Semantic reasoning method based on uncertainty knowledge graph

Abstract

The invention provides a semantic reasoning method based on an uncertainty knowledge graph, and relates to the technical field of text semantic reasoning, and the method comprises the steps: obtaining text semantic information through pattern matching; based on text semantic information, vectorizing the triple by using a knowledge graph of linear transformation to obtain an energy function; analyzing the energy function by using a triple evaluator to obtain a structural score and an uncertain score; performing linear fusion on the structural score and the uncertainty score to obtain a linear fusion score; and based on the uncertainty scores, calculating the confidence of each triad by using a confidence generator, and taking the linear fusion scores of the triads obtained by screening and the corresponding confidence as semantic reasoning results to complete semantic reasoning of the knowledge graph. According to the method, the problems that the knowledge extensibility and response speed of the knowledge graph are poor, the interpretability is poor, and the uncertainty of the knowledge is difficult to express are solved.

Description

Technical Field

[0001] This specification relates to the technical field of text semantic reasoning, and in particular to a semantic reasoning method based on uncertainty knowledge graph. Background Art

[0002] Semantic reasoning methods are one of the key technologies in the field of artificial intelligence. Their development background is rooted in the need for efficient management and utilization of massive knowledge, as well as the limitations of knowledge graphs in terms of attribute completeness and relationship richness. With the improvement of computing power and the advancement of machine learning technology, especially the application of deep learning in natural language processing, semantic reasoning algorithms have been able to simulate the human reasoning process and have shown their importance in many fields such as intelligent search, automated reasoning, and intelligent question answering. These algorithms include not only traditional methods based on logical rules, but also modern technologies such as graph representation learning, neural networks, and hybrid reasoning, which together promote the ability of artificial intelligence in understanding and processing complex semantic relationships. Knowledge graphs are widely used in decision support, recommendation systems, and other fields with their excellent knowledge expression and logical reasoning capabilities. However, in practical applications, the knowledge scalability and response speed of knowledge graphs are unsatisfactory, and in professional fields, the uncertainty of specific professional knowledge is often ignored. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a semantic reasoning method based on an uncertainty knowledge graph, which solves the problems of poor knowledge scalability and response speed, poor interpretability, and difficulty in expressing knowledge uncertainty of the knowledge graph.

[0004] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a semantic reasoning method based on uncertainty knowledge graph, comprising:

[0005] S1: Use pattern matching to obtain text semantic information;

[0006] S2: Based on the text semantic information, the triples are vectorized using a linearly transformed knowledge graph to obtain an energy function;

[0007] S3: Analyze the energy function using a triplet evaluator to obtain a structural score and an uncertainty score;

[0008] S4: linearly fuse the structural score and the uncertainty score to obtain a linear fusion score;

[0009] S5: Based on the uncertainty score, the confidence of each triple is calculated using the confidence generator, and the linear fusion score and corresponding confidence of the screened triple are used as the semantic reasoning result to complete the semantic reasoning of the knowledge graph.

[0010] The beneficial effects of the present invention are as follows: the processor uses the knowledge graph of linear transformation to vectorize the text semantic information, uses the triple evaluator to obtain the structural score and uncertainty score, and uses the confidence generator to calculate the confidence of each triple to obtain the semantic reasoning result. In this way, the structural uncertainty of the uncertainty knowledge graph and the uncertainty of the knowledge itself can be considered at the same time, a large number of entities and relationships and their uncertainties can be mapped into low-dimensional, dense real-valued vectors, and the knowledge graph completion method can be used to make decision recommendations that include the degree of uncertainty, thereby improving the efficiency and accuracy of decision-making.

[0011] Furthermore, the S1 includes:

[0012] Use the original symptom data to build a symptom library;

[0013] Based on the symptom library, pattern matching is used to extract the original symptom data to obtain text semantic information.

[0014] In this way, the accuracy and efficiency of symptom acquisition can be improved.

[0015] Furthermore, the expression of the energy function is:

[0016]

[0017] Where E(h,r,t) represents the energy function, (h,r,t) represents a triplet of text semantic information, H represents the first entity matrix, R represents the relationship matrix, and T represents the second entity matrix. represents the square of the norm of the matrix F, h represents the elements of the first entity matrix, r represents the elements of the relationship matrix, and t represents the elements of the second entity matrix; the first entity matrix, the relationship matrix and the second entity matrix belong to text semantic information.

[0018] The triple semantic features are extracted using the knowledge graph embedding module based on linear transformation, and the relationship between two entities is corresponded to the linear transformation of the entity matrix, which solves the problem that the existing model performs poorly on non-injective relations in knowledge graphs.

[0019] Furthermore, the expression of the structural score is:

[0020]

[0021] φ stru (E(h,r,t))=p stru ·E(h,r,t)+b stru ;

[0022] The expression for the uncertainty fraction is:

[0023]

[0024] φ unce (E(h,r,t))=p unce ·E(h,r,t)+b unce ;

[0025] Among them, Q stru represents a structural fraction, e represents a natural constant, φ stru represents the linear transformation of triple energy scores, E(h,r,t) represents the energy function, and p stru The first training parameter representing the structural score, b stru The second training parameter, Q, represents the structural score unce represents the uncertainty fraction, φ unce represents the linear transformation of triple energy fractions, p unce The first training parameter representing the uncertainty score, b unce A second training parameter representing an uncertainty score, h represents an element of the first entity matrix, r represents an element of the relationship matrix, and t represents an element of the second entity matrix.

[0026] Furthermore, the expression of the linear fusion score is:

[0027] score = α·Q stru +β·Q unce ;

[0028] Among them, score represents the linear fusion score, α represents the weight of the structural score, and Q stru represents the structural score, β represents the weight of the uncertainty score, Q unce Represents the uncertainty score.

[0029] Furthermore, the triple evaluator is obtained through training, and its first loss function is:

[0030] Γ Evaluator =Γ Pos +Γ Neg ;

[0031]

[0032] Among them, Γ Evaluator Represents the result of the first loss function, Γ Pos represents the loss function of positive triples, Γ Neg represents the loss function for negative triples, Q stru represents the structural score, Q uncerepresents the uncertainty score, w represents the confidence of the triple, S represents the set of positive triples, S' represents the set of negative triples, h represents the elements of the first entity matrix, r represents the elements of the relationship matrix, and t represents the elements of the second entity matrix.

[0033] A triple evaluator is designed that considers both structural information and uncertainty information. The evaluator learns the energy scores calculated by the knowledge graph embedding module respectively, while taking into account the uncertainty of the graph structure and knowledge content to more accurately represent the confidence of the triple.

[0034] Furthermore, the confidence generator is obtained through training, and its second loss function is:

[0035]

[0036] Among them, Γ Conf Represents the result of the second loss function, Q unce represents the uncertainty score, w represents the confidence of the tuple, S represents the set of positive triples, h represents the elements of the first entity matrix, r represents the elements of the relationship matrix, and t represents the elements of the second entity matrix.

[0037] Design a confidence generator. The confidence generator can also be regarded as a confidence prediction model based on triple scoring. Unlike the evaluator, the confidence generator does not share parameters with the evaluator and can efficiently predict the uncertainty of the knowledge graph. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0039] Figure 1 It is an exemplary flow chart of a semantic reasoning method based on an uncertain knowledge graph as shown in some embodiments of this specification. DETAILED DESCRIPTION

[0040] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0041] Embodiment 1

[0042] Figure 1is an exemplary flow chart of a semantic reasoning method based on an uncertain knowledge graph according to some embodiments of this specification. Figure 1 As shown, the process includes the following steps. In some embodiments, the process can be executed by a processor.

[0043] S1: Use pattern matching to obtain text semantic information.

[0044] Pattern matching is an algorithm that searches for matches in a target string from a main string.

[0045] In some embodiments, the processor may use matching patterns to search for text, search for a required small section of text from a large section of text, and obtain text semantic information.

[0046] Text semantic information is information reflecting entities in text semantics and the relationships between them. For example, the text semantic information may include positive triples and negative triples formed by the first entity matrix, the relationship matrix, and the second entity matrix.

[0047] A positive triple is a correct triple. For example, a positive triple may include types such as typical symptoms, general symptoms, and rare symptoms. For example, a symptom of a cold (disease) is a cough (symptom), which is a set of positive triples.

[0048] Negative triples are wrong triples. For example, negative triples may include types such as typical symptoms, general symptoms, and rare symptoms. For example, the symptom of gastritis (disease) is cough (symptom), which is a group of negative triples.

[0049] In some embodiments, the processor may construct a symptom library using the original symptom data; based on the symptom library, the original symptom data may be extracted using pattern matching to obtain text semantic information.

[0050] Symptom raw data is data reflecting disease text and corresponding symptoms.

[0051] The symptom database is a database constructed from raw symptom data.

[0052] In some embodiments, the processor can utilize professional medical knowledge such as clinical guidelines and relevant personnel to upload symptom conditions and obtain raw symptom data.

[0053] In some embodiments, the processor may extract raw symptom data from a symptom library based on pattern matching to obtain textual semantic information.

[0054] S2: Based on the text semantic information, the triples are vectorized using the linearly transformed knowledge graph to obtain an energy function.

[0055] The linear transformation knowledge graph is a knowledge graph that contains entities and corresponding relationships.

[0056] In some embodiments, the processor can obtain triples based on text semantic information, entities and corresponding relationships therein; and use the triples to perform linear transformation to obtain a linearly transformed knowledge graph.

[0057] The energy function is a linear transformation function between two entities corresponding to the entity matrix.

[0058] In some embodiments, the expression of the energy function may be:

[0059]

[0060] Where E(h,r,t) represents the energy function, (h,r,t) represents a triplet of text semantic information, H represents the first entity matrix, R represents the relationship matrix, and T represents the second entity matrix. represents the square of the norm of the matrix F, h represents the elements of the first entity matrix, r represents the elements of the relationship matrix, and t represents the elements of the second entity matrix; the first entity matrix, the relationship matrix and the second entity matrix belong to text semantic information.

[0061] The F-norm of a matrix is ​​the square root of the sum of the squares of each element in the matrix. It can be used to measure the size of a matrix (the similarity between HR and T). The smaller the result, the more similar they are.

[0062] In some embodiments, when the relationship matrix is ​​a singular matrix, the head entity may be linked to multiple tail entities, and the tail entity may be linked to multiple head entities. The processor can model various types of relationships based on the non-injective properties of the triple relationships retained therein, such as non-injective, symmetric, asymmetric, reverse and combination.

[0063] S3: Analyze the energy function using a triplet evaluator to obtain a structural score and an uncertainty score.

[0064] The triplet evaluator is a neural network model used to obtain structural and uncertainty scores.

[0065] In some embodiments, the triple evaluator can be trained by multiple labeled first training samples. For example, multiple labeled first training samples can be input into the initial triple evaluator, and the first loss function can be constructed by the label and the result of the initial triple evaluator. The parameters of the initial triple evaluator are iteratively updated based on the first loss function by gradient descent or other methods. When the preset conditions are met, the model training is completed and a trained triple evaluator is obtained. The preset conditions can be that the first loss function converges, the number of iterations reaches a threshold, etc.

[0066] In some embodiments, the first training sample may include a historical energy function. The labels may be corresponding structural scores and uncertainty scores. The labels may be manually annotated.

[0067] In some embodiments, the processor may train the triple evaluator using the first loss function to obtain a trained triple evaluator.

[0068] In some embodiments, the expression of the first loss function may be:

[0069] Γ Evaluator =Γ Pos +Γ Neg ;

[0070]

[0071]

[0072] Among them, Γ Evaluator Represents the result of the first loss function, Γ Pos represents the loss function of positive triples, Γ Neg represents the loss function for negative triples, Q stru represents the structural score, Q unce represents the uncertainty score, w represents the confidence of the triple, S represents the set of positive triples, S' represents the set of negative triples, h represents the elements of the first entity matrix, r represents the elements of the relationship matrix, and t represents the elements of the second entity matrix.

[0073] In some embodiments, the processor may utilize a first loss function to encourage positive triples to be close to 1 and negative triples to be close to 0; to encourage positive triples to be close to true confidence and to encourage negative triples to be close to 0.

[0074] In this way, positive and negative examples in uncertain knowledge graphs can be distinguished.

[0075] The structural score is a score that measures the correctness of a triple from the perspective of the knowledge graph structure.

[0076] In some embodiments, the expression of the structural score may be:

[0077]

[0078] φ stru (E(h,r,t))=p stru ·E(h,r,t)+b stru ;

[0079] Among them, Q stru represents a structural fraction, e represents a natural constant, φ strurepresents the linear transformation of triple energy scores, E(h,r,t) represents the energy function, and p stru The first training parameter representing the structural score, b stru A second training parameter representing the structural score.

[0080] The uncertainty score is a score that measures the confidence of a triple from the perspective of the knowledge itself.

[0081] In some embodiments, the expression for the uncertainty score may be:

[0082]

[0083] φ unce (E(h,r,t))=p unce ·E(h,r,t)+b unce ;

[0084] Among them, Q unce represents the uncertainty fraction, φ unce represents the linear transformation of triple energy fractions, p unce The first training parameter representing the uncertainty score, b unce A second training parameter representing an uncertainty score, h represents an element of the first entity matrix, r represents an element of the relationship matrix, and t represents an element of the second entity matrix.

[0085] S4: Linearly fuse the structural score and the uncertainty score to obtain a linear fusion score.

[0086] The linear fusion score is data that reflects the energy strength of each triple in the text semantic information.

[0087] In some embodiments, the expression of the linear fusion score may be:

[0088] score = α·Q stru +β·Q unce ;

[0089] Among them, score represents the linear fusion score, α represents the weight of the structural score, and Q stru represents the structural score, β represents the weight of the uncertainty score, Q unce Represents the uncertainty score.

[0090] S5: Based on the uncertainty score, the confidence of each triple is calculated using the confidence generator, and the linear fusion score and corresponding confidence of the screened triple are used as the semantic reasoning result to complete the semantic reasoning of the knowledge graph.

[0091] The confidence generator is a neural network model used to obtain the confidence of a triple.

[0092] In some embodiments, the confidence generator can be trained by multiple labeled second training samples. For example, multiple labeled second training samples can be input into the initial confidence generator, and the second loss function can be constructed by the labels and the results of the initial confidence generator. The parameters of the initial confidence generator are iteratively updated by gradient descent or other methods based on the second loss function. When the preset conditions are met, the model training is completed and a trained confidence generator is obtained. Among them, the preset conditions can be that the second loss function converges, the number of iterations reaches a threshold, etc.

[0093] In some embodiments, the second training sample may include an uncertainty score. The label may be a confidence level of the corresponding triple. The label may be manually annotated.

[0094] In some embodiments, the processor may train the confidence generator using the second loss function to obtain a trained confidence generator.

[0095] In some embodiments, the expression of the second loss function may be:

[0096]

[0097] Among them, Γ Conf Represents the result of the second loss function, Q unce represents the uncertainty score, w represents the confidence of the tuple, S represents the set of positive triples, h represents the elements of the first entity matrix, r represents the elements of the relationship matrix, and t represents the elements of the second entity matrix.

[0098] In some embodiments, the processor may utilize a confidence generator to approximate the true confidence value of a triple, the confidence generator not sharing parameters with the triple evaluator.

[0099] The semantic reasoning result is a result reflecting the relationship between entities in the text semantic information. For example, the semantic reasoning result may include the linear fusion score and corresponding confidence of each triple reflecting the entity relationship.

[0100] In some embodiments, workers can assist in work decision-making based on semantic reasoning results.

[0101] In some embodiments of this specification, the processor uses a linearly transformed knowledge graph to vectorize text semantic information, uses a triple evaluator to obtain a structural score and an uncertainty score, and uses a confidence generator to calculate the confidence of each triple to obtain a semantic reasoning result. In this way, the structural uncertainty of the uncertainty knowledge graph and the uncertainty of the knowledge itself can be considered at the same time, a large number of entities and relationships and their uncertainties can be mapped into low-dimensional, dense real-valued vectors, and a knowledge graph completion method can be used to make decision recommendations that include the degree of uncertainty, thereby improving the efficiency and accuracy of decision-making.

[0102] Embodiment 2

[0103] In some embodiments, a semantic reasoning method based on an uncertainty knowledge graph includes:

[0104] S1: Use pattern matching to obtain textual semantic information of medical disease symptoms;

[0105] S2: Based on the text semantic information of the medical disease symptoms, the triples are vectorized using a linearly transformed knowledge graph to obtain an energy function;

[0106] S3: Analyze the energy function using a triplet evaluator to obtain a structural score and an uncertainty score;

[0107] S4: linearly fuse the structural score and the uncertainty score to obtain a linear fusion score;

[0108] S5: Based on the uncertainty score, the confidence of each triple is calculated using a confidence generator, and the linear fusion score and corresponding confidence of the screened triples are used as the semantic reasoning results of the medical disease symptoms to complete the semantic reasoning of the knowledge graph of medical disease symptoms.

[0109] In some embodiments, the S1 includes:

[0110] Use the original symptom data to build a symptom library;

[0111] Based on the symptom library, pattern matching is used to extract the original symptom data to obtain textual semantic information of medical disease symptoms.

[0112] In some embodiments, the energy function is expressed as:

[0113]

[0114] Where E(h,r,t) represents the energy function, (h,r,t) represents a triplet of semantic information of medical disease symptom text, H represents the first entity matrix, R represents the relationship matrix, and T represents the second entity matrix. Indicates that h represents the elements of the first entity matrix, r represents the elements of the relationship matrix, and t represents the elements of the second entity matrix; the first entity matrix, the relationship matrix and the second entity matrix belong to the semantic information of medical disease symptom text.

[0115] In some embodiments, the structural score is expressed as:

[0116]

[0117] φ stru (E(h,r,t))=p stru ·E(h,r,t)+b stru ;

[0118] The expression for the uncertainty fraction is:

[0119]

[0120] φ unce (E(h,r,t))=p unce ·E(h,r,t)+b unce ;

[0121] Among them, Q stru represents a structural fraction, e represents a natural constant, φ stru represents, E(h,r,t) represents the energy function, p stru The first training parameter representing the structural score, b stru The second training parameter, Q, represents the structural score unce represents the uncertainty fraction, φ unce Indicates that p unce The first training parameter representing the uncertainty score, b unce A second training parameter representing an uncertainty score, h represents an element of the first entity matrix, r represents an element of the relationship matrix, and t represents an element of the second entity matrix.

[0122] In some embodiments, the linear fusion score is expressed as:

[0123] score = α·Q stru +β·Q unce ;

[0124] Among them, score represents the linear fusion score, α represents the weight of the structural score, and Q stru represents the structural score, β represents the weight of the uncertainty score, Q unce Represents the uncertainty score.

[0125] In some embodiments, the triple evaluator is obtained by training, and its first loss function is:

[0126] Γ Evaluator =Γ Pos +Γ Neg ;

[0127]

[0128] Among them, Γ Evaluator Represents the result of the first loss function, Γ Pos It means that Γ Neg Indicates that Q stru represents the structural score, Q unce Represents the uncertainty score, w represents, S represents the set of positive triples, S' represents the set of negative triples, h represents the elements of the first entity matrix, r represents the elements of the relationship matrix, and t represents the elements of the second entity matrix.

[0129] In some embodiments, the confidence generator is obtained through training, and its second loss function is:

[0130]

[0131] Among them, Γ Conf Represents the result of the second loss function, Q unce Represents the uncertainty score, w represents, S represents the set of positive triples, h represents the elements of the first entity matrix, r represents the elements of the relationship matrix, and t represents the elements of the second entity matrix.

Claims

1. A semantic reasoning method based on uncertainty knowledge graph, characterized in that: include: S1: Use pattern matching to obtain text semantic information; S2: Based on the text semantic information, the triples are vectorized using a linearly transformed knowledge graph to obtain an energy function; S3: Analyze the energy function using a triplet evaluator to obtain a structural score and an uncertainty score; S4: linearly fuse the structural score and the uncertainty score to obtain a linear fusion score; S5: Based on the uncertainty score, the confidence of each triple is calculated using the confidence generator, and the linear fusion score and corresponding confidence of the screened triple are used as the semantic reasoning result to complete the semantic reasoning of the knowledge graph.

2. The semantic reasoning method based on uncertainty knowledge graph according to claim 1 is characterized in that: The S1 includes: Use the original symptom data to build a symptom database; Based on the symptom library, pattern matching is used to extract the original symptom data to obtain text semantic information.

3. The semantic reasoning method based on uncertainty knowledge graph according to claim 1 is characterized in that: The expression of the energy function is: Where E(h,r,t) represents the energy function, (h,r,t) represents a triplet of text semantic information, H represents the first entity matrix, R represents the relationship matrix, and T represents the second entity matrix. represents the square of the norm of the matrix F, h represents the elements of the first entity matrix, r represents the elements of the relationship matrix, and t represents the elements of the second entity matrix; the first entity matrix, the relationship matrix and the second entity matrix belong to text semantic information.

4. The semantic reasoning method based on uncertainty knowledge graph according to claim 1 is characterized in that: The expression of the structural score is: φ stru (E(h,r,t))=p stru ·E(h,r,t)+b stru ; The expression for the uncertainty fraction is: φ unce (E(h,r,t))=p unce ·E(h,r,t)+b unce ; Among them, Q stru represents a structural fraction, e represents a natural constant, φ stru represents the linear transformation of triple energy scores, E(h,r,t) represents the energy function, and p stru The first training parameter representing the structural score, b stru The second training parameter, Q, represents the structural score unce represents the uncertainty fraction, φ unce represents the linear transformation of triple energy fractions, p unce The first training parameter representing the uncertainty score, b unce A second training parameter representing an uncertainty score, h represents an element of the first entity matrix, r represents an element of the relationship matrix, and t represents an element of the second entity matrix.

5. The semantic reasoning method based on uncertainty knowledge graph according to claim 1 is characterized in that: The expression of the linear fusion score is: score=α·Q stru +β·Q unce ; Among them, score represents the linear fusion score, α represents the weight of the structural score, and Q stru represents the structural score, β represents the weight of the uncertainty score, Q unce Represents the uncertainty score.

6. The semantic reasoning method based on uncertainty knowledge graph according to claim 1 is characterized in that: The triplet evaluator is obtained through training, and its first loss function is: C Evaluator =C Pos +C Neg ; Among them, Γ Evaluator Represents the result of the first loss function, Γ Pos represents the loss function of positive triples, Γ Neg represents the loss function for negative triples, Q stru represents the structural score, Q unce represents the uncertainty score, w represents the confidence of the triple, S represents the set of positive triples, S' represents the set of negative triples, h represents the elements of the first entity matrix, r represents the elements of the relationship matrix, and t represents the elements of the second entity matrix.

7. The semantic reasoning method based on uncertainty knowledge graph according to claim 1 is characterized in that: The confidence generator is obtained through training, and its second loss function is: Among them, Γ Conf Represents the result of the second loss function, Q unce represents the uncertainty score, w represents the confidence of the tuple, S represents the set of positive triples, h represents the elements of the first entity matrix, r represents the elements of the relationship matrix, and t represents the elements of the second entity matrix.

Images