Competitive hypothesis decision analysis method and device based on knowledge reasoning

Abstract

The application provides a competitive hypothesis decision analysis method and device based on knowledge reasoning, which comprises the following steps: using a pre-trained language model to encode and decode a problem to be analyzed to obtain a knowledge representation and a hypothesis condition of the problem; obtaining a plurality of candidate evidences related to the problem, extracting a knowledge representation feature from each candidate evidence, fusing the candidate evidence and the hypothesis condition, obtaining an effective evidence, and obtaining an argumentation matrix between the effective evidence and the hypothesis condition; based on the evidence feature and the hypothesis condition feature after the knowledge representation, the hypothesis condition, the effective evidence and the argumentation matrix are optimized through iterative reasoning; according to the optimized matrix feature, a key evidence is calculated from the optimized effective evidence, argumentation parameters of the key evidence are reversely calculated, and a reverse index of the key evidence having a greater influence degree on the final conclusion of the problem is obtained; and the problem is analyzed according to the final hypothesis condition, the key evidence and the reverse index.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and natural language processing, and in particular to a competitive hypothesis decision analysis method and apparatus based on knowledge reasoning. Background Technology

[0002] The competing hypothesis is an eight-step analysis method based on important findings from cognitive psychology, decision analysis and other scientific methods, namely: (1) identify the hypothesis; (2) list the evidence; (3) create a matrix diagram; (4) improve the matrix; (5) question the hypothesis; (6) question the evidence; (7) report the conclusion; (8) find the reverse indicators for future observation.

[0003] The competitive hypothesis method has the advantage of being unbiased and requires the tester to establish a hypothesis matrix, with the entire process being traceable. However, because the tester must establish and improve the hypothesis matrix during the analysis, and the data size of the hypothesis matrix is ​​proportional to the hypotheses and evidence data, the competitive hypothesis method involves a large data size, is time-consuming, and is prone to errors during computation and analysis when dealing with large-scale data. Summary of the Invention

[0004] In view of the above problems, the present invention provides a competitive hypothesis decision analysis method and apparatus based on knowledge reasoning.

[0005] According to the first aspect of this disclosure, a competitive hypothesis decision analysis method based on knowledge reasoning is provided, comprising: using a pre-trained language model to encode and decode the problem to be analyzed, obtaining a knowledge representation and hypothesis conditions of the problem; acquiring multiple candidate pieces of evidence related to the problem to form an evidence set, extracting features from the knowledge representation of each candidate piece of evidence, fusing the candidate evidence with the hypothesis conditions, calculating the degree of argumentation of the candidate evidence for the hypothesis conditions, obtaining valid evidence, and an argumentation matrix between the valid evidence and the hypothesis conditions; optimizing the hypothesis conditions, valid evidence, and argumentation matrix through iterative reasoning based on the evidence features and hypothesis condition features after knowledge representation; calculating key evidence from the optimized valid evidence based on the optimized matrix features, reversely calculating the argumentation parameters of the key evidence, and obtaining a reverse indicator that the key evidence has a significant impact on the final conclusion of the problem; and performing decision analysis on the problem based on the final hypothesis conditions, key evidence, and reverse indicator.

[0006] According to embodiments of this disclosure, the pre-trained language model is a BERT pre-trained language model, which includes an encoder and a decoder.

[0007] According to embodiments of this disclosure, encoding and decoding the problem to be analyzed includes: pre-constructing a BERT pre-trained language model; inputting the problem into the BERT pre-trained language model based on the text features of the problem, encoding it using an attention mechanism to obtain a knowledge representation of the problem, and then decoding the features of the problem using an attention mechanism to obtain possible assumptions about the problem.

[0008] According to embodiments of this disclosure, knowledge representation features are extracted for each candidate piece of evidence, and the candidate evidence and hypothesis conditions are fused. Specifically, this includes: extracting text features of each candidate piece of evidence using a BERT pre-trained language model; and fusing the text features of each candidate piece of evidence with the hypothesis conditions based on an attention mechanism to obtain a knowledge representation of the candidate evidence containing hypothesis condition information.

[0009] According to embodiments of this disclosure, the degree of argumentation of candidate evidence for the assumptions is categorized into five levels: unbiased, supportive, strongly supportive, opposing, and strongly opposing; the degree of importance and relevance of candidate evidence to the issue is categorized into three levels: high, moderate, and low.

[0010] According to embodiments of this disclosure, iterative reasoning specifically includes: performing a classification task on the features of two valid pieces of evidence through at least two fully connected layers to determine whether there is redundancy between the two valid pieces of evidence; if so, deleting the one of the two valid pieces of evidence with the lower ranking based on importance and relevance; fusing the features of one valid piece of evidence and other candidate evidence of invalid evidence in the evidence set with the assumption conditions through an attention mechanism, and then performing a classification task with the fully connected layers to determine whether there is an inference relationship between the other candidate evidence and the valid evidence under the assumption conditions; if so, adding the other candidate evidence to the valid evidence set.

[0011] According to embodiments of this disclosure, key evidence is calculated from optimized valid evidence based on optimized matrix features, including: obtaining the optimized matrix features of multiple valid pieces of evidence, calculating the weighted mean square error of the matrix features of each valid piece of evidence, and identifying at least one valid piece of evidence with the largest weighted mean square error of the matrix features as key evidence.

[0012] According to embodiments of this disclosure, the reverse calculation of the argumentation parameters of key evidence includes: using the knowledge representation of the optimized assumptions as the knowledge representation of the assumption fusion problem, repeatedly calculating the degree of argumentation of candidate evidence for the assumptions until the key evidence is calculated from the optimized valid evidence, and reversely calculating the argumentation parameters of the key evidence.

[0013] The second aspect of this disclosure provides a competitive hypothesis decision analysis device based on knowledge reasoning, comprising: an encoding / decoding module for encoding and decoding the problem to be analyzed using a pre-trained language model to obtain a knowledge representation and hypothesis conditions of the problem; a valid evidence acquisition module for acquiring multiple candidate pieces of evidence related to the problem to form an evidence set, extracting features from the knowledge representation of each candidate piece of evidence, fusing the candidate evidence with the hypothesis conditions, calculating the degree of argumentation of the candidate evidence for the hypothesis conditions, obtaining valid evidence, and an argumentation matrix between the valid evidence and the hypothesis conditions; an iterative optimization module for optimizing the hypothesis conditions, valid evidence, and argumentation matrix through iterative reasoning based on the evidence features and hypothesis condition features after knowledge representation; a key evidence acquisition module for calculating key evidence from the optimized valid evidence based on the optimized matrix features, reversely calculating the argumentation parameters of the key evidence, and obtaining a reverse indicator that the key evidence has a significant impact on the final conclusion of the problem; and a decision analysis module for performing decision analysis on the problem based on the final hypothesis conditions, key evidence, and reverse indicator.

[0014] Compared with the prior art, the competitive hypothesis decision analysis method and apparatus based on knowledge reasoning provided by the present invention have at least the following beneficial effects:

[0015] (1) This invention uses the BERT pre-trained language model to represent the knowledge of the text, which can greatly enrich the semantic information of the text;

[0016] (2) Knowledge representation of the analysis problem, assumptions, and evidence obtained from natural language processing tools. Based on the knowledge representation of the analysis problem, the corresponding assumptions can be obtained through feature encoding and decoding, which also enriches the contextual semantic representation of the assumptions in the context of the problem;

[0017] (3) Based on the knowledge representation of assumptions and evidence, attention mechanisms can be used to enrich the semantic and argumentative information between evidence and assumptions;

[0018] (4) For knowledge reasoning tasks, evidence-based knowledge representation can better determine the relationship between evidence, enrich the contextual information of evidence, and optimize the evidence set by fusing feature information. At the same time, knowledge representation of assumptions can also enrich the information on the ability of evidence to prove assumptions by using attention mechanisms, thereby enabling efficient and accurate decision analysis of the problem. Attached Figure Description

[0019] The above and other objects, features and advantages of the present invention will become more apparent from the following description of embodiments of the invention with reference to the accompanying drawings, in which:

[0020] Figure 1A flowchart illustrating a knowledge-based reasoning-based competitive hypothesis decision analysis method according to an embodiment of the present disclosure is shown schematically.

[0021] Figure 2 A flowchart illustrating the process of obtaining assumptions according to embodiments of the present disclosure is shown schematically.

[0022] Figure 3 A flowchart illustrating a feature fusion process according to an embodiment of the present disclosure is shown schematically;

[0023] Figure 4 A flowchart illustrating an iterative reasoning process according to an embodiment of the present disclosure is shown schematically;

[0024] Figure 5 A flowchart illustrating the key evidence determination process according to embodiments of this disclosure is shown schematically;

[0025] Figure 6 A flowchart illustrating the reverse computation process according to an embodiment of the present disclosure is shown schematically.

[0026] Figure 7 A block diagram of a knowledge-based reasoning competitive hypothesis decision analysis apparatus according to an embodiment of the present disclosure is shown schematically. Detailed Implementation

[0027] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the invention. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the invention for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0028] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0029] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0030] When using expressions such as "at least one of A, B, and C", they should generally be interpreted in accordance with the meaning that is commonly understood by a person skilled in the art (e.g., "a system having at least one of A, B, and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B, and C, etc.).

[0031] Figure 1 A flowchart illustrating a knowledge-based reasoning-based competitive hypothesis decision analysis method according to an embodiment of the present disclosure is shown.

[0032] like Figure 1 As shown, the knowledge-based reasoning-based competitive hypothesis decision analysis method of this embodiment may include operations S110 to S150.

[0033] In the S110 operation, a pre-trained language model is used to encode and decode the problem to be analyzed, thereby obtaining the knowledge representation and assumptions of the problem.

[0034] Preferably, the pre-trained language model can be a BERT pre-trained language model, which is an autoencoder pre-trained language model with bidirectional feature representation, including an encoder and a decoder. Using BERT as the word embedding representation, the question q to be analyzed is input into the BERT pre-trained language model for bidirectional encoding.

[0035] BERT swept 11 NLP tasks through pre-training and fine-tuning, mainly because its model architecture is based on the Transformer model and uses a masked language model. Compared to traditional RNN models, the Transformer model is more efficient and can capture longer-range dependencies. It also overcomes the limitation of RNN models not being able to perform parallel computation because the Transformer model uses a self-attention mechanism, which eliminates the dependency of the number of operations required to compute contextual information on the distance in the sequence. Furthermore, the masked language model allows each encoding to be indirectly represented in multiple layers of context, achieving deep bidirectional representation. This framework enables BERT to capture more meaningful contextual information than other pre-trained language models, thus allowing BERT's representational capabilities to represent the textual features of the task.

[0036] In operation S120, multiple candidate pieces of evidence related to the problem are obtained to form an evidence set. For each candidate piece of evidence, knowledge representation is performed to extract features. The candidate evidence is then fused with the assumptions to calculate the degree of argumentation of the candidate evidence to the assumptions, resulting in valid evidence and the argumentation matrix between the valid evidence and the assumptions.

[0037] In operation S130, based on the evidence features and hypothesis features after knowledge representation, the hypothesis, valid evidence, and argument matrix are optimized through iterative reasoning.

[0038] In operation S140, based on the optimized matrix characteristics, key evidence is calculated from the optimized valid evidence. The argumentation parameters of the key evidence are then calculated in reverse to obtain the inverse indicators that show the significant influence of the key evidence on the final conclusion of the problem.

[0039] In operation S150, decision analysis is performed on the problem based on the final assumptions, key evidence, and reverse indicators.

[0040] Through the embodiments of this disclosure, evidence-based knowledge representation, through feature information fusion, can better determine the relationship between evidence, enrich the contextual information of evidence, and optimize the evidence set; at the same time, with the knowledge representation of the assumptions, under the use of the attention mechanism, it can also enrich the representation information of the evidence' ability to prove the assumptions, thereby enabling efficient and accurate decision analysis of the problem to be analyzed.

[0041] Figure 2 A flowchart illustrating the process of obtaining assumptions according to an embodiment of the present disclosure is shown.

[0042] like Figure 2 As shown, in the above operation S110, the problem to be analyzed is encoded and decoded, which may specifically include operations S1101 to S1102.

[0043] In operation S1101, a BERT pre-trained language model is pre-built.

[0044] In operation S1102, based on the textual features of the question, the question is input into the BERT pre-trained language model, encoded using the attention mechanism to obtain the knowledge representation of the question, and then the attention mechanism is used to decode the features of the question to obtain the possible assumptions of the question.

[0045] Based on the textual features of the problem, both the encoder and decoder utilize attention mechanisms to obtain possible assumptions about the problem. Specifically, the knowledge representation and assumptions of the problem are calculated according to the following formula:

[0046] d q =V q r q σ(Ψ(K q r q Q q r q )) (1)

[0047] r h，q =V h，qr h σ(Ψ(K h，q r h Q h，q d q (2)

[0048] In formula (1), σ is the softmax function; Ψ is the multilayer perceptron calculation method; K q V q Q q These are the weight matrices of the encoder for problem q under the self-attention mechanism; r q It is a characteristic representation of problem q; d q It is the knowledge representation of the features of problem q after passing through the self-attention mechanism; in formula (2), K h，q V h，q Q h，q These are the weight matrices of the decoder's assumption h and the problem q under the self-attention mechanism, respectively; r h It is a characteristic representation of the assumption condition h; r h，q It is the knowledge representation of the fusion problem q based on the assumptions h.

[0049] Thus, through the attention mechanism, a knowledge representation is obtained that integrates the assumptions and the contextual information of the current problem.

[0050] Specifically, the problem to be analyzed could be, for example, "Mapo Tofu is a dish of which cuisine?" Decision analysis can be performed on this problem, and through the above operation S110, eight possible assumptions can be obtained, such as those from Shandong, Sichuan, Guangdong, Jiangsu, Fujian, Zhejiang, Hunan, and Anhui.

[0051] Figure 3 A flowchart illustrating a feature fusion process according to an embodiment of the present disclosure is shown schematically.

[0052] like Figure 3 As shown, in the above operation S120, knowledge representation is performed to extract features for each candidate piece of evidence, and the candidate evidence and the hypothesis conditions are fused. Specifically, this may include operations S1201 to S1202.

[0053] In operation S1201, the text features of each candidate piece of evidence are extracted using a BERT pre-trained language model.

[0054] In operation S1202, the text features of each candidate piece of evidence are fused with the hypothetical conditions based on the attention mechanism to obtain a knowledge representation of the candidate evidence containing the hypothetical condition information.

[0055] Specifically, in the above operation S120, after feature fusion, the valid evidence and the argument matrix between the valid evidence and the hypothesis are calculated according to the following formula:

[0056] r e，h =V e，h r e σ(Φ(K e，h r e ,Q e，h r h，q )) (3)

[0057]

[0058] r e，E =V e，E r e σ(Φ(K e，E r e ,Q e，E r E )) (5)

[0059] r e =W e，h r e，h +W e，E r e，E (6)

[0060] w eh =σ(W w [r e ;r h，q ]) (7)

[0061] d(e,h)=W d w eh (8)

[0062] I eq ,r eq =σ((W i ,W r )[r e ;d q ]) (9)

[0063] di(e,q)=W i I eq (10)

[0064] dr(e,q)=W r r eq (11)

[0065]

[0066] WI E′，H ={i eq |e∈E′} (13)

[0067] WR E′，H ={r eq|eeE′} (14)

[0068] In formula (3), Φ is the cosine similarity calculation method; e∈E, e is a candidate piece of evidence in the evidence set E; K e，h V e，h Q e，h These are the weight matrices for the self-attention mechanism applied to candidate evidence e under hypothesis h; r e It is the characteristic representation of candidate evidence e; r e，h It is the knowledge representation of candidate evidence e after fusing with hypothesis h; in formula (4), e j It refers to each candidate piece of evidence enumerated in evidence set E; It is the feature representation of each candidate piece of evidence; r E The knowledge representation of evidence set E; in formula (5), K e，E V e，E Q e，E These are the weight matrices of the evidence set E when performing a self-attention mechanism on candidate evidence e; r e，E It is the knowledge representation of candidate evidence e after fusing information from other evidence in evidence set E; in formula (6), W e，h W e，E It is the weight matrix of candidate evidence e after fusing the information of hypothesis h and evidence set E. Based on formula (6), the feature representation r of candidate evidence e is updated. e This allows it to integrate the knowledge representation of the assumption condition h and the feature information of the evidence set E; in formula (7), W w It is the weight matrix of the candidate evidence's ability to prove the hypothesis, w eh This indicates the persuasive power of candidate evidence e for hypothesis h; in formula (8), W d A 5-level classification matrix representing the degree of argumentation. d(e, h) is the degree of argumentation of candidate evidence e for hypothesis h; in formula (9), W i W r These are the weight matrices representing the importance and relevance of candidate evidence e to question q; i eq r eq Let W represent the importance and relevance of candidate evidence e to question q, respectively; in formula (10), W i A three-level classification matrix representing the importance of evidence to the issue. di(e, q) represents the importance of candidate evidence e to question q; in formula (11), W r A three-level classification matrix representing the degree of relevance of evidence to the issue. dr(e, q) represents the degree of relevance of candidate evidence e to question q; in formula (12), W E′，HLet represent the argument matrix of the valid evidence set E′ to the hypothesis set H; α is a hyperparameter used to judge the degree of argument validity of candidate evidence e; in formula (13), WI E′，H Represents the importance matrix of the valid evidence set E′ to the hypothesis set H; in formula (14), WR E′，H Let E' represent the correlation matrix between the valid evidence set E′ and the hypothesis condition set H.

[0069] In this embodiment of the disclosure, the degree of argumentation of candidate evidence for the hypothesis can specifically include five categories: unbiased, supportive, strongly supportive, opposed, and strongly opposed. Wherein, N represents unbiased, C represents supportive, CC represents strongly supportive, I represents opposed, and II represents strongly opposed. The formula for calculating the degree of argumentation f(e, h) of candidate evidence e for hypothesis h is as follows:

[0070]

[0071] Based on the above formulas (3)-(14), the importance and relevance of candidate evidence e to question q can be categorized into three types: high, medium, and low. Here, H represents high importance, M represents medium importance, and L represents low importance. The formulas for calculating importance fi(e, q) and relevance fr(e, q) are as follows:

[0072]

[0073]

[0074] Finally, the argument score for each candidate piece of evidence e against hypothesis h is calculated, also known as the Score. A higher Score indicates a greater likelihood of proof. The formula for calculating the Score is as follows:

[0075]

[0076]

[0077]

[0078] Score e，h =w eh h(e,h)ln hi(e,h)ln hr(e,h) (21)

[0079] Through the above operation S120, by using the attention mechanism to fuse the characteristics of evidence and the characteristics of hypotheses, effective hypotheses, evidence, and the persuasive power, importance, and relevance of the evidence can be obtained. Based on the above example, the effective hypotheses, evidence, and the persuasive power, importance, and relevance of the evidence are shown in Table 1 below.

[0080] Table 1

[0081]

[0082] Among them, valid evidence E1: Mapo tofu originated in the first year of the Tongzhi reign of the Qing Dynasty (1862). There was a shop called "Chen Xingsheng Restaurant" near Wanfu Bridge in Chengdu. The owner, Chen Chunfu (Chen Senfu), died early, and the small restaurant was run by his wife. The female owner had a slightly pockmarked face and was known as "Chen Mapo".

[0083] Valid evidence E2: Early Mapo Tofu used vegetable oil and beef. The cooking method was to first fry a large spoonful of vegetable oil in a pan until cooked, then add a handful of chili flakes, followed by beef. The beef was cooked until dry and tender before adding fermented black beans. Tofu was then added, a little water was added and stirred a few times to mix well. Finally, the pan was covered and simmered over low heat until the sauce thickened. Before serving, ground Sichuan peppercorns were sprinkled on top.

[0084] Valid evidence E3: The main ingredients are seasonings and tofu. The main ingredients include tofu, minced beef (pork can also be used), chili peppers, and Sichuan peppercorns. The numbing sensation comes from Sichuan peppercorns, and the spiciness comes from chili peppers. This dish highlights the characteristics of "numbing and spicy".

[0085] Valid evidence E4: The poem "Jin Cheng Zhu Zhi Ci" says, "Chen Ma Po's name is still widely known, her tofu is the most exquisite, the curtains sway by Wanfu Bridge, and the spring wine is drunk with the gentleman." "Chengdu Tonglan" records that Chen Ma Po Tofu was listed as a famous food in Chengdu at the end of the Qing Dynasty.

[0086] Assumption H1: Shandong cuisine. Assumption H2: Sichuan cuisine. Assumption H3: Cantonese cuisine. Assumption H4: Jiangsu cuisine. Assumption H5: Fujian cuisine. Assumption H6: Zhejiang cuisine. Assumption H7: Hunan cuisine. Assumption H8: Anhui cuisine.

[0087] Figure 4 A flowchart illustrating an iterative reasoning process according to an embodiment of the present disclosure is shown schematically.

[0088] like Figure 4 As shown, in the above operation S130, iterative reasoning may specifically include operations S1301 to S1302.

[0089] In operation S1301, the features of two valid pieces of evidence are processed through at least two fully connected layers to perform a classification task, and it is determined whether there is redundancy between the two valid pieces of evidence. If so, the one with the lower ranking in terms of importance and relevance is deleted.

[0090] In operation S1302, a valid piece of evidence and other candidate evidence that are not valid evidence in the evidence set are respectively fused with the hypothesis through the attention mechanism, and then classified with the fully connected layer to determine whether there is an inference relationship between the other candidate evidence and the valid evidence under the hypothesis. If so, the other candidate evidence is added to the valid evidence set.

[0091] Specifically, after obtaining the possible hypotheses, evidence valid for the current hypotheses, and the argument matrix through the above operation S120, the above operation S130 iteratively updates and optimizes the valid evidence and argument matrix through iterative reasoning. The specific calculation formula is as follows:

[0092]

[0093]

[0094] r h，E =V h，E r′ h σ(Ψ(K h，E r′ h Q h，E r E )) (twenty four)

[0095] r h =W h，q r h，q +W h，E r h，E (25)

[0096] In formula (22), W ee and W′ ee These are the i-th valid pieces of evidence, e i And the jth valid evidence e j The weight matrix used in merge calculations; Indicates the i-th valid piece of evidence e i And the jth valid evidence e j The characteristic representation; argmax is a function that evaluates the parameters of a function; redu(e i e j There are two valid pieces of evidence. i e j The classification results are used to determine whether merging is necessary. If the result is 1, it indicates that there is information redundancy between the two valid pieces of evidence. In this case, the valid piece of evidence with the lower ranking based on importance and relevance is deleted. In formula (23), W″ ee It is the weight matrix for the additional task of evidence reasoning; These are other candidate pieces of evidence that are not valid evidence in evidence set E; It is valid evidence. i Other candidate evidence The classification result determines whether there is a reasoning connection. If the result is 1, it indicates other candidate evidence. In valid evidence e i The following has inference value, so it is added to the valid evidence set E′; in formula (24), K h，E V h，E Q h，E These are the weight matrices of the evidence set E when applying a self-attention mechanism to hypothesis h; r h，E It is a knowledge representation of the hypothesis h that incorporates information from the evidence set E; in formula (25), W h，q W h，E It is the weight matrix that is fused again after the assumption condition h is separately fused with the information from the problem q and the evidence set E, and the feature representation r of the assumption condition h is updated. h This enables the knowledge representation of the hypothesis h, which integrates the feature information of the problem q and the evidence set E.

[0097] Then, the latest r h As r h，q Substitute the values ​​into formulas (3) to (25) to complete one iteration of calculation. Obtain the updated valid evidence and argument matrix according to formulas (12) to (14). Delete redundant evidence according to the result of formula (18) and add valuable valid evidence according to the result of formula (19). Through multiple rounds of iterative updates, the results of the assumptions, evidence, and argumentation capability matrix tend to stabilize, at which point the update is considered complete.

[0098] Through the above operation S130, based on the above example, the final optimized assumptions, valid evidence, and the persuasive power, importance, and relevance of the valid evidence are shown in Table 2 below.

[0099] Table 2

[0100] H1 H2 H3 H4 H5 H6 H7 H8 importance Relevance E1’ II CC II II II II I II H H E2’ N CC N N N N N N H H E3’ N CC N N N N N N H H E4’ N CC N N N N N N H H Score -47.7 -4.83 -52.85 -58.16 -58.73 -47.51 -39.51 -58.31

[0101] Among them, the optimized valid evidence E1': Mapo Tofu is a traditional famous dish in Sichuan.

[0102] Optimized valid evidence E2': Sichuan cuisine is one of the four major traditional cuisines of the Han Chinese, one of the eight major cuisines of China, and a culmination of Chinese culinary traditions. Sichuan cuisine is divided into local Sichuan cuisine and Shanghai-style Sichuan cuisine. Local Sichuan cuisine includes Sichuan-style dishes, noodles and snacks, hot pot, etc.

[0103] Optimized valid evidence E3': Mapo tofu originated in the first year of the Tongzhi reign of the Qing Dynasty (1862). Near Wanfu Bridge in Chengdu, there was a shop originally named "Chen Xingsheng Restaurant". The owner, Chen Chunfu (Chen Senfu), died early, and the small restaurant was run by his wife. The female owner had a slightly pockmarked face and was known as "Chen Mapo".

[0104] Optimized valid evidence E4': Chengdu, abbreviated as "Rong", also known as Rongcheng or Jincheng, is a prefecture-level city, provincial capital, sub-provincial city, megacity, core city of the Chengdu-Chongqing Twin-City Economic Circle, a national important high-tech industrial base, commercial and logistics center, comprehensive transportation hub, and an important central city in western China.

[0105] Assumption H1: Shandong cuisine. Assumption H2: Sichuan cuisine. Assumption H3: Cantonese cuisine. Assumption H4: Jiangsu cuisine. Assumption H5: Fujian cuisine. Assumption H6: Zhejiang cuisine. Assumption H7: Hunan cuisine. Assumption H8: Anhui cuisine.

[0106] Figure 5 A flowchart illustrating the key evidence determination process according to an embodiment of this disclosure is shown.

[0107] like Figure 5 As shown, in the above operation S140, key evidence is calculated from the optimized valid evidence based on the optimized matrix features, which may specifically include operations S1401 to S1402.

[0108] In operation S1401, the matrix features of multiple valid pieces of evidence after optimization are obtained, and the weighted mean square error of the matrix features of each valid piece of evidence is calculated.

[0109] In operation S1402, at least one valid piece of evidence with the largest weighted mean square error of the matrix features is identified as key evidence.

[0110] Therefore, the weighted mean squared error is calculated for the argument matrix of each optimized valid piece of evidence. The larger the result, the more prominent the criticality of that valid piece of evidence. The formula for calculating the weighted mean squared error K(e) of the matrix features of each valid piece of evidence is as follows:

[0111]

[0112] in, The valid evidence e in the above formula (21) is related to the j-th assumption h. j The score for the argument.

[0113] Through the above operation S1402, based on the above example, the optimized effective evidence E1' has the highest weighted mean square error calculation result, proving that effective evidence E1' is the key evidence in the current analysis results.

[0114] Figure 6 A flowchart illustrating the reverse computation process according to an embodiment of the present disclosure is shown schematically.

[0115] like Figure 6As shown, in the above operation S140, the argumentation parameters of key evidence are calculated in reverse, which may also include operation S1403.

[0116] In operation S1403, the knowledge representation of the optimized assumptions is used as the knowledge representation of the assumption fusion problem. The operation steps of repeatedly calculating the degree of argumentation of candidate evidence for the assumptions until the key evidence is calculated from the optimized valid evidence are used to calculate the argumentation parameters of the key evidence in reverse.

[0117] Therefore, by modifying the relative probability of the assumptions, and based on the linear computational properties, the matrix characteristics of the key evidence are obtained through reverse calculation.

[0118] Through the above operation S1403, based on the above example, the argument scores of the two assumptions H2 and H7 in the calculation results are tied for the highest. Therefore, the argument scores of assumptions H1 and H7 are set to be the same. While keeping other calculation results unchanged, the calculation is performed in reverse according to the calculation formulas (7)-(26). and Ultimately, two inverse indicators were obtained. and

[0119] When the persuasive power of valid evidence E1' for hypotheses H2 and H7 is respectively and When the valid evidence E1' is substituted into formula (15), the degree of argumentation of the assumptions H2 and H7 is II and CC, respectively. When other calculation results remain unchanged, the final argumentation scores of the assumptions H2 and H7 are consistent, which means that they have an impact on the final conclusion.

[0120] Finally, through the above operation S150, based on the final valid evidence and assumptions, and by integrating the reverse indicators of key evidence, a decision analysis is conducted on the problem to obtain all assumptions and their relative probabilities, key evidence, and reverse indicators for future observation.

[0121] Based on the above example, a decision analysis of the question "Mapo Tofu is a dish of which cuisine?" yields a result that most likely points to hypothesis H2, and the key evidence obtained during the reasoning process is E1. In the argumentation of key evidence E1, if the degree of argumentation for hypothesis H2 and H7 is II (strongly opposed) and CC (strongly supported) respectively, it may affect the result.

[0122] Therefore, the knowledge-based competitive hypothesis decision analysis method provided in this disclosure can perform efficient and accurate decision analysis on problems.

[0123] Figure 7A block diagram of a knowledge-based reasoning competitive hypothesis decision analysis apparatus according to an embodiment of the present disclosure is shown schematically.

[0124] like Figure 7 As shown, the knowledge-based reasoning competitive hypothesis decision analysis device 700 of this embodiment includes an encoding / decoding module 710, an effective evidence acquisition module 720, an iterative optimization module 730, a key evidence acquisition module 740, and a decision analysis module 750.

[0125] The encoding / decoding module 710 is used to encode and decode the problem to be analyzed using a pre-trained language model to obtain the knowledge representation and assumptions of the problem.

[0126] The effective evidence acquisition module 720 is used to acquire multiple candidate pieces of evidence related to the problem to form an evidence set, extract features from each candidate piece of evidence using knowledge representation, fuse the candidate evidence with the assumptions, calculate the degree of argumentation of the candidate evidence to the assumptions, obtain the effective evidence, and the argumentation matrix between the effective evidence and the assumptions.

[0127] The iterative optimization module 730 is used to optimize the hypothesis, valid evidence, and argument matrix through iterative reasoning based on the evidence features and hypothesis features after knowledge representation.

[0128] The key evidence acquisition module 740 is used to calculate key evidence from the optimized valid evidence based on the optimized matrix features, and to reverse calculate the argumentation parameters of the key evidence to obtain reverse indicators that indicate the degree to which the key evidence has a significant impact on the final conclusion of the issue.

[0129] The decision analysis module 750 is used to perform decision analysis on the problem based on the final assumptions, key evidence, and adverse indicators.

[0130] It should be noted that the embodiments of the device section are similar to those of the method section, and the technical effects achieved are also similar. For specific details, please refer to the above-mentioned method embodiment section, which will not be repeated here.

[0131] According to embodiments of this disclosure, any plurality of modules among the encoding / decoding module 710, the valid evidence acquisition module 720, the iterative optimization module 730, the key evidence acquisition module 740, and the decision analysis module 750 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules can be combined with at least part of the functionality of other modules and implemented in one module. According to embodiments of this disclosure, at least one of the encoding / decoding module 710, the valid evidence acquisition module 720, the iterative optimization module 730, the key evidence acquisition module 740, and the decision analysis module 750 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the encoding / decoding module 710, the valid evidence acquisition module 720, the iterative optimization module 730, the key evidence acquisition module 740, and the decision analysis module 750 may be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0132] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0133] Those skilled in the art will understand that the features described in the various embodiments and / or claims of the present invention can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments and / or claims of the present invention can be combined or combined in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or combinations fall within the scope of the present invention.

[0134] The embodiments of the present invention have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of the invention. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. The scope of the invention is defined by the appended claims and their equivalents. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the invention, and all such substitutions and modifications should fall within the scope of the invention.

Claims

1. A competitive hypothesis decision analysis method based on knowledge reasoning, characterized in that, include: Using a pre-trained language model, the problem to be analyzed is encoded and decoded to obtain the knowledge representation and assumptions of the problem; Multiple candidate pieces of evidence related to the problem are obtained to form an evidence set. Each candidate piece of evidence is represented by knowledge and features are extracted. The candidate pieces of evidence are fused with the hypothesis. The degree of argumentation of the candidate pieces of evidence to the hypothesis is calculated to obtain valid evidence and the argumentation matrix between the valid evidence and the hypothesis. Based on the evidence features and assumption features after knowledge representation, the assumptions, valid evidence, and argument matrix are optimized through iterative reasoning. Based on the optimized matrix features, key evidence is calculated from the optimized valid evidence. The argumentation parameters of the key evidence are then calculated in reverse to obtain a reverse indicator that the key evidence has a significant impact on the final conclusion of the issue. as well as Decision analysis is conducted on the problem based on the final assumptions, key evidence, and reverse indicators.

2. The method according to claim 1, wherein, The pre-trained language model is a BERT pre-trained language model, which includes an encoder and a decoder.

3. The method according to claim 2, wherein, Encoding and decoding the problem to be analyzed includes: Pre-build BERT pre-trained language models; Based on the textual features of the question, the question is input into the BERT pre-trained language model, encoded using an attention mechanism to obtain a knowledge representation of the question, and then the features of the question are decoded using an attention mechanism to obtain possible assumptions about the question.

4. The method according to claim 2, wherein, The step of extracting features from each candidate piece of evidence using knowledge representation and fusing the candidate evidence with the hypothesis conditions specifically includes: The text features of each candidate piece of evidence are extracted using the BERT pre-trained language model. Based on the attention mechanism, the text features of each candidate piece of evidence are fused with the hypothesis conditions to obtain a knowledge representation of the candidate evidence containing hypothesis condition information.

5. The method according to claim 1, wherein, The degree of argumentation of the candidate evidence for the hypothesis is divided into five categories: unbiased, supportive, strongly supportive, opposed, and strongly opposed. The importance and relevance of the candidate evidence to the issue are categorized into three levels: high, medium, and low.

6. The method according to claim 1, wherein, The iterative reasoning specifically includes: The features of two valid pieces of evidence are passed through at least two fully connected layers to perform a classification task, and it is determined whether there is redundancy between the two valid pieces of evidence. If so, the one of the two valid pieces of evidence with the lower ranking in terms of importance and relevance is deleted. One valid piece of evidence and other candidate evidence that are not valid evidence in the evidence set are respectively fused with features by an attention mechanism and the assumption conditions, and then the full-connected layer is used to perform a classification task to determine whether there is an inference relationship between the other candidate evidence and the valid evidence under the assumption conditions. If so, the other candidate evidence is added to the valid evidence set.

7. The method according to claim 1, wherein, The process of calculating key evidence from optimized valid evidence based on optimized matrix features includes: Obtain the optimized matrix features of multiple valid pieces of evidence, and calculate the weighted mean square error of the matrix features of each valid piece of evidence; The most valid piece of evidence with the largest weighted mean square error of the matrix features is identified as the key evidence.

8. The method according to claim 1, wherein, The reverse calculation of the argumentation parameters of the key evidence includes: The knowledge representation of the optimized assumptions is used as the knowledge representation of the assumption fusion problem. The degree of argumentation of the candidate evidence for the assumptions is repeatedly calculated until the operation step of calculating the key evidence from the optimized valid evidence is completed. The argumentation parameters of the key evidence are calculated in reverse.

9. A competitive hypothesis decision analysis device based on knowledge reasoning, characterized in that, include: The encoding / decoding module is used to encode and decode the problem to be analyzed using a pre-trained language model to obtain the knowledge representation and assumptions of the problem. The effective evidence acquisition module is used to acquire multiple candidate pieces of evidence related to the problem to form an evidence set, extract features from each candidate piece of evidence using knowledge representation, fuse the candidate evidence with the hypothesis, calculate the degree of argumentation of the candidate evidence to the hypothesis, obtain effective evidence, and the argumentation matrix between the effective evidence and the hypothesis. The iterative optimization module is used to optimize the assumptions, valid evidence, and argument matrix through iterative reasoning based on the evidence features and assumption features after knowledge representation. The key evidence acquisition module is used to calculate key evidence from optimized valid evidence based on optimized matrix features, reverse calculate the argumentation parameters of the key evidence, and obtain a reverse indicator that the key evidence has a significant impact on the final conclusion of the issue. as well as The decision analysis module is used to perform decision analysis on the problem based on the final assumptions, key evidence, and reverse indicators.

Images