Knowledge-enhanced medical illusion static detection and correction method and system

Abstract

The invention provides a knowledge-enhanced medical illusion static detection and correction method and system, and relates to the technical field of medical illusion detection.The method comprises the steps that first feature information is obtained; determining a causal association strength coefficient; constructing a dynamic knowledge element network; obtaining first evidence information according to the first feature information and the dynamic knowledge element network; obtaining a first consistency coefficient; determining a timeliness coefficient and a confidence coefficient of the first evidence information; processing the causal association strength coefficient, the first consistency coefficient, the timeliness coefficient and the confidence coefficient through a trained medical illusion detection model to generate a medical illusion detection probability; determining whether the first feature information needs to be corrected; and when it is determined that the first feature information needs to be corrected, correcting the first feature information through a trained medical illusion correction model. According to the invention, the accuracy of medical illusion static detection and correction can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical hallucination detection, and particularly relates to a knowledge-enhanced medical hallucination static detection and correction method and system. BACKGROUND

[0002] In the related art, the static detection and correction of medical hallucinations can be realized by means of artificial detection and correction by professional technicians, that is, mainly relying on human factors, and the workload of the detection work is very large, therefore, excessive reliance on human factors may not guarantee the accuracy of the detection and correction results, and may not guarantee the efficiency of the detection and correction.

[0003] The information disclosed in the background section of this application is only intended to deepen the understanding of the general background of the application and should not be regarded as acknowledging or implying in any form that this information constitutes prior art known to those skilled in the art. SUMMARY

[0004] The present application provides a knowledge-enhanced medical hallucination static detection and correction method and system, which can solve the technical problems that the related art cannot guarantee the accuracy of the detection and correction results, and cannot guarantee the efficiency of the detection and correction.

[0005] According to a first aspect of the present application, a knowledge-enhanced medical hallucination static detection and correction method is provided, comprising: processing the to-be-detected text by a feature extraction model to obtain first feature information; determining a causal correlation strength coefficient according to the first feature information; constructing a dynamic knowledge element network; obtaining first evidence information according to the first feature information and the dynamic knowledge element network; processing the first feature information and the first evidence information by a knowledge-enhanced NLI model to obtain a first consistency coefficient; determining a timeliness coefficient and a confidence coefficient of the first evidence information according to the dynamic knowledge element network; processing the causal correlation strength coefficient, the first consistency coefficient, the timeliness coefficient and the confidence coefficient by a trained medical hallucination detection model to generate a medical hallucination detection probability; determining whether the first feature information needs to be corrected according to the medical hallucination detection probability; in the case of determining that correction is needed, correcting the first feature information by a trained medical hallucination correction model.

[0006] According to the present application, the causal correlation strength coefficient is determined according to the first feature information, comprising: According to the first feature information, a first causality verb is acquired; According to the first causality verb, a strong causality verb recognition result, a relevance verb recognition result and a fact verb recognition result are determined; According to the strong causality verb recognition result, the relevance verb recognition result and the fact verb recognition result, a causality correlation strength coefficient is determined.

[0007] According to the application, according to the dynamic knowledge element network, a timeliness coefficient and a confidence coefficient of first evidence information are determined, comprising: According to the dynamic knowledge element network, a last update time, an information source type and information research sample data of first evidence information are acquired; According to the last update time, a timeliness coefficient is determined; According to the information source type and the information research sample data, a confidence coefficient is determined.

[0008] According to the application, according to the information source type and the information research sample data, a confidence coefficient is determined, comprising: According to the information source type, a highest confidence recognition result, a high confidence recognition result, a medium confidence recognition result and a low confidence recognition result are determined; According to the highest confidence recognition result, the high confidence recognition result, the medium confidence recognition result and the low confidence recognition result, an information source confidence coefficient is determined; According to the information research sample data, a gender sample recognition result and an age sample recognition result are determined; According to the gender sample recognition result and the age sample recognition result, an information sample confidence coefficient is determined; According to the information source confidence coefficient and the information sample confidence coefficient, a confidence coefficient is determined.

[0009] According to the application, the training step of the medical hallucination detection model comprises: A training sample data set is acquired; According to the training sample data set, sample feature information and sample evidence information of sample feature information are acquired; A sample causality correlation strength coefficient of the sample feature information is determined; A sample consistency coefficient of the sample feature information and the sample evidence information is determined; A sample timeliness coefficient and a sample confidence coefficient of sample evidence information are determined; According to the medical hallucination detection model, the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient are processed to determine the sample medical hallucination detection probability of the sample feature information; An actual medical hallucination detection result of the sample feature information is obtained. According to the actual medical hallucination detection result, the sample medical hallucination detection probability, the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient, a training loss function of the medical hallucination detection model is determined. According to the training loss function of the medical hallucination detection model, the medical hallucination detection model is trained to obtain a trained medical hallucination detection model.

[0010] According to the present application, according to the actual medical hallucination detection result, the sample medical hallucination detection probability, the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient, a training loss function of the medical hallucination detection model is determined, including: according to the formula: , determining the training loss function of the medical hallucination detection model wherein, is the actual medical hallucination detection result of the kth sample feature information, , is the sample medical hallucination detection probability of the kth sample feature information, is the sample causal correlation strength coefficient of the kth sample feature information, is a preset causal correlation strength coefficient threshold, is the sample consistency coefficient of the kth sample feature information and the i th sample evidence information of the sample feature information, is a preset consistency coefficient threshold, is the sample timeliness coefficient of the i th sample evidence information of the kth sample feature information, is a preset timeliness coefficient threshold, is the sample confidence coefficient of the i th sample evidence information of the kth sample feature information, is a preset confidence coefficient threshold, n is the number of sample evidence information of the sample feature information, i≤n, K is the number of sample feature information, k≤K, i, n, k and K are all positive integers.

[0011] According to the present application, the training step of the medical hallucination correction model comprises: A second training sample data set is obtained. According to the second training sample data set, second sample feature information and second sample evidence information of the second sample feature information are obtained; According to the medical hallucination correction model, the second sample feature information and the second sample evidence information are processed to obtain sample correction feature information; According to the trained medical hallucination detection model, the sample correction feature information is processed to obtain a sample correction medical hallucination detection probability; According to the sample correction medical hallucination detection probability, a second training loss function of the medical hallucination correction model is determined; According to the training loss function of the medical hallucination correction model, the medical hallucination detection model is trained to obtain a trained medical hallucination correction model.

[0012] According to the present application, according to the sample correction medical hallucination detection probability, a second training loss function of the medical hallucination correction model is determined, including: according to the formula: , determining the second training loss function of the medical hallucination correction model wherein, is the sample correction medical hallucination detection probability of the jth sample correction feature information, m is the number of sample correction feature information, j≤m, j and m are positive integers.

[0013] According to the second aspect of the present application, a knowledge enhanced medical hallucination static detection and correction system is provided, comprising: a feature information module for processing the text to be detected by a feature extraction model to obtain first feature information; a causal strength module for determining a causal correlation strength coefficient according to the first feature information; a dynamic knowledge module for constructing a dynamic knowledge element network; an evidence information module for obtaining first evidence information according to the first feature information and the dynamic knowledge element network; a consistency coefficient module for processing the first feature information and the first evidence information by a knowledge enhanced NLI model to obtain a first consistency coefficient; an evidence coefficient module for determining a timeliness coefficient and a confidence coefficient of the first evidence information according to the dynamic knowledge element network; a detection probability module for processing the causal correlation strength coefficient, the first consistency coefficient, the timeliness coefficient and the confidence coefficient by a trained medical hallucination detection model to generate a medical hallucination detection probability; a determination and correction module for determining whether the first feature information needs to be corrected according to the medical hallucination detection probability; a hallucination correction module configured to correct the first feature information by a trained medical hallucination correction model if it is determined that the correction is needed.

[0014] Technical effects: According to the present application, the text to be detected can be divided into multiple first feature information, and the causal correlation strength of the first feature information, the consistency of the first feature information and the corresponding first evidence information, and the timeliness and confidence of the first evidence information can be accurately analyzed, and the analysis results are processed by the medical hallucination detection model to obtain the medical hallucination detection probability. According to the medical hallucination detection probability, it is determined whether the first feature information needs to be corrected. In the case that the first feature information needs to be corrected, the first feature information is corrected by the medical hallucination correction model, which improves the accuracy of medical hallucination static detection and correction. When determining the training loss function, the training loss function of the medical hallucination detection model can be determined according to the actual medical hallucination detection result, the sample medical hallucination detection probability, the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient. In the calculation process, the possible influence of the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient on the medical hallucination detection result can be determined to determine the error influence of the above data on the sample medical hallucination detection probability, and based on the influence and the error of the sample medical hallucination detection probability, the training loss function is set. In the training process, the training loss function is reduced, and the accuracy of the medical hallucination detection model is improved more targeted. When determining the second training loss function, the second training loss function of the medical hallucination correction model can be determined according to the sample correction medical hallucination detection probability. In the training process, the correctness of the sample correction feature information processed by the medical hallucination correction model can be improved, and the performance of the medical hallucination correction model can be improved.

[0015] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, but not limiting the present application. Other features and aspects of the present application will be more apparent from the following detailed description of exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other embodiments according to these drawings without creating any inventive labor; Figure 1 Exemplarily, a flowchart of a knowledge-enhanced medical hallucination static detection and correction method according to an embodiment of the present application is shown; Figure 2A schematic diagram for determining a causality correlation strength coefficient according to an embodiment of the present application is exemplarily shown. Figure 3 A schematic diagram for determining a timeliness coefficient and a confidence coefficient of first evidence information according to an embodiment of the present application is exemplarily shown. Figure 4 A block diagram of a knowledge-enhanced medical hallucination static detection and correction system according to an embodiment of the present application is exemplarily shown. DETAILED DESCRIPTION

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in connection with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0018] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and some embodiments can not be described again for the same or similar concepts or processes.

[0019] Figure 1 A flowchart of a knowledge-enhanced medical hallucination static detection and correction method according to an embodiment of the present application is exemplarily shown, and the method comprises: Step S1, processing a to-be-detected text by a feature extraction model to obtain first feature information; Step S2, determining a causality correlation strength coefficient according to the first feature information; Step S3, constructing a dynamic knowledge element network; Step S4, obtaining first evidence information according to the first feature information and the dynamic knowledge element network; Step S5, processing the first feature information and the first evidence information by a knowledge-enhanced NLI model to obtain a first consistency coefficient; Step S6, determining a timeliness coefficient and a confidence coefficient of first evidence information according to the dynamic knowledge element network; Step S7, processing the causality correlation strength coefficient, the first consistency coefficient, the timeliness coefficient, and the confidence coefficient by a trained medical hallucination detection model to generate a medical hallucination detection probability; Step S8, determining whether the first feature information needs to be corrected according to the medical hallucination detection probability; Step S9, in the case of determining that correction is needed, correcting the first feature information by a trained medical hallucination correction model.

[0020] According to the knowledge-enhanced medical hallucination static detection and correction method of the embodiment of the present application, the text to be detected can be divided into a plurality of first characteristic information, the causal correlation strength of the first characteristic information, the consistency of the first characteristic information and the corresponding first evidence information, and the timeliness and confidence of the first evidence information are accurately analyzed, and the analysis results are processed through a medical hallucination detection model to obtain a medical hallucination detection probability. According to the medical hallucination detection probability, it is determined whether the first characteristic information needs to be corrected. In the case that the first characteristic information needs to be corrected, the first characteristic information is corrected through a medical hallucination correction model, thereby improving the accuracy of medical hallucination static detection and correction.

[0021] According to an embodiment of the present application, in step S1, the first characteristic information is obtained by processing the text to be detected through a feature extraction model.

[0022] For example, the complete medical text (such as a diagnosis suggestion) generated by the LLM, i.e. the text to be detected, is input into the feature extraction model, and the text to be detected is divided into independent and verifiable first characteristic information, such as the content of the text to be detected is “the patient can take ibuprofen (200mg, three times a day) to relieve pain and inflammation, but needs to pay attention to its gastrointestinal side effects, and is recommended to be taken with food”, and the text to be detected is processed through the feature extraction model to obtain a plurality of first characteristic information, such as “ibuprofen, for, pain relief”, “ibuprofen, for, inflammation relief”, “ibuprofen, has side effects, gastrointestinal irritation”, etc.

[0023] According to an embodiment of the present application, in step S2, the causal correlation strength coefficient is determined according to the first characteristic information.

[0024] Figure 2 An exemplary schematic diagram for determining the causal correlation strength coefficient according to the embodiment of the present application is shown.

[0025] According to an embodiment of the present application, step S2 includes: Step S21, acquiring a first causality verb according to the first characteristic information; Step S22, determining a strong causal verb recognition result, a relevance verb recognition result and a fact verb recognition result according to the first causality verb; Step S23, determining a causal correlation strength coefficient according to the strong causal verb recognition result, the relevance verb recognition result and the fact verb recognition result.

[0026] For example, a causal verb in the first feature information is extracted, that is, a first causal verb such as "treatment", "cause", "inhibition", "accompany", "is", "is", and the like; it is judged that the first causal verb belongs to which one of a strong causal verb (an explicit causal verb directly asserting that an event, behavior or entity causes, prevents or changes another event or state, such as treatment, cause, cause, induce, prevent, inhibit, aggravate, relieve, cure, and the like), a correlation verb (a causal verb without explicit assertion, only indicating a statistical connection or co-occurrence between two things, such as accompanying, coefficient, higher / lower probability), and a fact verb (a verb that claims to describe a static, objective attribute, definition, component or numerical value, such as "is"), if it belongs to a strong causal verb, the strong causal verb is 1, otherwise it is 0. The determination method of the correlation verb recognition result and the fact verb recognition result is the same as that of the strong causal verb recognition result, which will not be repeated here; if the strong causal verb recognition result is 1, the causal correlation strength coefficient is 2, if the correlation verb recognition result is 1, the causal correlation strength coefficient is 1, and if the fact verb recognition result is 1, the causal correlation strength coefficient is 3.

[0027] According to an embodiment of the present application, step S3, a dynamic knowledge element network is constructed.

[0028] For example, data sources are integrated from UMLS, DrugBank, clinical guidelines, UpToDate, and the latest literature (continuously crawled through PubMed API), a graph database is used to store the data sources, the edges in the graph database represent the relationship, and are attached with a list of evidence sources, last update time and the like. Attributes, and through NLP information extraction technology, automatically extract triples from new literature, and after manual or high-quality model review, integrate into the graph to realize real-time update of the dynamic knowledge element network.

[0029] According to an embodiment of the present application, step S4, first evidence information is obtained according to the first feature information and the dynamic knowledge element network.

[0030] For example, for the first feature information, a graph query is performed in the dynamic knowledge element network to obtain a related evidence graph, and the first evidence information is extracted according to the related evidence graph.

[0031] According to an embodiment of the present application, step S5, the first feature information and the first evidence information are processed by a knowledge-enhanced NLI model to obtain a first consistency coefficient.

[0032] For example, the consistency of the first feature information and the first evidence information is judged by using a natural language inference model pre-trained with medical knowledge, that is, a knowledge-enhanced NLI model such as MedNLI or a BioBERT / ClinicalBERT-based NLI model, to obtain a first consistency coefficient. The higher the consistency of the first feature information and the first evidence information, the greater the first consistency coefficient.

[0033] According to an embodiment of the present application, in step S6, a timeliness coefficient and a confidence coefficient of the first evidence information are determined according to the dynamic knowledge element network.

[0034] Figure 3 An exemplary schematic diagram for determining the timeliness coefficient and the confidence coefficient of the first evidence information according to an embodiment of the present application is shown.

[0035] According to an embodiment of the present application, step S6 includes: In step S61, the last update time, the information source type, and the information research sample data of the first evidence information are obtained according to the dynamic knowledge element network. In step S62, the timeliness coefficient is determined according to the last update time. In step S63, the confidence coefficient is determined according to the information source type and the information research sample data.

[0036] For example, the last update time (the time length from the last update of the evidence information to the current time), the information source, and the information research sample data of each first evidence information are determined through the information attached in the “edge” in the dynamic knowledge element network. The timeliness coefficient is determined according to the ratio of the preset update time (which can be set to 6 months) and the last update time. The greater the timeliness coefficient, the stronger the timeliness of the first evidence information. The confidence of the first evidence information is evaluated according to the information source type and the information research sample data to determine the confidence coefficient.

[0037] According to an embodiment of the present application, step S63 includes: In step S631, the highest confidence recognition result, the high confidence recognition result, the medium confidence recognition result, and the low confidence recognition result are determined according to the information source type. In step S632, the information source confidence coefficient is determined according to the highest confidence recognition result, the high confidence recognition result, the medium confidence recognition result, and the low confidence recognition result. In step S633, the gender sample recognition result and the age sample recognition result are determined according to the information research sample data. In step S634, the information sample confidence coefficient is determined according to the gender sample recognition result and the age sample recognition result. Step S635, according to the information source confidence coefficient and the information sample confidence coefficient, determine the confidence coefficient.

[0038] For example, if the information source type belongs to authoritative guide, systematic review and Meta analysis, the highest confidence identification result is 4, otherwise, the highest confidence identification result is 0, if the information source type belongs to large-scale randomized controlled trial, the high confidence identification result is 3, otherwise, the high confidence identification result is 0, if the information source type belongs to observational study or cohort study, the medium confidence identification result is 2, otherwise, the medium confidence identification result is 0, if the information source type belongs to case report or expert opinion, the low confidence identification result is 1, otherwise, the low confidence identification result is 0; according to the sum of the highest confidence identification result, the high confidence identification result, the medium confidence identification result and the low confidence identification result, determine the information source confidence coefficient, the larger the information source confidence coefficient, the higher the information source confidence; according to the information research sample data, determine the gender sample identification result and the age sample identification result, such as, according to the information research sample data to determine the gender covered by the information research sample, if the information research sample covers men and women, the gender sample identification result is 1, otherwise 0, if the information research sample covers all ages (such as children, young, middle-aged and old), the age sample identification result is 1, otherwise, the ratio of the age covered by the information research sample to the whole age determines the age sample identification result, such as, the age covered by the information research sample is “young”, the corresponding age sample identification result is 1 / 4; according to the sum of the gender sample identification result and the age sample identification result, determine the information sample confidence coefficient; according to the sum of the information source confidence coefficient and the information sample confidence coefficient, determine the confidence coefficient.

[0039] According to one embodiment of the present application, step S7, the causal association strength coefficient, the first consistency coefficient, the timeliness coefficient and the confidence coefficient are processed by the trained medical illusion detection model to generate a medical illusion detection probability.

[0040] For example, the medical illusion detection model belongs to a kind of neural network model, including: data preprocessing and input layer, feature extraction layer, the medical illusion detection model is trained through historical data pair, so that the medical illusion detection model can output medical illusion detection probability.

[0041] According to one embodiment of the present application, the training step of the medical illusion detection model includes: Obtain the training sample data set; According to the training sample data set, obtain sample feature information and sample evidence information of sample feature information; determine a sample causality correlation strength coefficient of the sample feature information; determine a sample consistency coefficient of the sample feature information and the sample evidence information; determine a sample timeliness coefficient and a sample confidence coefficient of the sample evidence information; process the sample causality correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient according to the medical hallucination detection model, and determine a sample medical hallucination detection probability of the sample feature information; obtain an actual medical hallucination detection result of the sample feature information; determine a training loss function of the medical hallucination detection model according to the actual medical hallucination detection result, the sample medical hallucination detection probability, the sample causality correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient; train the medical hallucination detection model according to the training loss function of the medical hallucination detection model, and obtain a trained medical hallucination detection model.

[0042] For example, the historical data verified by expert detection is acquired as a training sample data set; in the training sample data set, sample feature information used for training and sample evidence information of the sample feature information are acquired; a sample causal correlation strength coefficient of the sample feature information is determined, and the determination manner of the sample causal correlation strength coefficient is similar to the determination manner of the causal correlation strength coefficient, which will not be repeated here; a sample consistency coefficient of the sample feature information and the sample evidence information is determined, and the determination manner of the sample consistency coefficient is similar to the determination manner of the first consistency coefficient, which will not be repeated here; a sample timeliness coefficient and a sample confidence coefficient of the sample evidence information are determined, and the determination manners of the sample timeliness coefficient and the sample confidence coefficient are similar to the determination manners of the timeliness coefficient and the confidence coefficient, which will not be repeated here; the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient are processed according to the medical hallucination detection model, and a probability that the sample feature information belongs to a medical environment, that is, a sample medical hallucination detection probability, is determined; a result whether the sample feature information belongs to a medical hallucination verified by expert detection in an actual practice process is acquired, that is, an actual medical hallucination detection result; when the sample feature information belongs to a medical hallucination, the actual medical hallucination detection result is equal to 1; when the sample feature information does not belong to a medical hallucination, the actual medical hallucination detection result is equal to 0; a training loss function of the medical hallucination detection model is determined according to the actual medical hallucination detection result, the sample medical hallucination detection probability, the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient; the medical hallucination detection model is trained according to the training loss function of the medical hallucination detection model, the accuracy of the medical hallucination detection model for medical environment detection is improved, and a trained medical hallucination detection model is obtained.

[0043] According to one embodiment of the application, the training loss function of the medical hallucination detection model is determined according to the actual medical hallucination detection result, the sample medical hallucination detection probability, the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient, and comprises: determining the training loss function of the medical hallucination detection model according to formula (1) , (1) wherein, is an actual medical hallucination detection result of the kth sample feature information, , is a sample medical hallucination detection probability of the kth sample feature information, is a sample causal correlation strength coefficient of the kth sample feature information, is a preset causal correlation strength coefficient threshold, a sample consistency coefficient of the kth sample feature information and the ith sample evidence information of the sample feature information, a preset consistency coefficient threshold, a sample timeliness coefficient of the ith sample evidence information of the kth sample feature information, a preset timeliness coefficient threshold, a sample confidence coefficient of the ith sample evidence information of the kth sample feature information, a preset confidence coefficient threshold, n is the number of sample evidence information of the sample feature information, i≤n, K is the number of sample feature information, k≤K, i, n, k and K are positive integers.

[0044] According to one embodiment of the present application, a ratio of the sample consistency coefficient of the kth sample feature information and the ith sample evidence information of the sample feature information to the preset consistency coefficient threshold, the greater the ratio, the greater the sample consistency coefficient of the kth sample feature information and the ith sample evidence information of the sample feature information, and the preset consistency coefficient threshold can be set to 0.9, a ratio of the sample timeliness coefficient of the ith sample evidence information of the kth sample feature information to the preset timeliness coefficient threshold, the greater the ratio, the greater the sample timeliness coefficient of the ith sample evidence information of the kth sample feature information, and the preset timeliness coefficient threshold can be set to 1, a ratio of the sample confidence coefficient of the ith sample evidence information of the kth sample feature information to the preset confidence coefficient threshold, the greater the ratio, the greater the sample confidence coefficient of the ith sample evidence information of the kth sample feature information, and the preset confidence coefficient threshold can be set to 1, a ratio of the sample causal correlation strength coefficient of the kth sample feature information to the preset causal correlation strength coefficient threshold, the greater the ratio, the greater the sample causal correlation strength coefficient of the kth sample feature information, and the preset causal correlation strength coefficient threshold can be set to 1, The sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient are negatively correlated with the size of the sample medical hallucination detection probability, and the sample causal correlation strength coefficient is positively correlated with the size of the sample medical hallucination detection probability. For example, the greater the sample consistency coefficient, the higher the consistency of the first feature information and the first evidence information, the lower the probability that the first feature information is proved by the first evidence information to be medical hallucination, and the smaller the sample medical hallucination detection probability. The greater the sample timeliness coefficient, the more recent the update time of the first evidence information, the higher the credibility of the first evidence information, the higher the evidence strength, the lower the probability that the first feature information is proved by the first evidence information to be medical hallucination, and the smaller the sample medical hallucination detection probability. The greater the sample confidence coefficient, the more authoritative the source of the first evidence information and the more comprehensive the research sample coverage of the first evidence information, the higher the credibility of the first evidence information, the higher the evidence strength, the lower the probability that the first feature information is proved by the first evidence information to be medical hallucination, and the smaller the sample medical hallucination detection probability. The greater the sample causal correlation strength coefficient, the higher the evidence strength required by the first feature information, the greater the probability that the first feature information is determined to be medical hallucination due to insufficient evidence strength, and the greater the sample medical hallucination detection probability.

[0045] According to one embodiment of the present application, The error of the actual medical hallucination detection result of the kth sample feature information and the sample medical hallucination detection probability is calculated by The weighted sum of the error of the actual medical hallucination detection result and the sample medical hallucination detection probability is obtained. In the training process, the training loss function is reduced, the accuracy of the medical hallucination detection model for medical environment detection is improved, and the precision of the medical hallucination detection model is improved.

[0046] In this way, the training loss function of the medical hallucination detection model can be determined according to the actual medical hallucination detection result, the sample medical hallucination detection probability, the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient. In the calculation process, the possible influence of the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient on the medical hallucination detection result can be determined to determine the error influence of the above data on the sample medical hallucination detection probability, and the training loss function is set based on the influence and the error of the sample medical hallucination detection probability. In the training process, the training loss function is reduced, and the precision of the medical hallucination detection model is more targeted improved.

[0047] According to one embodiment of the present application, step S8, whether the first feature information needs to be corrected is determined according to the medical hallucination detection probability.

[0048] For example, when the medical hallucination detection probability is greater than or equal to 0.9, it is determined that the first feature information needs to be corrected.

[0049] According to one embodiment of the present application, in the case of determining that correction is needed, the first feature information is corrected by the trained medical hallucination correction model.

[0050] For example, the medical hallucination correction model is a deep learning neural network model such as a convolutional neural network model, which can correct the first feature information detected as medical hallucination based on the input first evidence information. The corrected first feature information may still belong to medical hallucination, and therefore the medical hallucination correction model can be trained so that the corrected first feature information does not belong to medical hallucination.

[0051] According to one embodiment of the present application, the training step of the medical hallucination correction model comprises: obtaining a second training sample data set; obtaining second sample feature information and second sample evidence information of the second sample feature information according to the second training sample data set; processing the second sample feature information and the second sample evidence information according to the medical hallucination correction model to obtain sample corrected feature information; processing the sample corrected feature information according to the trained medical hallucination detection model to obtain a sample corrected medical hallucination detection probability; determining a second training loss function of the medical hallucination correction model according to the sample corrected medical hallucination detection probability; training the medical hallucination detection model according to the training loss function of the medical hallucination correction model to obtain the trained medical hallucination correction model.

[0052] For example, various sample feature information that has been detected as medical hallucination and its corresponding evidence information are obtained as the second training sample data set; second sample feature information and second sample evidence information of the second sample feature information are obtained according to the second training sample data set; the second sample feature information and the second sample evidence information are processed according to the medical hallucination correction model to obtain corrected second sample feature information, i.e., sample corrected feature information; the sample corrected feature information is processed according to the trained medical hallucination detection model to obtain a probability that the sample corrected feature information is detected as medical hallucination, i.e., a sample corrected medical hallucination detection probability; a second training loss function of the medical hallucination correction model is determined according to the sample corrected medical hallucination detection probability; the medical hallucination detection model is trained according to the training loss function of the medical hallucination correction model to improve the accuracy of the medical hallucination correction model, and the trained medical hallucination correction model is obtained.

[0053] According to one embodiment of the present application, the medical illusion correction model is trained according to the sample correction medical illusion detection probability, and a second training loss function of the medical illusion correction model is determined according to the sample correction medical illusion detection probability, including: determining the second training loss function of the medical illusion correction model according to formula (2) , (2) wherein, is the sample correction medical illusion detection probability of the jth sample correction feature information, m is the number of sample correction feature information, j≤m, j and m are positive integers.

[0054] According to one embodiment of the present application, in the training process of the medical illusion correction model, is minimized, then is as close to 1 as possible, so that is minimized, so that the probability of the sample correction feature information processed by the medical illusion correction model being detected as a medical environment by the trained medical illusion detection model is reduced, thereby improving the correctness of the sample correction feature information processed by the medical illusion correction model and improving the performance of the medical illusion correction model.

[0055] In this way, the second training loss function of the medical illusion correction model can be determined according to the sample correction medical illusion detection probability, and in the training process, the correctness of the sample correction feature information processed by the medical illusion correction model can be improved, and the performance of the medical illusion correction model can be improved.

[0056] According to the knowledge-enhanced medical hallucination static detection and correction method of the embodiment of the present application, the text to be detected can be divided into a plurality of first characteristic information, the causal correlation strength of the first characteristic information, the consistency of the first characteristic information and the corresponding first evidence information, and the timeliness and confidence of the first evidence information are accurately analyzed, and the analysis results are processed through the medical hallucination detection model to obtain a medical hallucination detection probability. According to the medical hallucination detection probability, it is determined whether the first characteristic information needs to be corrected. In the case that the first characteristic information needs to be corrected, the first characteristic information is corrected through the medical hallucination correction model, thereby improving the accuracy of medical hallucination static detection and correction. When determining the training loss function, the actual medical hallucination detection result, the sample medical hallucination detection probability, the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient can be used to determine the training loss function of the medical hallucination detection model. In the calculation process, the possible influence of the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient on the medical hallucination detection result can be used to determine the error influence of the above data on the sample medical hallucination detection probability, and based on the influence and the error of the sample medical hallucination detection probability, the training loss function is set. In the training process, the training loss function is reduced, and the accuracy of the medical hallucination detection model is improved more targeted. When determining the second training loss function, the sample correction medical hallucination detection probability can be used to determine the second training loss function of the medical hallucination correction model. In the training process, the correctness of the sample correction characteristic information processed by the medical hallucination correction model can be improved, and the performance of the medical hallucination correction model can be improved.

[0057] Figure 4 An exemplary block diagram of a knowledge-enhanced medical hallucination static detection and correction system according to an embodiment of the present application is shown. The system includes: A characteristic information module for processing the text to be detected by a feature extraction model to obtain first characteristic information; A causal strength module for determining a causal correlation strength coefficient according to the first characteristic information; A dynamic knowledge module for constructing a dynamic knowledge network; An evidence information module for obtaining first evidence information according to the first characteristic information and the dynamic knowledge network; A consistency coefficient module for processing the first characteristic information and the first evidence information through a knowledge-enhanced NLI model to obtain a first consistency coefficient; An evidence coefficient module for determining a timeliness coefficient and a confidence coefficient of the first evidence information according to the dynamic knowledge network; The detection probability module is configured to process the causal correlation strength coefficient, the first consistency coefficient, the timeliness coefficient and the confidence coefficient by using a trained medical hallucination detection model to generate a medical hallucination detection probability. The determination correction module is configured to determine whether the first feature information needs to be corrected according to the medical hallucination detection probability. The hallucination correction module is configured to correct the first feature information by using a trained medical hallucination correction model in the case of determining that the correction is needed.

[0058] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions loaded thereon for performing various aspects of the present application.

[0059] Those skilled in the art will understand that the embodiments of the application shown in the above description and the drawings are only examples and do not limit the application. The purpose of the application has been fully and effectively achieved. The function and structural principle of the application has been shown and explained in the embodiments, and the implementation of the application can be any modification or modification without departing from the principle.

Claims

1. A knowledge-augmented medical hallucination static detection and correction method, characterized in that, The method comprises the following steps: processing the to-be-detected text through a feature extraction model to obtain first feature information; determining a causal correlation strength coefficient according to the first feature information; constructing a dynamic knowledge element network; obtaining first evidence information according to the first feature information and the dynamic knowledge element network; processing the first feature information and the first evidence information through a knowledge-enhanced NLI model to obtain a first consistency coefficient; determining a timeliness coefficient and a confidence coefficient of the first evidence information according to the dynamic knowledge element network; processing the causal correlation strength coefficient, the first consistency coefficient, the timeliness coefficient and the confidence coefficient through a trained medical illusion detection model to generate a medical illusion detection probability; determining whether the first feature information needs to be corrected according to the medical illusion detection probability; in the case of determining that correction is needed, correcting the first feature information through a trained medical illusion correction model.

2. The knowledge-augmented medical hallucination static detection and correction method of claim 1, wherein, According to the first feature information, the causal correlation strength coefficient is determined, comprising: obtaining a first causality verb according to the first feature information; determining a strong causal verb recognition result, an association verb recognition result and a fact verb recognition result according to the first causality verb; determining a causal correlation strength coefficient according to the strong causal verb recognition result, the association verb recognition result and the fact verb recognition result.

3. The knowledge-augmented medical hallucination static detection and correction method of claim 1, wherein, According to the dynamic knowledge element network, the timeliness coefficient and the confidence coefficient of the first evidence information are determined, comprising: obtaining the last update time, the information source type and the information research sample data of the first evidence information according to the dynamic knowledge element network; determining the timeliness coefficient according to the last update time; determining the confidence coefficient according to the information source type and the information research sample data.

4. The knowledge-augmented medical hallucination static detection and correction method of claim 3, wherein, According to the information source type and the information research sample data, the confidence coefficient is determined, comprising: determining a highest confidence recognition result, a high confidence recognition result, a medium confidence recognition result and a low confidence recognition result according to the information source type; determining an information source confidence coefficient according to the highest confidence recognition result, the high confidence recognition result, the medium confidence recognition result and the low confidence recognition result; determining a gender sample recognition result and an age sample recognition result according to the information research sample data; determining an information sample confidence coefficient according to the gender sample recognition result and the age sample recognition result; determining the confidence coefficient according to the information source confidence coefficient and the information sample confidence coefficient.

5. The knowledge-augmented medical hallucination static detection and correction method of claim 1, wherein, The training steps of the medical illusion detection model comprise: obtaining a training sample data set; obtaining sample feature information and sample evidence information of the sample feature information according to the training sample data set; determining a sample causal correlation strength coefficient of the sample feature information; determining a sample consistency coefficient of the sample feature information and the sample evidence information; determining a sample timeliness coefficient and a sample confidence coefficient of the sample evidence information; According to the medical hallucination detection model, the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient are processed to determine the sample feature information sample medical hallucination detection probability; Obtain the actual medical hallucination detection result of the sample feature information; According to the actual medical hallucination detection result, the sample medical hallucination detection probability, the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient, the training loss function of the medical hallucination detection model is determined; According to the training loss function of the medical hallucination detection model, the medical hallucination detection model is trained to obtain the trained medical hallucination detection model.

6. The knowledge-augmented medical hallucination static detection and correction method of claim 5, wherein, According to the actual medical hallucination detection result, the sample medical hallucination detection probability, the sample causal correlation strength coefficient, the sample consistency coefficient, the sample timeliness coefficient and the sample confidence coefficient, the training loss function of the medical hallucination detection model is determined, including: according to formula: ， A training loss function for determining a medical hallucination detection model , wherein is an actual medical hallucination detection result of kth sample feature information, , is a sample medical hallucination detection probability of kth sample feature information, is a sample causal correlation strength coefficient of kth sample feature information, is a preset causal correlation strength coefficient threshold, is a sample consistency coefficient of kth sample feature information and ith sample evidence information of the sample feature information, is a preset consistency coefficient threshold, is a sample timeliness coefficient of ith sample evidence information of kth sample feature information, is a preset timeliness coefficient threshold, is a sample confidence coefficient of ith sample evidence information of kth sample feature information, is a preset confidence coefficient threshold, n is a number of sample evidence information of sample feature information, i≤n, K is a number of sample feature information, k≤K, and i, n, k, and K are positive integers.

7. The knowledge-augmented medical hallucination static detection and correction method of claim 1, wherein, The training step of the medical hallucination correction model comprises: Obtain the second training sample data set; According to the second training sample data set, obtain the second sample feature information and the second sample evidence information of the second sample feature information; According to the medical hallucination correction model, the second sample feature information and the second sample evidence information are processed to obtain sample correction feature information; According to the trained medical hallucination detection model, the sample correction feature information is processed to obtain the sample correction medical hallucination detection probability; According to the sample correction medical hallucination detection probability, the second training loss function of the medical hallucination correction model is determined; According to the training loss function of the medical hallucination correction model, the medical hallucination detection model is trained to obtain the trained medical hallucination correction model.

8. The knowledge-augmented medical hallucination static detection and correction method of claim 7, wherein, According to the sample correction medical hallucination detection probability, the second training loss function of the medical hallucination correction model is determined, including: according to formula: ， Determine the second training loss function of the medical hallucination correction model wherein, is the sample correction medical hallucination detection probability of the jth sample correction feature information, m is the number of sample correction feature information, j≤m, j and m are both positive integers.

9. A knowledge-augmented medical hallucination static detection and correction system configured to perform the method of any one of claims 1-8, wherein, Including: The feature information module is used for processing the to-be-detected text through the feature extraction model to obtain the first feature information; The causal strength module is used for determining the causal correlation strength coefficient according to the first feature information; The dynamic knowledge module is used for constructing a dynamic knowledge element network; The evidence information module is used for obtaining the first evidence information according to the first feature information and the dynamic knowledge element network; The consistency coefficient module is used for processing the first feature information and the first evidence information through the knowledge enhanced NLI model to obtain the first consistency coefficient; The evidence coefficient module is used for determining the timeliness coefficient and the confidence coefficient of the first evidence information according to the dynamic knowledge element network; The detection probability module is used for processing the causal correlation strength coefficient, the first consistency coefficient, the timeliness coefficient and the confidence coefficient through the trained medical hallucination detection model to generate a medical hallucination detection probability; The determination correction module is used for determining whether the first feature information needs to be corrected according to the medical hallucination detection probability; An illusion correction module is configured to correct the first feature information by a trained medical illusion correction model in a case where it is determined that correction is needed. An illusion correction module is configured to correct the first feature information by a trained medical illusion correction model in a case where it is determined that correction is needed.

Images