sequence includes negative causality confidence. Negative causality confidence refers to the degree of credibility with which the negative causality layer validates the existence of negative causal relationships in the output text.

[0036] In some implementations, the causal layer can also include a non-causal layer, a statistical correlation layer, a temporal causal layer, an intervention causal layer, and a counterfactual causal layer.

[0037] The non-causal layer is a validation level used to verify that no causal relationship exists in the output text. It can validate random events where there is no statistical correlation between variables and no temporal order, such as the result of a coin toss or winning lottery numbers. The confidence sequence includes the non-causal confidence level. The non-causal confidence level refers to the degree of confidence that the non-causal layer verifies the absence of a causal relationship in the output text.

[0038] The statistical association layer is a validation layer used to verify the existence of statistical associations in the output text. For example, statements such as "ice cream sales are positively correlated with the number of drownings" and "the incidence of lung cancer is higher in smokers than in non-smokers" are statistically associated. The confidence sequence includes statistical confidence levels. Statistical confidence level refers to the degree of certainty with which the statistical association layer verifies the existence of statistical associations in the output text.

[0039] The temporal causal layer is a verification level used to validate the existence of temporal relationships in the output text. For example, "it rains first, then the ground gets wet" and "smoking first, then lung cancer" demonstrate a "cause precedes effect" temporal relationship. The confidence sequence includes temporal confidence. Temporal confidence represents the degree of credibility of the temporal causal layer's validation of the existence of temporal relationships in the output text.

[0040] The intervention causality layer is a validation layer used to verify the existence of interventional changes in the output text. An interventional change refers to an active alteration of the variable value in the causal entity, resulting in a change in the variable in the effect entity. For example, statements such as "the medication group recovered 20% faster than the control group" or "sales decreased by 15% after the price increase" demonstrate interventional changes. The confidence sequence includes the intervention confidence level. The intervention confidence level refers to the degree of credibility of the interventional change in the output text verified by the intervention causality layer.

[0041] The counterfactual causal layer is a verification level used to verify the existence of counterfactual relationships in the output text. A counterfactual relationship means that if the variable in the cause entity does not occur, then the variable in the effect entity will also not occur; that is, the cause entity is a necessary condition for the effect entity. For example, "If the patient does not smoke, he will not get lung cancer" and "If advertising is not run, product sales will not increase" contain counterfactual relationships. The confidence sequence includes the counterfactual confidence level. The counterfactual confidence level refers to the degree of credibility of the counterfactual causal layer's verification of the existence of counterfactual relationships in the output text.

[0042] Hierarchical verification refers to verifying data sequentially from the lowest to the highest level of the causal hierarchy. The levels of the causal hierarchy can be preset based on experience. Figure 3 This is an exemplary schematic diagram illustrating the hierarchical verification of a causal chain according to some embodiments of this specification. For example... Figure 3 As shown, the levels of the causal layer, from low to high, can be set as follows: no causal layer L0, expression perception layer L1, statistical association layer L2, temporal causal layer L3, intervention causal layer L4, counterfactual causal layer L5, and negative causal layer L6.

[0043] In some embodiments, the processor can verify whether the causal-free layer L0 is true; in response to the causal-free layer L0 being true, the causal-free confidence level is set to 1, and the confidence levels corresponding to higher-level causal layers are all set to 0; in response to the causal-free layer L0 not being true, the causal-free confidence level is set to 0, and then the verification of the next level of causal layer is performed. In some embodiments, the processor can determine that the causal-free layer is true based on the absence of a causal chain in the output text; otherwise, the causal-free layer is not true.

[0044] In some embodiments, the processor can perform semantic coherence detection on the output text through the expression perception layer L1 to determine the semantic distance between the causal entity and the effect entity in the context; subjective confidence is negatively correlated with semantic distance. In some embodiments, the processor can determine the logical consistency of the causal chain through the expression perception layer L1 using a local rule engine, and use the value of logical consistency as the subjective confidence. The local rule engine can be preset based on experience. In some embodiments, the expression perception layer can also be verified in other ways. For example, it can be calculated based on the perplexity of the language model.

[0045] In some embodiments, the processor can determine subjective confidence based on the confidence level of the text generation algorithm, the citation reliability of the output text, and the logical consistency of the output text through the expression perception layer. For further explanation of the above, see [link to relevant documentation]. Figure 2 Related explanations.

[0046] In some embodiments, the processor can determine the Pearson correlation coefficient through statistical software or a programming library via the statistical correlation layer L2; based on the Pearson correlation coefficient, it can determine the significance probability value (also known as the r-value) through simulated sampling; in response to the r-value being less than a first preset threshold (e.g., 0.05), it is determined that the statistical correlation layer L2 is valid, and the square of the Pearson correlation coefficient is used as the statistical confidence level, and then proceeds to the next level of causal layer; otherwise, it is determined that the statistical correlation layer L2 is invalid, and the statistical confidence level and the confidence level corresponding to the higher-level causal layer are both set to 0.

[0047] In some embodiments, the processor can take the p-value of the Granger causality test through the temporal causality layer L3; in response to the occurrence time of the causal entity being earlier than the occurrence event of the effect entity, and the p-value being less than a second preset threshold (e.g., 0.05), the processor determines that the temporal causality layer L3 is valid, and uses the difference between 1 and the p-value as the temporal confidence level, and then proceeds to the next level of causality layer; otherwise, the processor determines that the temporal causality layer L3 is invalid, and sets the temporal confidence level and the confidence level corresponding to the higher-level causality layer to 0.

[0048] In some embodiments, the processor can simulate multiple randomized controlled experiments by intervening in causal layer L4; in response to the existence of an intervention effect in multiple randomized controlled experiments, it determines that the intervention causal layer L4 is valid, takes the Cohen effect size (i.e., Cohen's d effect size) and normalizes it, uses the normalized Cohen effect size as the intervention confidence level, and proceeds to the next level of causal layer; otherwise, it determines that the intervention causal layer L4 is invalid, and sets the intervention confidence level and the confidence level corresponding to the higher-level causal layer to 0.

[0049] In some embodiments, the processor can determine the probability of necessity (PN) through the counterfactual causal layer L5; in response to PN being greater than or equal to a third preset threshold (e.g., 0.5), the processor determines that the counterfactual causal layer L5 is valid, and uses the value of PN as the negative causal confidence level, and then proceeds to the next level of causal layer; otherwise, the processor determines that the counterfactual causal layer L5 is invalid, and sets the negative causal confidence level and the confidence level corresponding to the higher-level causal layer to 0.

[0050] In some embodiments, the processor can retrieve a negative causal table based on causal chains through a negative causal layer L6. If no negative causal chain exists with a similarity greater than a fourth preset threshold, the negative causal layer is deemed invalid, and a negative causal confidence score of 0 is output. Conversely, if a negative causal layer exists, the negative causal layer is deemed valid, and the maximum similarity score is used as the negative causal confidence score. The negative causal table includes multiple negative causal chains, each consisting of a cause entity and an effect entity with a negative causal relationship. The negative causal table can be preset based on experience. In some embodiments, the negative causal layer can also be verified in other ways. For example, it can be implemented based on causal intervention statistical methods; or, for example, based on a preset logical contradiction rule template.

[0051] In some embodiments, the processor can also determine the negative causality probability of a causal chain through a negative causality layer; and determine the negative causality confidence level based on the negative causality probability of the causal chain. For further explanation of the above, see [link to relevant documentation]. Figure 3 Related explanations.

[0052] Step 230: Determine the confidence level of the output text based on the confidence sequence.

[0053] Credibility refers to a comprehensive evaluation index of the authenticity, rationality, and reliability of the output text generated by a text generation algorithm.

[0054] In some embodiments, the processor can obtain the confidence level of the output text based on a weighted sum of the confidence levels in the confidence sequence. The weights corresponding to each causal layer can be preset based on experience. For example, the processor can assign lower weights to causal layers with lower mapping levels and higher weights to causal layers with higher mapping levels.

[0055] In some embodiments, credibility can also be determined in other ways. For example, by fitting multi-dimensional scores based on statistical methods or physical models. In some embodiments, the weights corresponding to each causal layer can be adjusted according to the scenario configuration; computational tasks can also be assigned to cloud or local nodes for execution.

[0056] In some embodiments, the processor may determine the confidence level of the output text based on a confidence sequence, the inherent reliability of multiple causal layers, and the weights of multiple causal layers.

[0057] Inherent reliability refers to the inherent reliability or basic weight of a causal layer in scientific argumentation. The inherent reliability and weights of multiple causal layers can be preset based on experience. In some embodiments, the inherent reliability of a causal layer is positively correlated with its level. For example, the inherent reliability of a no-causal layer L0 can be set to 0, the inherent reliability of an expression-perception layer L1 can be set to 0.5, the inherent reliability of a statistical correlation layer L2 can be set to 0.7, the inherent reliability of a temporal causal layer L3 can be set to 0.8, the inherent reliability of an interventional causal layer L4 can be set to 0.9, the inherent reliability of a counterfactual causal layer L5 can be set to 0.9, and the inherent reliability of a negative causal layer L6 can be set to 0.95. In some embodiments, the weight of a causal layer is positively correlated with its level.

[0058] In some embodiments, the processor determines that the confidence level of the output text is positively correlated with the confidence level value, inherent reliability, and weight of each causal layer in the confidence level sequence. For example, the processor can calculate the confidence level of the output text using formula (1):

[0059] in, Indicates the credibility of the output text; Indicates the number of causal layers; , , Representing the causal layer respectively The corresponding weights, confidence levels, and inherent reliability, where k ranges from 1 to n. This is the weight decay factor. 0.8 can be taken.

[0060] In some embodiments of this specification, credibility is calculated by comprehensively considering confidence sequences, inherent reliability, and hierarchical weights, which effectively reduces the interference of pseudo-associations in low-quality data and significantly improves the scientificity and accuracy of text credibility assessment.

[0061] Some embodiments in this specification employ a multi-layered causal structure to progressively verify the causal chain. This mechanism delves deeper step by step, from subjective language expression to intervention mechanisms and preventative effects, not only capturing surface features but also verifying objective, deep-seated causality. This effectively eliminates spurious correlations and significantly enhances the rigor and accuracy of the output text credibility assessment.

[0062] It should be noted that the above description of process 200 is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to process 200 under the guidance of this specification. However, these modifications and changes remain within the scope of this specification.

[0063] Figure 4 This is an exemplary schematic diagram illustrating the determination of subjective confidence levels according to some embodiments of this specification. In some embodiments, such as Figure 4 As shown, the processor can determine subjective confidence based on the confidence of the text generation algorithm, the citation reliability of the output text, and the logical consistency of the output text through the expression perception layer.

[0064] The confidence level of a text generation algorithm refers to the reliability of its output. Citation reliability refers to the degree to which external sources, documents, or bases used in the information are authentic, accurate, and verifiable. For example, citation reliability refers to the authenticity and authority of the references cited in the output text. Logical consistency refers to the degree to which statements, reasoning steps, and conclusions within the output text are coherent and consistent, without contradictions.

[0065] The confidence score of a text generation algorithm can be the average of its token confidence scores. Token confidence score represents the degree of certainty with which the text generation algorithm predicts each token. In some embodiments, the processor can directly obtain the confidence scores of multiple tokens output by the text generation algorithm; the average of these multiple token confidence scores is used as the overall confidence score of the text generation algorithm. Citation reliability is positively correlated with the number of citations from authoritative sources (such as journal articles, industry standards, official reports, etc.). For example, each citation from an authoritative source adds 0.1 points, with an upper limit of 0.3 points for citation reliability. Logical consistency can be determined using a Natural Language Inference (NLI) model. For example, for each output text, if the NLI model verifies contextual consistency, the logical consistency value is increased by 0.2; otherwise, it is decreased by 0.2.

[0066] In some embodiments, the processor may normalize the confidence of the text generation algorithm, the citation reliability of the output text, and the logical consistency, and use the sum of the normalized confidence of the text generation algorithm, the citation reliability of the output text, and the logical consistency as the subjective confidence.

[0067] In some embodiments, the determination of subjective confidence levels can also be achieved in other ways. For example, it can be achieved based on statistical methods such as logistic regression models.

[0068] In some embodiments of this specification, algorithmic confidence, citation reliability, and logical consistency are fused at the expression perception layer to determine subjective confidence. This mechanism quantifies and evaluates text quality from multiple dimensions, enabling early filtering and reliable transmission, laying the foundation for multi-layered causal verification, and thus improving the verification accuracy and robustness of the framework.

[0069] Figure 5This is an exemplary schematic diagram illustrating the determination of negative causality confidence based on some embodiments of this specification. In some embodiments, the processor may determine the negative causality probability of a causal chain through a negative causality layer; and determine the negative causality confidence based on the negative causality probability of the causal chain.

[0070] Negative causal probability refers to the probability that a negative causal relationship holds true in the output text.

[0071] In some embodiments, the processor may calculate the probability of an effect entity occurring when it is not exposed to the causal entity (denoted as the first probability) and the probability of an effect entity occurring when it is exposed to the causal entity (denoted as the second probability); determine the negative causal probability based on the first probability and the second probability; and, in response to the negative causal probability being greater than a fifth preset threshold (e.g., 0.05), use the negative causal probability as the negative causal confidence level; otherwise, set the negative causal confidence level to 0.

[0072] In some embodiments, the first probability is the ratio of the number of samples in the unexposed group that developed the disease to the total number of samples in the unexposed group, where the unexposed group is a multi-group randomized controlled trial conducted without exposure to the causal agent; the second probability is the ratio of the number of samples in the exposed group that developed the disease to the total number of samples in the exposed group, where the exposed group is a multi-group randomized controlled trial conducted with exposure to the causal agent. In some embodiments, the processor can use the difference between the first and second probabilities as the negative causal probability. For example, for the negative causal relationship "infection rate decreased by 30% after vaccination", the unexposed group is the unvaccinated samples, the first probability is the ratio of the number of unvaccinated samples that developed infection to the total number of unvaccinated samples; the exposed group is the vaccinated samples, and the second probability is the ratio of the number of vaccinated samples that developed infection to the total number of vaccinated samples.

[0073] In some embodiments, the probability of negative causality can also be determined in other ways. For example, it can be obtained by looking up historical statistical data in a table; or, for example, by prediction based on a deep learning network.

[0074] In some embodiments of this specification, negative causal confidence is calculated by negative causal probability, transforming static text into a dynamic intervention evaluation mechanism, effectively filtering false prevention relationships, and outputting quantitative indicators, thereby improving the depth and accuracy of causal chain credibility evaluation.

[0075] In some embodiments, the processor may determine the domain to which the output text belongs based on a preset algorithm; and determine at least one confidence level through at least one causal layer based on the domain to which the output text belongs.

[0076] In some embodiments, the processor performs domain identification based on a preset algorithm to determine the domain to which the output text belongs.

[0077] The domain to which the output text belongs refers to the knowledge or application category tag of the output text, such as the medical field, the financial field, the consumer field, the legal field, etc.

[0078] In some embodiments, the processor can input the output text into a domain classifier to obtain the domain to which the output text belongs. The domain classifier can be a model based on a deep learning architecture such as BERT.

[0079] In some embodiments, a domain classifier can be obtained through training. The training samples for the domain classifier include a large number of texts from various domains, with the training labels corresponding to the domain.

[0080] In some embodiments, determining the domain to which the output text belongs can also be achieved in other ways. For example, classification can be performed based on machine learning models of other natural language processing methods.

[0081] In some embodiments, the processor may pre-build a domain knowledge base containing multiple pre-defined mapping functions for core domains; load the mapping function corresponding to the domain to which the output text belongs from the domain knowledge base; and determine at least one confidence level based on the mapping function. The mapping function may be pre-defined.

[0082] In some embodiments, for the medical field, the mapping function for the statistical correlation layer L2 is "e2=min(1, r2×1.2)" and the mapping function for the intervention causality layer L4 is "e4=min(1, d×1.5)", where e2 and e4 represent the statistical confidence level and the intervention confidence level, respectively, r represents the Pearson correlation coefficient, and d represents Cohen's d effect size. For information on the Pearson correlation coefficient and Cohen's d effect size, see [link to relevant documentation]. Figure 2 Step 220 related explanation. In some embodiments, for the medical field, the processor can increase the weights corresponding to the statistical correlation layer L2. Weights corresponding to the intervention causal layer L4 .

[0083] In some embodiments, specifically for the financial field, the processor can lower a second preset threshold (e.g., from 0.05 to 0.025). In response to a significance probability value less than the second preset threshold, it is determined that the temporal causal layer L3 is valid; the mapping function of the temporal causal layer L3 is "e3=1". The mapping function for the counterfactual causal layer L5, p×2", is e5=PN×1.1, where e3 and e5 represent the time series confidence and counterfactual confidence, respectively, p represents the significance probability value, and PN represents the necessary probability. For more information on significance probability and necessary probability, please refer to [link to relevant documentation]. Figure 2 Step 220 related explanation. In some embodiments, for the financial field, the processor can increase the weight corresponding to the counterfactual causality layer L5. .

[0084] In some embodiments, for the consumer domain, the mapping function of the expression perception layer L1 is to increase the upper limit of the citation reliability of the output text (e.g., to 0.4), and the mapping function of the statistical association layer L2 is "e2=r 2 ", where r represents the Pearson correlation coefficient and e² represents the statistical confidence level. For information on the reliability of the output text citations, please refer to..." Figure 3 Related explanations.

[0085] In some embodiments, specifically in the legal field, when calculating citation reliability through the expression perception layer L1, only legal provisions and judicial interpretations are considered authoritative sources. Citation reliability is positively correlated with the number of authoritative sources; for example, adding one authoritative source increases citation reliability by 0.2, and the upper limit of citation reliability is set to 0.4. When determining the intervention confidence level through the intervention causality layer L4, randomized controlled trials select only judicial precedents.

[0086] In some embodiments, the mapping functions and related parameters in the domain knowledge base can be added or adjusted according to the scenario configuration; the data source can also be local collection or remote acquisition.

[0087] In some embodiments of this specification, the confidence level is determined by identifying the domain to which the text belongs and calling the corresponding mapping rules accordingly. This enables the confidence level calculation to dynamically adapt to the preference mechanisms of different domains for evidence strength, thereby improving the accuracy of confidence level assessment and enhancing the versatility and practicality of the solution in different vertical scenarios.

[0088] In some embodiments, the processor can determine the social consensus rate of the output text based on the confidence sequence, the user's individual cognitive level, cultural adaptation factors, and social influence weights.

[0089] An individual's cognitive level refers to the level of causality that an individual can understand and accept. This level can be determined by consulting a pre-defined cognitive table. The pre-defined cognitive table includes various individual types and their corresponding cognitive levels. Individual type refers to the user type used in the text generation algorithm. The processor can automatically match individual types based on user registration information, professional certifications, questionnaires, etc.

[0090] As an example only, the preset cognitive table may include: the individual cognitive level of the general public can be set to 2, indicating that the general public only accepts causal layers at statistical correlation level L2 and lower; the individual cognitive level of industry practitioners can be set to 3, indicating that industry practitioners only accept causal layers at statistical correlation level L3 and lower; the individual cognitive level of domain engineers or analysts can be set to 4, indicating that domain engineers or analysts only accept causal layers at statistical correlation level L4 and lower; the individual cognitive level of domain experts or researchers can be set to 5, indicating that domain experts or researchers only accept causal layers at statistical correlation level L5 and lower; and the individual cognitive level of decision-makers or policymakers can be set to 6, indicating that the general public accepts causal layers at statistical correlation level L6 and lower (i.e., accepts all levels of causal layers).

[0091] The cultural fit factor is a quantitative value reflecting the bias in the understanding of causal expressions by different individual types based on their cultural background. In some embodiments, the cultural fit factor can be set to (0.8, 1.2), and the processor can calibrate the cultural fit factor based on the Power Distance Index (PDI) of the Hofstede cultural dimension model. The PDI measures the degree of acceptance and expectation of unequal power distribution among members of a society with less power. The PDI for different individual types can be pre-set based on experience or obtained through questionnaire surveys. The cultural fit factor is positively correlated with the PDI. For example, in response to a PDI greater than the sixth pre-set threshold (e.g., 70), the cultural fit factor is set to 1.1; in response to a PDI less than the seventh pre-set threshold (e.g., 50), the cultural fit factor is set to 0.9. If the cultural background is unclear, the cultural fit factor is set to 1.0.

[0092] Social influence weight refers to a quantitative value reflecting the influence of different individual types within a group, ranging from 0 to 1. The processor can determine social influence weight based on factors such as expert title, user authentication level, and historical decision accuracy. For example, the social influence weight of a leading expert in the field can be 0.6 to 0.7; the social influence weight of an industry practitioner can be 0.2 to 0.3; and the social influence weight of the general public can be 0.1 to 0.2.

[0093] Social consensus rate refers to the degree to which an output text is generally accepted and recognized by the general public (including all types of individuals). In some embodiments, the processor can determine the social consensus rate based on a confidence sequence, the user's individual cognitive level, cultural adaptation factors, and social influence weights, using a preset algorithm.

[0094] In some embodiments, the preset algorithm can be expressed as formula (2) and formula (3):

[0095]

[0096] in, Indicates the rate of social consensus; This indicates the user's individual cognitive level; Indicates cultural adaptation factor; Indicates the weight of social influence; Representing causal layer The corresponding weights; Representing causal layer Its inherent reliability. Regarding... and See also Figure 2 Related explanations.

[0097] In some embodiments of this specification, the social consensus rate is calculated by comprehensively considering individual cognitive levels, cultural factors, and influence weights. By utilizing a calculation mechanism that quantifies multidimensional user differences, an objective quantitative comparison of different texts is achieved, providing a reliable scientific basis for the final decision.

[0098] In some embodiments, the text generation algorithm includes multiple sub-algorithms. These sub-algorithms are one of the following: recurrent neural networks (RNNs), large-scale natural language generation models based on the Transformer architecture, expert systems, etc. For further explanation of the above, please refer to [link to relevant documentation]. Figure 2 .

[0099] Figure 6 This is an exemplary flowchart illustrating the determination of causal chains and path strength according to some embodiments of this specification. Figure 6 As shown, process 600 includes steps 610-630 as described below. In some embodiments, process 600 may be executed by a processor.

[0100] Step 610: Extract multiple sub-causal chains from the output text of multiple sub-algorithms.

[0101] A sub-causal chain refers to a causal chain extracted from the output text of a single sub-algorithm model.

[0102] The method for extracting sub-causal chains is similar to that for extracting causal chains. For information on causal chain extraction methods, please refer to [link to relevant documentation]. Figure 2 Related explanations.

[0103] Step 620: Construct a knowledge graph based on multiple sub-causal chains.

[0104] A knowledge graph is a structured semantic knowledge base that expresses the relationships between cause entities and effect entities in the form of a graph. In some embodiments, the nodes of a knowledge graph include cause entities and effect entities, and the edges of the knowledge graph represent the relationships between cause entities and effect entities. The relationship between cause entities and effect entities can be that the cause entity leads to the occurrence of the effect entity, and the direction of the edge is the direction from the cause entity to the effect entity.

[0105] In some embodiments, the processor can acquire effect entities and cause entities from multiple sub-causal chains, align effect entities and / or cause entities using entity alignment techniques (such as node matching based on semantic similarity); merge identical effect entities and / or cause entities into a single node, and connect cause entities and effect entities to form edges, thereby obtaining an initial knowledge graph (such as...). Figure 7 As shown); based on the initial knowledge graph, calculate the confidence of each edge in the initial knowledge graph. In response to an edge having a confidence of less than the eighth preset threshold (e.g., 0.5), delete that edge (i.e., retain edges with a confidence greater than the eighth preset threshold), and delete nodes not connected by edges to obtain the knowledge graph (e.g., ...). Figure 8 (As shown). A node can be either an effect entity or a cause entity. The same node can also be both an effect entity and a cause entity.

[0106] Figure 7 This is an exemplary schematic diagram of an initial knowledge graph according to some embodiments of this specification. Figure 7 and Figure 8 For example, see the illustration. Figure 7 As shown, the example contains three sub-causal chains: the first sub-causal chain extracted by the first sub-algorithm (including cause entity 1 and effect entity 1), the second sub-causal chain extracted by the second sub-algorithm (including cause entity 2 and effect entity 2), and the third sub-causal chain extracted by the third sub-algorithm (including cause entity 3 and effect entity 3). Using entity alignment technology, all cause entities and / or effect entities are aligned. The effect entity 1 of the first sub-causal chain and the cause entity 2 of the second sub-causal chain are identical, so they are merged into one node; the effect entity 2 of the second sub-causal chain and the effect entity 3 of the third sub-causal chain are identical, so they are merged into one node. After merging, the knowledge graph includes 4 nodes; connecting the cause entities and effect entities forms a structure as shown below. Figure 7 The knowledge graph shown.

[0107] In some embodiments, the processor can determine the confidence level of an edge based on the weight of the sub-algorithm corresponding to the edge and the confidence level of the sub-algorithm. For example, the processor can determine the confidence level of an edge using formula (4):

[0108] in, Representing an edge Confidence level; Representing an edge The number of sub-algorithms; Representing an edge The weights of the corresponding sub-algorithms; Representing an edge The confidence level of the corresponding sub-algorithm. The weight of a sub-algorithm refers to its weight among at least one sub-algorithm corresponding to this edge. The weight of a sub-algorithm can be set based on its historical accuracy (e.g., if sub-algorithm 1 has an accuracy of 90%, then its weight is 0.9). The confidence level of the sub-algorithm. This can be the confidence score output by the sub-algorithm itself. When the sub-algorithm does not output a confidence score, if there are edges in the sub-algorithm's output... ,but Set to 1; if no edge exists in the output of the sub-algorithm. ,but Set to 0. For example, such as... Figure 7 As shown, the edge The sub-algorithm corresponding to 1 includes sub-algorithm 1 and sub-algorithm 2, then Set to 2.

[0109] Figure 8 This is an exemplary schematic diagram of a knowledge graph according to some embodiments of this specification. For example, Figure 7 and Figure 8 As shown, if the edges with a confidence level greater than the eighth preset threshold include S1, S2, S4, and S5, then the corresponding paths (i.e., causal chains) include path S1, path S2, path S1-S4, path S2-S4, and path S5.

[0110] Step 630: Based on the knowledge graph, determine at least one causal chain and its corresponding path strength.

[0111] A path is a logical route from a cause entity to a result entity. A path may include at least one edge. Path strength is a quantitative indicator of the reliability of a path.

[0112] In some embodiments, the processor may, based on a knowledge graph, treat at least one path from a cause entity to an effect entity as at least one causal chain; and use the statistical value (such as average, product, minimum, etc.) of the confidence of at least one edge in each path as the path strength.

[0113] In some embodiments, a set of cause and effect entities may include at least one path; for each set of cause and effect entities, the processor may calculate the credibility corresponding to at least one edge in each path, and take the average credibility corresponding to the at least one edge as the credibility corresponding to that path; the credibility of the output text is obtained by weighted summation of the credibility of the at least one path corresponding to the set of cause and effect entities; the weights are positively correlated with the path strength. For example, as shown... Figure 8As shown, cause entity 1 and effect entity 2 include path S5 and path S1-S4; the credibility of path S5 is the credibility of edge S5 (cause entity 1 - effect entity 2); the credibility of the causal chain S1-S4 is the average of the credibility of edge S1 (cause entity 1 - effect entity 1) and the credibility of edge S4 (cause entity 2 - effect entity 2); the final credibility of cause entity 1 and effect entity 2 is the weighted sum of the credibility of path S5 and the credibility of path S1-S4. For details on credibility calculation, please refer to [link to relevant documentation]. Figure 2 Related explanations.

[0114] In some embodiments of this specification, sub-causal chains are extracted by fusing the outputs of multiple sub-algorithms and constructing a causal knowledge graph, and the consensus level of multiple models is quantified using path strength. This mechanism effectively integrates the advantages of different models, significantly improves the coverage and robustness of causal chains, and thus provides a more reliable evaluation basis for the final text credibility.

[0115] In some embodiments, the processor can build a pollution pattern library; based on the causal chain, search the pollution pattern library to determine the vulnerability type of the causal chain.

[0116] The Pollution Pattern Library is a database reflecting various vulnerability types. It includes pattern feature vectors and corresponding vulnerability types. A vulnerability type refers to a logical flaw or error category present in the causal chain or output text. Vulnerability types include causal inversion, uncontrolled promiscuity, misuse of negative causality, survivorship bias, spurious correlation, appeal to authority, slippery slope fallacy, and sample bias. A pattern feature vector is a vector composed of sample causal entities and sample effect entities.

[0117] In some implementations, the processor can extract the cause and effect entities of multiple vulnerable sample texts using a Sentence-BERT pre-trained model, generating multiple sets of pattern feature vectors. The vector dimension is preset based on experience (e.g., set to 768). The label for each set of pattern feature vectors is the vulnerability type. The contaminated pattern library includes a similarity threshold corresponding to each vulnerability type. The similarity threshold reflects the minimum critical value of similarity between the vector to be matched and the pattern feature vector. Similarity is negatively correlated with vector distance. The similarity threshold can be preset based on experience or requirements, and the similarity threshold varies for each vulnerability type.

[0118] In some embodiments, the processor may also construct a pollution pattern library in other feasible ways. For example, a pattern mapping table containing various logical fallacy feature rules may be constructed based on an expert knowledge base.

[0119] In some embodiments, the processor can concatenate causal chains into natural language text, extract the natural language text using a Sentence-BERT pre-trained model to obtain a vector to be matched, the vector dimension of the vector to be matched being consistent with the vector dimension of the pattern feature vector; search a pollution pattern library based on the vector to be matched, determine the similarity between the vector to be matched and the pattern feature vector, and take the vulnerability type corresponding to the pattern feature vector with a similarity greater than the similarity threshold as the vulnerability type of the causal chain corresponding to the vector to be matched.

[0120] Some embodiments in this specification construct a pollution pattern library and perform searches based on causal chains, enabling accurate identification of logical vulnerability types through feature vector similarity calculations. This mechanism combines multi-model fusion with progressive verification rules, significantly reducing the false positive rate of vulnerabilities and effectively enhancing the reliability of the final audit results.

[0121] In some embodiments, the processor may determine a confidence range sequence based on interlayer parameters of multiple causal layers; and determine the confidence range of the output text based on the confidence range sequence of multiple causal layers.

[0122] Interlayer parameters refer to specific indicators used to determine confidence levels during the validation process of causal layers. For example, interlayer parameters include the Pearson correlation coefficient for statistical correlation layer L2, Cohen's d effect size for interventional causality layer L4, and the necessary probability for counterfactual causality layer L5.

[0123] The confidence range sequence is a set of fluctuation intervals of confidence based on multiple causal layers.

[0124] In some embodiments, for each causal layer, the processor can determine the confidence range of that causal layer based on the error of its inter-layer parameters. The upper limit of the confidence range is the sum of the confidence level and the error, and the lower limit is the difference between the confidence level and the error. The error can be preset empirically and directly output through the causal layer. See [link to documentation] for the method of determining the confidence level. Figures 2-5 Related explanations.

[0125] In some embodiments, the processor can also derive the confidence intervals of each causal layer based on resampling techniques. For example, the output text is generated multiple times using a text generation algorithm, and the inter-layer parameters are obtained during the multiple generation processes. The confidence is calculated based on the maximum and minimum values ​​of the inter-layer parameters to obtain the upper and lower limits of the confidence range.

[0126] In some embodiments, determining the confidence range sequence can also be achieved in other feasible ways, such as determining the upper and lower limits of the confidence range based on a machine learning model.

[0127] The credibility range is the range of values ​​for the credibility of the output text.

[0128] In some embodiments, the processor can calculate a weighted average of the confidence ranges for each causal layer. For example, the lower bound of the confidence scores for each layer can be summed using weighted averages to obtain the lower bound of the confidence score, and the upper bound of the confidence scores for each layer can be summed using weighted averages to obtain the upper bound of the confidence score. The weight settings can be found in [reference needed]. Figure 2 Step 230 related instructions.

[0129] In some embodiments, the processor can substitute the upper limit of the confidence range into formula (1) to calculate the upper limit of the confidence of the output text, and substitute the lower limit of the confidence range into formula (1) to calculate the lower limit of the confidence of the output text; and determine the confidence range of the output text based on the upper limit and lower limit of the confidence of the output text.

[0130] In some embodiments, the processor may also determine the confidence range of the output text through other feasible methods. For example, by using Monte Carlo simulation, values ​​are randomly sampled within the confidence range of each causal layer and fused together to obtain the probability interval of confidence as the confidence range.

[0131] Some embodiments in this specification determine the confidence range by introducing interlayer parameters and confidence range sequences. This mechanism can reflect the fluctuation edges of multi-causal layer data. This more realistically reflects the quality of evidence, and is particularly suitable for scenarios with small sample sizes or high data noise, significantly enhancing the reliability of the evaluation results.

[0132] The basic concepts have been described above. Obviously, for those skilled in the art, the detailed disclosure above is merely illustrative and does not constitute a limitation of this specification. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this specification. Such modifications, improvements, and corrections are suggested in this specification and therefore remain within the spirit and scope of the exemplary embodiments described herein.

[0133] Furthermore, this specification uses specific terms to describe embodiments thereof. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic associated with at least one embodiment of this specification. Therefore, it should be emphasized and noted that references to "an embodiment," "one embodiment," or "an alternative embodiment" in different locations throughout this specification do not necessarily refer to the same embodiment. Moreover, certain features, structures, or characteristics in one or more embodiments of this specification can be appropriately combined.

[0134] Furthermore, the order of processing elements and sequences, the use of numbers and letters, or other names described in this specification are not intended to limit the order of the processes and methods described herein. Although various examples have been discussed in the foregoing disclosure of some embodiments of the invention that are currently considered useful, it should be understood that such details are for illustrative purposes only. For example, while the system components described above can be implemented by hardware devices, they can also be implemented solely by software solutions, such as installing the described system on existing servers or mobile devices.

[0135] Similarly, it should be noted that, in order to simplify the description disclosed in this specification and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of the embodiments in this specification may sometimes combine multiple features into a single embodiment, drawing, or description thereof. In fact, the embodiments have fewer features than all the features of the single embodiment disclosed above.

[0136] In some embodiments, numbers describing the quantity of components and attributes are used. It should be understood that such numbers used in the description of embodiments are modified in some examples with the terms "approximately," "approximately," or "generally." Unless otherwise stated, "approximately," "approximately," or "generally" indicates that the numbers are allowed to vary by ±20%. Accordingly, in some embodiments, the numerical parameters used in the specification are approximate values, which may be changed depending on the characteristics required by individual embodiments. In some embodiments, numerical parameters should take into account specified significant digits and employ a general method of digit reservation. Although the numerical ranges and parameters used to confirm their breadth of range in some embodiments of this specification are approximate values, in specific embodiments, such values ​​are set as precisely as feasible.

[0137] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and should be considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.

Claims

1. An audit method of outputting text, characterized by, include: Extract causal chains from the output text of a text generation algorithm, wherein the causal chains include cause entities, effect entities, and relation types; Based on multiple causal layers, the causal chain is verified hierarchically to determine a confidence sequence. The confidence sequence includes a confidence level determined based on at least one of the multiple causal layers, which includes at least an expression perception layer and a negative causal layer. The confidence level of the output text is determined based on the confidence sequence.

2. The method according to claim 1, characterized in that, The step of performing hierarchical progressive verification of the causal chain based on multiple causal layers to determine the confidence sequence includes: The subjective confidence level is determined through the expression perception layer based on the confidence level of the text generation algorithm, the citation reliability of the output text, and the logical consistency of the output text.

3. The method according to claim 1, characterized in that, The step of performing hierarchical progressive verification of the causal chain based on multiple causal layers to determine the confidence sequence includes: The negative causal probability of the causal chain is determined through the negative causal layer. The negative causal confidence level is determined based on the negative causal probability of the causal chain.

4. The method according to claim 1, characterized in that, The method further includes: Based on the confidence sequence, the user's individual cognitive level, cultural adaptation factor, and social influence weight, the social consensus rate of the output text is determined.

5. The method according to claim 1, characterized in that, The text generation algorithm includes multiple sub-algorithms, and the method further includes: Extract multiple sub-causal chains from the output text of the various sub-algorithms; Based on the multiple sub-causal chains, a knowledge graph is constructed. The nodes of the knowledge graph include causal entities and effect entities, and the edges of the knowledge graph represent the relationship between causal entities and effect entities. Based on the knowledge graph, at least one causal chain and its corresponding path strength are determined.

6. The method according to claim 1, characterized in that, The method further includes: Construct a pollution pattern library; Based on the causal chain, the pollution pattern library is searched to determine the vulnerability type of the causal chain.

7. The method according to claim 1, characterized in that, The step of performing hierarchical progressive verification on the causal chain to determine the confidence sequence includes: The domain to which the output text belongs is determined based on a preset algorithm; Based on the domain to which the output text belongs, at least one confidence level is determined through the at least one causal layer.

8. The method according to claim 1, characterized in that, The method further includes: Based on the interlayer parameters of the multiple causal layers, a confidence range sequence is determined; The confidence range of the output text is determined based on the confidence range sequence of the multiple causal layers.

9. The method according to claim 1, characterized in that, Determining the credibility of the output text based on the confidence sequence includes: The confidence level of the output text is determined based on the confidence sequence, the inherent reliability of the plurality of causal layers, and the weights of the plurality of causal layers.

10. An auditing system for output text, characterized in that, include: The extraction module is configured to extract causal chains from the output text of a text generation algorithm, the causal chains including causal entities, effect entities, and relation types; The verification and determination module is configured to perform hierarchical progressive verification of the causal chain based on multiple causal layers to determine a confidence sequence, wherein the confidence sequence includes confidence levels determined based on the multiple causal layers, and the multiple causal layers include at least an expression perception layer and a negative causal layer; A confidence determination module is configured to determine the confidence of the output text based on the confidence sequence.

Images