Machine learning-based rule information extraction method, system and intelligent terminal

Abstract

The application discloses a rule information extraction method and system based on machine learning and an intelligent terminal. The method combines a domain pre-training BERT model and a rule-based natural language processing method, can extract multi-layer business process related semantic information from regulatory documents, and constructs a rule library based on the semantic information. The rule library constructed by the application can be used for business process compliance checking of an enterprise, so that the enterprise can detect business process violations, prevent potential violations in subsequent process execution, and save energy, time and financial resources of the enterprise.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, and in particular to a method, system, and smart terminal for extracting rule information based on machine learning. Background Technology

[0002] Business Process Management (BPM) provides a framework to support the business process lifecycle, from modeling, configuration, and development to evaluation. Business Process Compliance (BPC) is a crucial part of BPM, measuring whether an organization's business processes comply with all applicable laws, regulations, guidelines, and standards. Regulatory documentation refers to the applicable laws, regulations, guidelines, and standards that business processes must adhere to; examples include the Health Insurance Portability and Accountability Act (HIPAA), the Sarbanes-Oxley Act (SOX), and the General Data Protection Regulation (GDPR). If a company's business processes violate these regulations, it may lose investor confidence, face fines, and even criminal charges. Therefore, compliance with rules from various sources is essential for every organization to avoid substantial fines and improve business process transparency. However, in practice, reviewing and ensuring the consistency of an organization's business processes with regulatory documentation remains largely manual. Furthermore, the cost of manual review can increase significantly due to the constant evolution of regulatory documents.

[0003] Business process compliance checks can occur at different stages of the process lifecycle. During design, the process model is checked to ensure it conforms to a set of constraints. During runtime, the executing process instance is continuously monitored to detect and even predict violations. Finally, during the review phase, process compliance is checked against system-generated log files upon completion of the process instance execution. Design-phase compliance checks are a priority option because detecting any non-compliance at an early stage is crucial for saving the company effort, time, and financial resources. Several business process compliance methods based on process models have been proposed in the prior art. However, the constraints used by these methods are not derived from regulatory documents but rather from a set of pre-assumed constraint rules expressed using formal logic such as Linear Time Logic (LTL), Event Calculus (EC), or Complex Event Processing (CEP). Therefore, these methods cannot directly use regulatory documents for compliance checks.

[0004] However, regulatory documents are unstructured texts in natural language form. To achieve business process compliance checks, it is necessary to extract structured rules that the business processes must follow. Although some methods aim to extract process descriptions or compliance constraints from documents, formal legal norms require significant manual work and complex modeling languages ​​to obtain the semantics of regulatory documents. Therefore, how to extract structured business process rules from diverse regulatory documents to provide the preconditions for automated business process compliance checks is a pressing technical problem that needs to be solved. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention provides a method, system and smart terminal for extracting rule information based on machine learning, which can effectively solve the above problems.

[0006] The specific technical solution adopted in this invention is as follows:

[0007] In a first aspect, the present invention provides a method for extracting rule information based on machine learning, comprising the following steps:

[0008] S1: Input a regulatory document library containing the target regulatory documents. right The text content c of each regulatory document E is preprocessed to remove noise from the text and each regulatory document is split into independent rule statements sent.

[0009] S2: Fine-tune the BERT model on a general corpus of legal language, then connect the fine-tuned BERT model with a classifier to form a rule-based sentence classification model; train the rule-based sentence classification model to identify the sentence type of each rule sentence, and obtain the classification result t for each rule sentence sentence. r The classification results labels for the statement types include mandatory, prohibited, permitted, and definitional.

[0010] S3: Through constituent syntactic analysis and dependency syntactic analysis, semantic constituent extraction is performed on the rule statement `sent` after the statement type has been identified, and operation A is extracted from it. r Executor P r Condition C r Constraint O r And exception E r There are a total of five semantic components;

[0011] S4: Classify the statement type obtained from the same rule statement `sent` into `t`. r Combined with the semantic components, a rule record r = (t) is generated. r A r P r Cr O r E r U r f r ), where U r ∈A r ×(C r ∪O r () indicates the sequential relationship between an operation and a condition or an operation and a constraint. It indicates the execution relationship between executors and operations, that is, it specifies which executors must perform which operations; the rule base R consists of all rule records and is used as the basis for compliance checks in business processes.

[0012] As a preferred embodiment of the first aspect above, S2 specifically includes the following sub-steps:

[0013] S21: Collect general legal corpus and preprocess it to remove special characters and punctuation marks, forming a legal corpus;

[0014] S22: Construct a pre-training task for the masked language model. First, for each sentence S in the general corpus of the legal corpus, randomly mask some words and add [CLS] and [SEP] tags to the beginning and end of the sentence, respectively. Then, based on the BERT model, each sentence S in the legal corpus is input into the embedding layer. Through three vector layers—a character vector layer, a sentence vector layer, and a position vector layer—an input sequence of length n is obtained by adding these three vectors to obtain a composite vector representation Emb of size (n, 768), represented by the following formula:

[0015] Emb = E token +E segment +F position

[0016] Where E token Represents word vectors, E segment F represents the sentence vector. position Represents a position vector;

[0017] The synthesized vector representation Emb is then input into an encoder structure consisting of 12 Transformer layers. The attention of Emb is calculated using the multi-head self-attention mechanism of Transformer, and the final representation N is obtained through the encoder.

[0018] Finally, the final encoded representation N and the lexical positions of the occluded word are input into the multilayer perceptron to predict the occluded word.

[0019] In the task of pre-training a masked language model, a legal corpus is used to train the model to achieve fine-tuning of the parameters of the BERT model.

[0020] S23: Load the BERT model with fine-tuned parameters and connect it to the Softmax classifier to form a rule statement classification model; further fine-tune the rule statement classification model using a pre-labeled rule statement dataset, and continuously adjust the weights of the original model through backpropagation training to obtain a fine-tuned rule statement classification model; the rule statement classification results labeled in the rule statement dataset include four categories: mandatory, prohibited, permitted, and definitional.

[0021] S24: Using the fine-tuned rule statement classification model, classify the rule statement `sent` obtained in S1 to obtain the classification result `t` for each rule statement `sent`. r .

[0022] As a preferred option in the first aspect mentioned above, the statement type in S2 is specifically defined as follows:

[0023] The term "mandatory" refers to a situation where something is legally required to be enforced, and failure to meet this condition constitutes a violation.

[0024] The prohibited type refers to a situation where doing something is legally forbidden; if this condition is not met, a violation occurs.

[0025] The permitted type means that something is allowed to be done without violating the mandatory or prohibited type; if this condition is not met, a violation may occur.

[0026] The term "definition type" refers to legal provisions that define concepts.

[0027] As a preferred embodiment of the first aspect above, S3 specifically includes the following sub-steps:

[0028] S31: Perform text preprocessing on each rule statement sent, including word segmentation, part-of-speech tagging, and named entity recognition;

[0029] S32: Perform constituent parsing and dependency parsing on the preprocessed rule statement text using natural language processing tools to obtain constituent parsing trees T. c and T d Using query language in tree T c and T d Based on the extraction rules and keywords, five semantic components were extracted from the text, namely, Operation A. r Executor P r Condition C r Constraint O r And exception E r .

[0030] As a preferred embodiment of the first aspect mentioned above, the part-of-speech tagging tool is Stanford NLP, and the natural language analysis tool is Stanford NLP.

[0031] As a preferred embodiment of the first aspect above, the five semantic components in S3 are specifically defined as follows:

[0032] The operation refers to the specific content of the rule, that is, the mandatory, prohibited and permitted content defined in the rule;

[0033] The executor refers to the role responsible for the operation and execution;

[0034] The condition indicates the state when the rule applies;

[0035] The constraint represents a special case specified by the rule;

[0036] The exception indicates a situation where the rule does not apply.

[0037] As a preferred embodiment of the first aspect above, in step S32, the specific method for extracting the five types of semantic components from the text using a query language based on extraction rules and extraction keywords is as follows:

[0038] S321: Stanford Tregex is used as the query language, and the specific rule notation used is as follows:

[0039] A << B means that in the tree structure, A dominates B;

[0040] A < B means that in the tree structure, A directly governs B;

[0041] A$B indicates that A and B are sister species in the tree structure;

[0042] S322: Extract the executor P according to the following steps. r :

[0043] First, design the extraction executor P. r The required extraction rules and keywords, and the specific extraction rules for the executor include the following three types:

[0044] The first type: Subject dependency exists and

[0045] The second type: There is object dependency and passive voice, and PP < IN$( )

[0046] The third type: Exists object dependency and active voice, and...

[0047] The executor keyword is a keyword in the rule statement used to indicate the potential location of the executor, and its dependency relationship is represented in the dependency parse tree T. d superior, Representative noun phrases, Represents prepositional phrases, Represents prepositions;

[0048] Extract noun phrases from the rule statement `sent` that satisfy any of the executor extraction rules. And use it as the executor P r ;

[0049] S323: Extract condition C according to the following steps. r :

[0050] First, design the extraction condition C. r The required extraction rules and keywords, specifically the following two types of conditional extraction rules:

[0051]

[0052]

[0053] The condition keyword is used in the rule statement to indicate condition C. r Keywords for potential locations Represents a clause. Represents prepositional phrases;

[0054] Extract clauses from the rule statement `sent` that satisfy any one of the conditions for the extraction rule. or prepositional phrase And take it as condition C r ;

[0055] S324: Extract constraint O according to the following steps r :

[0056] First, design and extract constraints O r The required extraction rules and keywords, and the specific extraction rules include the following two types:

[0057]

[0058]

[0059] The constraint keywords are those used in the rule statement to indicate constraints O. r Keywords for potential locations Representative noun phrases, IN represents a prepositional phrase; IN represents a preposition.

[0060] Extract noun phrases from the rule statement `sent` that satisfy any one of the constraint extraction rules. or prepositional phrase and use it as constraint O r ;

[0061] S325: Extract exception E according to the following steps r :

[0062] First, design the extraction of exception E. r The required extraction rules and keywords, and the exception extraction rules specifically include the following three types:

[0063]

[0064]

[0065]

[0066] The exception keyword is used in the rule statement to indicate exception E. r Keywords for potential locations Represents a clause. Representative noun phrases, IN represents a prepositional phrase; IN represents a preposition.

[0067] Extract clauses from the rule statement `sent` that satisfy any one of the constraint extraction rules. Noun phrases or prepositional phrase And treat it as an exception E r ;

[0068] S326: Extract operation A according to the following steps. r :

[0069] First, design the extraction rules and keywords for modal verbs. The extraction rules for modal verbs specifically include the following three types:

[0070]

[0071]

[0072]

[0073] Where S represents a simple declarative sentence, VP represents a phrasal verb, and SINV represents a declarative sentence with subject inversion. SBAR represents a modal verb; SBAR represents a clause.

[0074] Redesign Operation A r The extraction rules are as follows: the extraction rules are for all verbs except modal verbs. Condition Cr and constraint O r Verb phrases other than VP;

[0075] Finally, extract the modal verbs in the rule statement `sent` that satisfy any one of the modal verb extraction rules. Then, extract the verb phrases VP in the rule statement `sent` that satisfy any one of the operation extraction rules and use them as operation A. r .

[0076] Secondly, the present invention provides a rule information extraction system based on machine learning, comprising:

[0077] The rule statement extraction module is used as input to a regulatory document library containing the target regulatory documents. right The text content c of each regulatory document E is preprocessed to remove noise from the text and each regulatory document is split into independent rule statements sent.

[0078] The classification model training module is used to fine-tune the BERT model on a general corpus of legal language. The fine-tuned BERT model is then connected to a classifier to form a rule-based sentence classification model. This model is trained to identify the sentence type of each rule statement (sent) and obtain the classification result (t) for each statement (sent). r The classification results labels for the statement types include mandatory, prohibited, permitted, and definitional.

[0079] The semantic component extraction module is used to extract semantic components from rule statements (sent) after the statement type has been identified through component syntactic analysis and dependency syntactic analysis, thereby extracting operation A. r Executor P r Condition C r Constraint O r And exception E r There are a total of five semantic components;

[0080] The rule base construction module is used to classify the statement type obtained from the same rule statement `sent` into categories `t`. r Combined with the semantic components, a rule record r = (t) is generated. r A r P r C r O r E r U r f r ), where U r ∈A r ×(Cr ∪O r () indicates the sequential relationship between an operation and a condition or an operation and a constraint. It indicates the execution relationship between executors and operations, that is, it specifies which executors must perform which operations; the rule base R consists of all rule records and is used as the basis for compliance checks in business processes.

[0081] Thirdly, the present invention provides a smart terminal, which includes a memory and a processor;

[0082] The memory is used to store computer programs;

[0083] The processor is configured to, when executing the computer program, implement the machine learning-based rule information extraction method as described in any of the first aspects.

[0084] As a preferred option in the third aspect mentioned above, the smart terminal is either a local terminal or a cloud terminal.

[0085] Compared to traditional text parsing methods, this invention offers the following advantages: 1. It automatically extracts semantic information from process models and regulatory documents, significantly reducing manual labor costs; 2. It uses deep learning and rule-based traditional natural language processing methods to extract semantics from regulatory texts, improving accuracy and providing more detailed violation information. Attached Figure Description

[0086] Figure 1 This is a flowchart of the present invention;

[0087] Figure 2 This is an illustrative diagram illustrating an embodiment of the present invention. Detailed Implementation

[0088] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0089] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0090] like Figure 1 As shown, in a preferred embodiment of the present invention, a method for extracting rule information based on machine learning is provided, which includes the following steps:

[0091] S1: Input a regulatory document library containing the target regulatory documents. right The text content c of each regulatory document E is preprocessed to remove noise from the text and each regulatory document is split into independent rule statements sent.

[0092] It should be noted that the regulatory document library in this invention needs to include a number of regulatory documents depending on the actual business process, and it must contain at least one regulatory document.

[0093] S2: Fine-tune the BERT model on a general corpus of legal language, then connect the fine-tuned BERT model with a classifier to form a rule-based sentence classification model; train the rule-based sentence classification model to identify the sentence type of each rule sentence, and obtain the classification result t for each rule sentence sentence. r The classification results labels for the statement types include mandatory, prohibited, permitted, and definition types.

[0094] In embodiments of the present invention, step S2 specifically includes the following sub-steps:

[0095] S21: Collect general legal corpus and preprocess it to remove special characters and punctuation marks, forming a legal corpus;

[0096] S22: Construct a pre-training task for the masked language model. First, for each sentence S in the general corpus of the legal corpus, randomly mask a portion of its words (up to 15% of the total words), and add [CLS] and [SEP] tags to the beginning and end of the sentence, respectively. Then, based on the BERT model, each sentence S in the legal corpus is input into the embedding layer. Through three vector layers—a character vector layer, a sentence vector layer, and a position vector layer—an input sequence of length n is obtained by adding these three vectors to obtain a composite vector representation Emb of size (n, 768), represented by the following formula:

[0097] Emb = E token +E segmennt +E position

[0098] Where E token Represents word vectors, E segment E represents the sentence vector. position Represents a position vector;

[0099] The synthesized vector representation Emb is then input into an encoder structure consisting of 12 Transformer layers. The multi-head self-attention mechanism of the Transformer is used to calculate the attention of Emb, and the final representation N is obtained through the encoder. The specific attention function is as follows:

[0100]

[0101] Where d represents the dimensions of vectors Q, K, and V;

[0102] Finally, the final encoded representation N and the lexical positions of the occluded word are input into the multilayer perceptron to predict the occluded word.

[0103] In the task of pre-training a masked language model, a legal corpus is used to train the model to achieve fine-tuning of the parameters of the BERT model.

[0104] S23: Load the BERT model with fine-tuned parameters and connect it to the Softmax classifier to form a rule statement classification model; further fine-tune the rule statement classification model using a pre-labeled rule statement dataset, and continuously adjust the weights of the original model through backpropagation training to obtain a fine-tuned rule statement classification model; the rule statement classification results labeled in the rule statement dataset include four categories: mandatory, prohibited, permitted, and definitional.

[0105] S24: Using the fine-tuned rule statement classification model, classify the rule statement `sent` obtained in S1. The model outputs a probability prediction vector. Search for the maximum value in the vector and take the classification result corresponding to the maximum value as the final output, that is, obtain the classification result `t` for each rule statement `sent`. r .

[0106] In embodiments of the present invention, the statement types in S2 above are specifically defined as follows:

[0107] The term "mandatory" refers to a situation where something is legally required to be enforced, and failure to meet this condition constitutes a violation.

[0108] The prohibited type refers to a situation where doing something is legally forbidden; if this condition is not met, a violation occurs.

[0109] The permitted type means that something is allowed to be done without violating the mandatory or prohibited type; if this condition is not met, a violation may occur.

[0110] The term "definition type" refers to legal provisions that define concepts.

[0111] S3: Through constituent syntactic analysis and dependency syntactic analysis, semantic constituent extraction is performed on the rule statement `sent` after the statement type has been identified, and operation A is extracted from it. r Executor P r Condition C r Constraint O r And exception E r There are a total of five semantic components.

[0112] In embodiments of the present invention, the above five semantic components are specifically defined as follows:

[0113] The operation refers to the specific content of the rule, that is, the mandatory, prohibited and permitted content defined in the rule;

[0114] The executor refers to the role responsible for the operation and execution;

[0115] The condition indicates the state when the rule applies;

[0116] The constraint represents a special case specified by the rule;

[0117] The exception indicates a situation where the rule does not apply.

[0118] In embodiments of the present invention, step S3 specifically includes the following sub-steps:

[0119] S31: Perform text preprocessing on each rule statement sent, including word segmentation, part-of-speech tagging, and named entity recognition;

[0120] S32: Perform constituent parsing and dependency parsing on the preprocessed rule statement text using natural language processing tools to obtain constituent parsing trees T. c and T d Using query language in tree T c and T d Based on the extraction rules and keywords, five semantic components were extracted from the text, namely, Operation A. r Executor P r Condition C r Constraint O r And exception E r .

[0121] Specifically, in S31 above, the part-of-speech tagging tool selected is Stanford NLP. Additionally, in S32 above, the natural language analysis tool can also be Stanford NLP.

[0122] In an embodiment of the present invention, the specific steps in S32 above, which involve using a query language to extract five types of semantic components from the text based on extraction rules and keywords, are as follows:

[0123] S321: Stanford Tregex is used as the query language, and the specific rule notation used is as follows:

[0124] A << B means that in the tree structure, A dominates B;

[0125] A < B means that in the tree structure, A directly governs B;

[0126] A$B indicates that A and B are sister species in the tree structure;

[0127] S322: Extract the executor P according to the following steps. r :

[0128] First, design the extraction executor P. r The required extraction rules and keywords, and the specific extraction rules for the executor include the following three types:

[0129] The first type: Subject dependency exists and

[0130] The second type: There is object dependency and passive voice, and PP < IN$( )

[0131] The third type: Exists object dependency and active voice, and...

[0132] The executor keywords are the keywords in the rule statements used to indicate the potential location of the executor (in this embodiment, executor keywords include physician, expert, company, judge, prosecutor, driver, officer, inspector), and their dependencies are represented in the dependency parse tree T. d superior, Representative noun phrases, Represents prepositional phrases, Represents prepositions;

[0133] Extract noun phrases from the rule statement `sent` that satisfy any of the executor extraction rules. (i.e., the bolded part) and use it as the executor P r ;

[0134] S323: Extract condition C according to the following steps. r :

[0135] First, design the extraction condition C. r The required extraction rules and keywords, specifically the following two types of conditional extraction rules:

[0136]

[0137]

[0138] The condition keyword is used in the rule statement to indicate condition C. rKeywords for potential locations (in this embodiment, conditional keywords include if, in case of, provided that, in the context of, who, whose, which) Represents a clause. Represents prepositional phrases;

[0139] Extract clauses from the rule statement `sent` that satisfy any one of the conditions for the extraction rule. or prepositional phrase (i.e., the bolded part) and use it as condition C. r ;

[0140] S324: Extract constraint O according to the following steps r :

[0141] First, design and extract constraints O r The required extraction rules and keywords, and the specific extraction rules include the following two types:

[0142]

[0143]

[0144] The constraint keywords are those used in the rule statement to indicate constraints O. r Keywords for potential locations (in this embodiment, constraint keywords include before, after, at least, at most, equal to, greatest, smallest, last of, least of), Representative noun phrases, IN represents a prepositional phrase; IN represents a preposition.

[0145] Extract noun phrases from the rule statement `sent` that satisfy any one of the constraint extraction rules. or prepositional phrase (i.e., the bolded black portion) and use it as constraint O r ;

[0146] S325: Extract exception E according to the following steps r :

[0147] First, design the extraction of exception E. r The required extraction rules and keywords, and the exception extraction rules specifically include the following three types:

[0148]

[0149]

[0150]

[0151] The exception keyword is used in the rule statement to indicate exception E. r Keywords for potential locations (exceptions in this embodiment include with the exception of, except for, derogation, apart from, and other than). Represents a clause. Representative noun phrases, IN represents a prepositional phrase; IN represents a preposition.

[0152] Extract clauses from the rule statement `sent` that satisfy any one of the constraint extraction rules. Noun phrases or prepositional phrase (i.e., the bolded part) and treat it as an exception E. r ;

[0153] S326: Extract operation A according to the following steps. r :

[0154] First, design the extraction rules and keywords for modal verbs. The extraction rules for modal verbs specifically include the following three types:

[0155]

[0156]

[0157]

[0158] Where S represents a simple declarative sentence, VP represents a phrasal verb, and SINV represents a declarative sentence with subject inversion. SBAR represents a modal verb; SBAR represents a clause.

[0159] Redesign Operation A r The extraction rules are as follows: the extraction rules are for all verbs except modal verbs. Condition C r and constraint O r Verb phrases other than VP;

[0160] Finally, extract the modal verbs in the rule statement `sent` that satisfy any one of the modal verb extraction rules. Then, extract the verb phrases VP in the rule statement `sent` that satisfy any one of the operation extraction rules and use them as operation A. r .

[0161] The Stanford Tregex query language syntax rules used above are summarized in the table below:

[0162] Table 1 Tregex Syntax Rules

[0163]

[0164] The extraction rules and keywords are detailed in the table below:

[0165] Table 2 Extraction Rules

[0166]

[0167] Note: MD: modality, NP: noun phrase, VP: verb phrase, IN: preposition, PP: prepositional phrase, SBAR: clause, SINV: declarative sentence with subject inversion.

[0168] Table 3 Keyword Extraction

[0169]

[0170] The bolded black portion in the rule table represents the extraction target.

[0171] S4: Classify the statement type obtained from the same rule statement `sent` into `t`. r Combined with the semantic components, a rule record r = (t) is generated. r A r P r C r O r E r U r f r ), where U r ∈A r ×(C r ∪O r () indicates the sequential relationship between an operation and a condition or an operation and a constraint. It indicates the execution relationship between executors and operations, that is, it specifies which executors must perform which operations; the rule base R consists of all rule records and is used as the basis for compliance checks in business processes.

[0172] When performing compliance checks, business processes can be compared with the corresponding rule records in the rule base to determine whether they conform to the rule records. If they do not conform, they can be considered as violations. Therefore, the rules in the rule base generated by this invention can be used to perform consistency checks with the enterprise's business processes and discover non-compliant business processes within the enterprise.

[0173] Furthermore, the above-described steps can be implemented as software functional units in the form of logical instructions in memory. When this software is sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0174] Therefore, in another embodiment of the present invention, based on the same inventive concept, a rule information extraction system based on machine learning is also provided, comprising:

[0175] The rule statement extraction module is used as input to a regulatory document library containing the target regulatory documents. right The text content c of each regulatory document E is preprocessed to remove noise from the text and each regulatory document is split into independent rule statements sent.

[0176] The classification model training module is used to fine-tune the BERT model on a general corpus of legal language. The fine-tuned BERT model is then connected to a classifier to form a rule-based sentence classification model. This model is trained to identify the sentence type of each rule statement (sent) and obtain the classification result (t) for each statement (sent). r The classification results labels for the statement types include mandatory, prohibited, permitted, and definitional.

[0177] The semantic component extraction module is used to extract semantic components from rule statements (sent) after the statement type has been identified through component syntactic analysis and dependency syntactic analysis, thereby extracting operation A. r Executor P r Condition C r Constraint O r And exception E r There are a total of five semantic components;

[0178] The rule base construction module is used to classify the statement type obtained from the same rule statement `sent` into categories `t`. r Combined with the semantic components, a rule record r = (t) is generated. r A r P r C r O r E r Ur f r ), where Ux∈A r ×(C r ∪O r () indicates the sequential relationship between an operation and a condition or an operation and a constraint. It indicates the execution relationship between executors and operations, that is, it specifies which executors must perform which operations; the rule base R consists of all rule records and is used as the basis for compliance checks in business processes.

[0179] The specific methods executed by each module in the above-mentioned machine learning-based rule information extraction system are the same as those in the machine learning-based rule information extraction methods described in the previous embodiments, and will not be described in detail here.

[0180] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the machine learning-based rule information extraction method provided in the above embodiments. The storage medium stores a computer program, which, when executed by a processor, implements the machine learning-based rule information extraction method as described in the foregoing scheme.

[0181] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a smart terminal corresponding to the machine learning-based rule information extraction method provided in the above embodiments, which includes a memory and a processor;

[0182] The memory is used to store computer programs;

[0183] The processor is configured to implement the machine learning-based rule information extraction method as described in the foregoing embodiments when executing the computer program.

[0184] It is understood that the aforementioned storage medium and memory can be random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage medium can also be any medium capable of storing program code, such as a USB flash drive, external hard drive, magnetic disk, or optical disk.

[0185] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0186] It is understandable that the aforementioned smart terminals can be local terminals or cloud terminals.

[0187] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the device and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0188] The following example demonstrates the implementation and technical effects of the machine learning-based rule information extraction method described in the preceding embodiments.

[0189] Example

[0190] The steps in this embodiment are the same as steps S1 to S4 described above in the specific implementation method, and will not be repeated here. The following section demonstrates some of the implementation processes and results:

[0191] To clearly demonstrate the results of the entire method of this invention, this embodiment uses an example of a compliance check on a telephone company's process model for acquiring new customers. This telephone company needs to comply with regulatory documents from various sources, including company regulations and the General Data Protection Regulation (GDPR). This embodiment first uses the machine learning-based rule information extraction method proposed in this invention to parse the aforementioned regulatory documents, and then further evaluates the compliance of the process model based on the parsed rules. The entire process model includes the following steps: First, the telephone company requests new data from the customer and verifies the data's accuracy. If the customer decides to change their phone number, the telephone company will communicate with the original telephone company and sign a contract with the customer. After completing the above steps and receiving payment, the telephone company sends a SIM card to the customer and activates it. If the process takes more than 30 days, it will be terminated.

[0192] This embodiment simply provides examples of rules extracted from regulatory documents. Specific rule constraints and their sources are shown in Table 4.

[0193] Table 4 Rules from Different Sources

[0194]

[0195] The inspection results of the process model are as follows: Figure 2 As shown in Table 4, this embodiment uses the aforementioned rule information extraction method to check whether the model conforms to the rules. It can be seen that this embodiment detected operation missing violations for both rules R1 and R4. Based on the results of this embodiment, analysts can modify the process model after analysis. For violations of R1, a sub-process needs to be added to the process; that is, the process will terminate when the process time exceeds 30 days. For violations of R4, an activity to "request user consent" is added to the process model. For rule R2, this embodiment detected an execution order error violation. To resolve this issue, an activity to "check the correctness of personal data" should be added after obtaining personal data. Finally, this embodiment detected an incorrect executor violation for R3. According to this rule, the SIM card should be activated by the telephone company, not the customer themselves. It is worth noting that R4 is a rule from GDPR, which differs from previous company regulations. When checking the process model, the participants in the rule records must be manually converted because participants in GDPR are not specific but rather a category. In this case, "data controller" needs to be converted to "telephone company" and "data subject" to "user" before the check.

[0196] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for extracting rule information based on machine learning, characterized in that... Includes the following steps: S1: Input a regulatory document library containing the target regulatory documents. ,right Each regulatory document text content Preprocessing is performed to remove noise from the text and to break each regulatory document down into individual rule statements. ; S2: Will The model is fine-tuned on a general corpus of legal terminology, and then the fine-tuned model is... The model is connected to the classifier to form a rule statement classification model; the rule statement classification model is trained to recognize each rule statement. The statement type is used to obtain each rule statement. Classification results The classification results labels for the statement types include mandatory, prohibited, permitted, and definitional. S3: Through constituent syntactic analysis and dependency syntactic analysis, the rule statements whose sentence types have been identified are processed. Semantic component extraction is performed to extract operations from them. Executor ,condition ,constraint and exceptions There are a total of five semantic components; S4: Combine the same rule statement The classification results of the statement types obtained in the process Combined with the semantic components, a rule record is generated. ,in This indicates the sequential relationship between operations and conditions, or operations and constraints. It indicates the execution relationship between executors and operations, that is, it specifies which executors must perform which operations; The rule base consists of all rule records. It is used as the basis for compliance checks in business processes.

2. The rule information extraction method based on machine learning according to claim 1, characterized in that, S2 specifically includes the following sub-steps: S21: Collect general legal corpus and preprocess it to remove special characters and punctuation marks, forming a legal corpus; S22: Construct a pre-training task for the masked language model, that is, firstly, train each sentence in the general corpus within the legal corpus. Randomly obscure some words and add [CLS] and [SEP] markers at the beginning and end of the sentence respectively; then based on The model will analyze each sentence in the legal corpus. The input embedding layer consists of three vector layers: a character vector layer, a sentence vector layer, and a position vector layer. An input sequence of length n is obtained by adding these three vectors to produce a composite vector representation of size (n, 768). The formula is as follows: in Representing word vectors, Represents sentence vectors, Represents a position vector; Then represent the composite vector The input is fed into an encoder structure consisting of 12 Transformer layers, and the multi-head self-attention mechanism of the Transformer is used for computation. The attention of the encoder is used to obtain the final representation. ; Finally, the final encoded representation The lexical positions of the occluded word are input into a multilayer perceptron to predict the occluded word; In the pre-training task of masked language models, a legal corpus is used to train the model. Fine-tuning of model parameters; S23: Loading parameters after fine-tuning The model is connected to a Softmax classifier to form a rule statement classification model. The rule statement classification model is further fine-tuned using a pre-labeled rule statement dataset. The weights of the original model are continuously adjusted through backpropagation training to obtain a fine-tuned rule statement classification model. The rule statement classification results labeled in the rule statement dataset include four categories: mandatory, prohibited, permitted, and definitional. S24: Using the fine-tuned rule statement classification model, input the rule statement obtained in S1. The rules are categorized to obtain each rule statement. Classification results .

3. The rule information extraction method based on machine learning according to claim 1, characterized in that, The statement types in S2 are specifically defined as follows: The term "mandatory" refers to a situation where something is legally required to be enforced, and failure to meet this condition constitutes a violation. The prohibited type refers to a situation where doing something is legally forbidden; if this condition is not met, a violation occurs. The permitted type means that something is allowed to be done without violating the mandatory or prohibited type; if this condition is not met, a violation may occur. The term "definition type" refers to legal provisions that define concepts.

4. The rule information extraction method based on machine learning according to claim 1, characterized in that, S3 specifically includes the following sub-steps: S31: For each rule statement Perform text preprocessing, including word segmentation, part-of-speech tagging, and named entity recognition; S32: Perform constituent parsing and dependency parsing on the preprocessed rule statement text using natural language processing tools to obtain constituent parsing trees. and Using query language in a tree and Based on the extraction rules and keywords, five semantic components were extracted from the text, namely operations. Executor ,condition ,constraint and exceptions .

5. The rule information extraction method based on machine learning according to claim 4, characterized in that, The part-of-speech tagging tool used is Stanford NLP.

6. The rule information extraction method based on machine learning according to claim 4, characterized in that, The five semantic components in S3 are specifically defined as follows: The operation refers to the specific content of the rule, that is, the mandatory, prohibited and permitted content defined in the rule; The executor refers to the role responsible for the operation and execution; The condition indicates the state when the rule applies; The constraint represents a special case specified by the rule; The exception indicates a situation where the rule does not apply.

7. The rule information extraction method based on machine learning according to claim 4, characterized in that, In step S32, the specific steps for extracting the five semantic components from the text using a query language based on extraction rules and keywords are as follows: S321: Stanford Tregex is used as the query language, and the specific rule notation used is as follows: A << B means that in the tree structure, A dominates B; A < B means that in the tree structure, A directly governs B; A $ B means that A and B are sisters in the tree structure; S322: Extract the executor according to the following steps : First, design the extraction executor The required extraction rules and keywords, and the specific extraction rules for the executor include the following three types: The first type: Subject dependency exists and NP < executor keyword The second type: There is object dependency and passive voice, and PP < IN $ (NP < actor keyword) The third type: Exists object dependency and active voice, and NP < actor keyword. The executor keyword is a keyword in the rule statement used to indicate the potential location of the executor, and its dependency relationship is represented in the dependency parse tree. Above, NP represents a noun phrase, PP represents a prepositional phrase, and IN represents a preposition; Extract the rule statement The noun phrase NP that satisfies any executor extraction rule is used as the executor. ; S323: Extract conditions according to the following steps : First, design the extraction conditions. Required extraction rules and keywords, specific extraction rules. Including the following two types: SBAR << Conditional Keywords PP << Conditional Keywords The condition keywords are those used in the rule statement to indicate conditions. Keywords for potential locations: SBAR stands for clause, PP stands for prepositional phrase; Extract the rule statement Clauses that satisfy any one of the condition extraction rules (SBAR or prepositional phrase PP) are used as conditions. ; S324: Extract constraints according to the following steps : First, design the extraction constraints. The required extraction rules and keywords, and the specific extraction rules include the following two types: NP < constraint keywords PP < (IN < constraint keyword) $NP The constraint keywords are those used in the rule statement to indicate constraints. Keywords for potential positions: NP represents noun phrases, PP represents prepositional phrases, and IN represents prepositions; Extract the rule statement The noun phrase NP or prepositional phrase PP that satisfies any constraint extraction rule is used as a constraint. ; S325: Extract exceptions following these steps : First design the extraction exception The required extraction rules and keywords, and the exception extraction rules specifically include the following three types: SBAR << Exception Keywords PP << Exception Keywords NP << (IN < exception keyword) The exception keyword is used in the rule statement to indicate exceptions. Keywords for potential positions: SBAR represents clause, NP represents noun phrase, PP represents prepositional phrase, and IN represents preposition; Extract the rule statement Clauses that satisfy any one of the constraint extraction rules (SBAR, NP, or PP) are considered exceptions. ; S326: Extract the operation according to the following steps. : First, design the extraction rules and keywords for modal verbs. The extraction rules for modal verbs specifically include the following three types: (S < VP) << MD ((SINV << S) < VP) << MD ((SBAR << S) < VP) << MD Where S represents a simple declarative sentence, VP represents a phrasal verb, SINV represents a declarative sentence with subject inversion, MD represents a modal verb, and SBAR represents a clause; Redesign Operation The extraction rules are as follows: the extraction rules are all MD except for modal verbs and conditional verbs. and constraints Verb phrases other than VP; Finally, extract the rule statements. The modal verb MD that satisfies any modal verb extraction rule is then used to extract the rule-based statements. Verb phrases (VPs) that satisfy any operation extraction rule are used as operations. .

8. A rule information extraction system based on machine learning, characterized in that, include: The rule statement extraction module is used as input to a regulatory document library containing the target regulatory documents. ,right Each regulatory document text content Preprocessing is performed to remove noise from the text and to break each regulatory document down into individual rule statements. ; The classification model training module is used to train the classification model. The model is fine-tuned on a general corpus of legal terminology, and then the fine-tuned model is... The model is connected to the classifier to form a rule statement classification model; The rule statement classification model is trained to recognize each rule statement. The statement type is used to obtain each rule statement. Classification results The classification results labels for the statement types include mandatory, prohibited, permitted, and definitional. The semantic component extraction module is used to extract semantic components from rule statements after identifying their types through component syntactic analysis and dependency syntactic analysis. Semantic component extraction is performed to extract operations from them. Executor ,condition ,constraint and exceptions There are a total of five semantic components; The rule base building module is used to combine the same rule statement The classification results of the statement types obtained in the process Combined with the semantic components, a rule record is generated. ,in This indicates the sequential relationship between operations and conditions, or operations and constraints. It indicates the execution relationship between executors and operations, that is, it specifies which executors must perform which operations; The rule base consists of all rule records. It is used as the basis for compliance checks in business processes.

9. A smart terminal, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement the rule information extraction method based on machine learning as described in any one of claims 1 to 7 when executing the computer program.

10. The smart terminal as described in claim 9, characterized in that, The smart terminal can be a local terminal or a cloud terminal.

Images