Parameter information extraction method and device and electronic equipment

Abstract

The embodiment of the invention provides a parameter information extraction method and device and electronic equipment, and relates to the technical field of data extraction, and the method comprises the steps: obtaining a target contract to be subjected to parameter information extraction; determining target positioning information corresponding to a target contract category to which the target contract belongs; based on the target positioning information, positioning a paragraph containing parameter information of the target parameter in the target contract to obtain a target paragraph corresponding to the target parameter; obtaining a plurality of parameter feature sentences corresponding to the target parameter; analyzing whether a target single sentence matched with any parameter feature sentence corresponding to the target parameter exists in the single sentences of the target paragraph corresponding to the target parameter; and in response to the existence of the matched target single sentence, extracting parameter information of a target parameter in the target contract based on the target single sentence. Therefore, the extraction precision of the parameter information in the contract can be improved.

Description

Technical Field

[0001] This application relates to the technical field of data extraction, and in particular, to a method, apparatus, and electronic device for extracting parameter information. Background Art

[0002] Currently, after a contract is signed and comes into effect, business personnel need to register the parameter information of various parameters in the contract for subsequent other data operations based on the extracted parameter information of various parameters. Taking a fund contract as an example, the parameter information of various parameters extracted from the fund contract can be applied to subsequent data operations such as fund settlement and safety valuation.

[0003] In related technologies, parameter information is extracted based on keywords. For example, for a parameter such as fund name, the keyword corresponding to this parameter is "the name of the purchased fund is". Then, when extracting parameter information with this keyword, the target single sentence determined to contain this keyword in the contract is "the name of the purchased fund is: XX Fund". At this time, the parameter information of this parameter extracted is "XX Fund". However, for contracts of the same category, there are diversities in the description statements used by different companies for the same type of parameter. For example, for a parameter such as fund name, the description statement used by Company A for this parameter is "the name of the purchased fund is XX Fund", the description statement used by Company B for this parameter is "the name of the fund purchased this time is XX Fund", and the description statement used by Company C for this parameter is "the fund purchased this time is named XX Fund", etc.

[0004] Therefore, if parameter information is extracted based on keywords in related technologies, it may lead to missing extraction of parameter information in the contract or even extracting incorrect parameter information, ultimately resulting in a low extraction accuracy for the parameter information in the contract. Summary of the Invention

[0005] The purpose of the embodiments of this application is to provide a method, apparatus, and electronic device for extracting parameter information to improve the extraction accuracy of parameter information in the contract. The specific technical solutions are as follows:

[0006] In a first aspect, the embodiments of this application provide a method for extracting parameter information, including:

[0007] Obtain a target contract for which parameter information extraction is to be performed; determine the target positioning information corresponding to the target contract category to which the target contract belongs; wherein, the target positioning information is used to indicate the target parameter for which parameter information extraction is to be performed in the contracts of the target contract category, and the text position of the parameter information of the target parameter in any contract; based on the target positioning information, in the target contract, locate the paragraph containing the parameter information of the target parameter to obtain the target paragraph corresponding to the target parameter; obtain a plurality of parameter feature sentences corresponding to the target parameter; wherein, each parameter feature sentence corresponding to the target parameter is: a candidate sentence for describing the parameter information of the target parameter in the contracts of the target contract category; among the plurality of parameter feature sentences corresponding to the target parameter, different parameter feature sentences are different; analyze whether there is a target single sentence in each single sentence of the target paragraph corresponding to the target parameter that matches any parameter feature sentence corresponding to the target parameter; in response to the existence of a matching target single sentence, based on the target single sentence, extract the parameter information of the target parameter in the target contract.

[0008] In a second aspect, an embodiment of the present application provides a parameter information extraction device, including:

[0009] A contract acquisition module, configured to acquire a target contract for which parameter information extraction is to be performed;

[0010] A positioning information determination module, configured to determine the target positioning information corresponding to the target contract category to which the target contract belongs; wherein, the target positioning information is used to indicate the target parameter for which parameter information extraction is to be performed in the contracts of the target contract category, and the text position of the parameter information of the target parameter in any contract;

[0011] A paragraph determination module, configured to, based on the target positioning information, locate, in the target contract, the paragraph containing the parameter information of the target parameter to obtain the target paragraph corresponding to the target parameter;

[0012] A feature sentence acquisition module, configured to acquire a plurality of parameter feature sentences corresponding to the target parameter; wherein, each parameter feature sentence corresponding to the target parameter is: a candidate sentence for describing the parameter information of the target parameter in the contracts of the target contract category; among the plurality of parameter feature sentences corresponding to the target parameter, different parameter feature sentences are different;

[0013] A single sentence analysis module, configured to analyze whether there is a target single sentence in each single sentence of the target paragraph corresponding to the target parameter that matches any parameter feature sentence corresponding to the target parameter;

[0014] An information extraction module, configured to, in response to the existence of a matching target single sentence, extract parameter information of the target parameter in the target contract based on the target single sentence.

[0015] In a third aspect, an embodiment of the present application provides an electronic device, including: a memory for storing a computer program; a processor for implementing any of the parameter information extraction methods when executing the program stored on the memory.

[0016] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, in which a computing program is stored, and when the computer program is executed by a processor, any of the parameter information extraction methods is implemented.

[0017] An embodiment of the present application also provides a computer program product containing instructions, which when running on a computer, causes the computer to execute any of the above-mentioned parameter information extraction methods.

[0018] Advantageous effects of the embodiments of the present application:

[0019] In the solution provided by the embodiments of the present application, corresponding positioning information is set for each contract category, and the positioning information is used to indicate the parameter for which parameter information is to be extracted in the contract of the contract category, and the text position of the parameter information of the parameter in any contract. In this way, after obtaining the target contract, the target positioning information corresponding to the target contract category to which the target contract belongs is determined, and thus, through the determined target positioning information, the target paragraph corresponding to the target parameter is determined in the target contract, and it is analyzed whether there is a target single sentence in each single sentence of the target paragraph that matches any parameter feature sentence corresponding to the target parameter; when there is a target single sentence, the parameter information of the target parameter in the target contract is extracted based on the target single sentence. Among them, there are multiple parameter feature sentences corresponding to a parameter in a contract category, and each parameter feature sentence corresponding to the parameter is: a candidate sentence for describing the parameter information of the parameter in the contract of the contract category, and among the multiple parameter feature sentences corresponding to the parameter, different parameter feature sentences are different. That is to say, there is a one-to-many correspondence between the parameter and the corresponding parameter feature sentence. When extracting parameter information, based on this correspondence, the matching of the target single sentence can be performed, which can improve the extraction accuracy of the parameter information and the extraction precision when extracting the parameter information in the contract. Description of the Drawings

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application, and those of ordinary skill in the art can also obtain other embodiments according to these drawings.

[0021] Figure 1 A flow chart of a parameter information extraction method provided in an embodiment of the present application;

[0022] Figure 2 A schematic diagram of a specific embodiment provided for the present application;

[0023] Figure 3 A schematic diagram of a process of extracting parameters using a Boolean value type provided in an embodiment of the present application;

[0024] Figure 4 A flow chart of a parameter extraction process using a string type provided in an embodiment of the present application;

[0025] Figure 5 A schematic diagram of a product interface provided in an embodiment of the present application;

[0026] Figure 6 A schematic diagram of the structure of a parameter information extraction device provided in an embodiment of the present application;

[0027] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field based on the present application belong to the scope of protection of the present application.

[0029] To facilitate understanding, the background technology involved in this application is first introduced as an example below.

[0030] After the contracts of private equity funds, bonds, stocks, trusts and other securities are signed and come into effect, business personnel (also known as operation personnel) need to register the parameter information of various parameters in the contracts so as to perform other data operations based on the parameter information of various parameters extracted subsequently. Taking the fund contract as an example, the parameter information of various parameters extracted from the fund contract can be applied to subsequent data operations related to business operation activities such as fund settlement and safe valuation. The process of registering the parameter information of the fund contract is as follows: First, the business personnel extract the key parameters from the contract, and then register them item by item in the contract registration system. However, due to factors such as human resources and professional skills, there are risks of false reporting and missing reporting in manual extraction and entry, and the efficiency is relatively low. To address the above problems, in the process of extracting the parameter information in the contract, current technical means are used to replace manual work. However, the existing general contract information extraction methods still have deficiencies. Exemplarily: (1) The extraction scope of elements (parameter information) is narrow, and only basic information such as fund name can be extracted, which cannot meet the requirements for extracting more complex key parameters. (2) The extraction accuracy of elements is low. Currently, when extracting elements in the contract, it is completed through technical means such as keyword and regular expression matching. However, due to the diversity of the content expressed in the contract, the above means often fail.

[0031] To solve the above technical problems and improve the extraction accuracy of the parameter information in the contract, the embodiments of the present application provide a parameter information extraction method, device and electronic device. Among them, this method is applicable to various application scenarios for extracting parameter information from contracts. For example, extracting parameter information from fund contracts, and for another example, extracting parameter information from stock contracts, etc. And this method can be applied to various electronic devices such as laptop computers and desktop computers, hereinafter referred to as electronic devices. Based on this, the embodiments of the present application do not limit the application scenarios and execution entities of this method.

[0032] The following will specifically describe a parameter information extraction method provided by the embodiments of the present application in conjunction with the accompanying drawings. As Figure 1 shown, this method may include the following steps:

[0033] S101: Obtain a target contract to be subjected to parameter information extraction.

[0034] In this application, when extracting parameters for a target contract, first obtain the target contract for which parameter information is to be extracted. For example, multiple contracts are stored in a contract system, and each contract is set with a registration status. The contract with the set registration status indicating that the contract has not been registered is determined as the target contract for which parameter information is to be extracted. Another example is that the contract specified by the user is determined as the target contract for which parameter information is to be extracted. This application does not limit the method for determining the target contract. Among them, the above target contract can be a fund contract, a bond contract, a stock contract, a labor contract, a sales contract, a lease contract, etc. In this regard, the embodiments of this application do not make specific limitations.

[0035] S102: Determine the target positioning information corresponding to the target contract category to which the target contract belongs. Among them, the target positioning information is used to indicate the target parameters for which parameter information is to be extracted in the contracts of the target contract category, and the text position of the parameter information of the target parameters in any contract.

[0036] In this application, the parameters for which parameter information needs to be extracted are different in contracts of different contract categories, and the parameters to be extracted are preset based on user requirements. For any contract of the same contract category, the chapter, paragraph, and other text positions where the parameter information of the parameters to be extracted is located in the contract are fixed. Therefore, for contracts of different contract categories, the corresponding positioning information for that contract category is set in advance. This positioning information is used to indicate the parameters for which parameter information is to be extracted in the contracts of that contract category, and the position of the parameter information of the parameters in any contract. Exemplarily, the text position of the reference information of the target parameter in any contract can be the paragraph position in any contract.

[0037] In this way, after obtaining the target contract, the target contract category to which the target contract belongs can be determined, and thus the target positioning information corresponding to the target contract category can be determined. Based on the determined target positioning information, the target parameters for which parameter information is to be extracted in the target contract and the text position of the parameter information of the target parameters in any contract can be determined. It should be noted that the embodiments of this application do not limit the number of target parameters to be extracted. That is, the number of target parameters can be one or at least two. When the number of target parameters is at least two, the at least two target parameters are parameters of different categories. For example, product code, product short name, product full name, custodian, etc. At this time, the following steps S103 - S106 are executed for each of the at least two target parameters.

[0038] S103: Based on the target positioning information, in the target contract, locate the paragraph containing the parameter information of the target parameter to obtain the target paragraph corresponding to the target parameter.

[0039] In this application, based on the target location information, the target parameters to be extracted in the target contract are determined. For the target parameters, based on the text location of the parameter information of the target parameter in the target contract, the paragraphs containing the parameter information of the target parameter are located in the target contract, and the target paragraphs corresponding to the target parameter are obtained.

[0040] Optionally, the above target location information further includes a start field for characterizing the start of each chapter in the target contract, and an end field for characterizing the end of the chapter. Thus, based on the start field and the end field, the text content of the target contract is divided into chapters, and multiple paragraphs corresponding to each chapter are obtained, so as to facilitate subsequent locating the paragraphs containing the parameter information of the target parameter based on the text location of the parameter information of the target parameter in each of the multiple paragraphs corresponding to the divided chapters, and obtaining the target paragraphs corresponding to the target parameter.

[0041] S104: Obtain multiple parameter feature sentences corresponding to the target parameter. Each parameter feature sentence corresponding to the target parameter is a candidate sentence for describing the parameter information of the target parameter in the contract of the target contract category; among the multiple parameter feature sentences corresponding to the target parameter, different parameter feature sentences are different.

[0042] In this application, parameter feature sentences corresponding to the target parameter in the contract of the target contract category are set in advance. Each parameter feature sentence is a candidate sentence for describing the parameter information of the target parameter in the contract of the target contract category. Exemplarily, taking the target parameter as "fund name" as an example, the parameter information of this target parameter is "XX Fund", then the candidate sentence for describing "XX Fund" can be "The name of the purchased fund is XX Fund". And considering the diversification of text description content, among the multiple parameter feature sentences corresponding to a target parameter, different parameter feature sentences are different. For example, the candidate sentences for describing "XX Fund" can be "The name of the purchased fund is XX Fund", "The name of the fund bought this time is XX Fund", "The fund purchased this time is named XX Fund", etc. That is to say, for the same target parameter, the multiple parameter feature sentences corresponding to it are different text contents for describing the same parameter information. Therefore, after determining the target parameter to be extracted, the multiple parameter feature sentences corresponding to this target parameter set in advance can be obtained.

[0043] Optionally, record the multiple parameter feature sentences corresponding to each target parameter in a parameter feature sentence set.

[0044] S105: Analyze whether there is a target single sentence in each single sentence of the target paragraph corresponding to the target parameter that matches any of the parameter feature sentences corresponding to the target parameter.

[0045] In this application, based on the target location information, the target paragraph corresponding to the target parameter can be determined, that is, it is determined that the parameter information of the target parameter exists in a single sentence in the target paragraph. Therefore, after determining multiple parameter feature sentences corresponding to the target parameter, analyze each single sentence in the target paragraph to see if there is a target single sentence that matches any of the parameter feature sentences corresponding to the target parameter. That is to say, for each single sentence in the target paragraph, if the single sentence matches any of the parameter feature sentences corresponding to the target parameter, it indicates that the single sentence contains the parameter information of the target parameter; if the single sentence does not match all of the parameter feature sentences corresponding to the target parameter, continue to analyze the next single sentence until a target single sentence is determined from each single sentence, or no target single sentence is determined from each single sentence, and the analysis ends.

[0046] Optionally, a parameter feature sentence blacklist is set for each target parameter, and the blacklist records sentences that have a similarity with the parameter feature sentence and meet the similarity threshold but are not the parameter feature sentence. In this way, when performing parameter feature sentence matching for the target single sentence, if the matched single sentence is a sentence recorded in the parameter feature sentence blacklist, directly ignore the single sentence and analyze the next single sentence to improve the determination efficiency of the target single sentence.

[0047] Optionally, the parameter feature sentence blacklist corresponding to each target parameter is stored in a parameter feature sentence blacklist set.

[0048] S106: In response to the existence of a matching target single sentence, based on the target single sentence, extract the parameter information of the target parameter in the target contract.

[0049] In this application, when there is a matching target single sentence, it indicates that the parameter information of the target parameter exists in the single sentence. At this time, based on the target single sentence, extract the parameter information of the target parameter in the target contract. It should be noted that considering the different parameter categories to which the target parameter belongs, the presentation forms of the corresponding extracted parameter information are different. Exemplarily, if the parameter category to which the target parameter belongs is a boolean value, that is, the extracted parameter information does not exist in the text content of the target single sentence, but is a judgment result obtained by judging whether the target parameter exists in the target single sentence. For example, when the judgment result is that the target parameter exists, a predetermined field value is used as the parameter information of the target parameter, and the predetermined field value is used to indicate the existence of the target parameter; if the parameter category to which the target parameter belongs is a string, that is, the extracted parameter information is the text content existing in the target single sentence. For example, when the target parameter is the fund name, the parameter information of the target parameter is XX Fund included in the target single sentence.

[0050] Optionally, in one embodiment, in response to the absence of a matching target single sentence in each single sentence of the target paragraph corresponding to the target parameter, that is, in the target paragraph, there is no single sentence that matches any of the parameter feature sentences corresponding to the target parameter, and the parameter information of the target parameter cannot be extracted. At this time, an alarm message indicating the failure of parameter information extraction of the target parameter is output to remind the user that the parameter information extraction fails. Among them, the presentation form of the above alarm message can be a predetermined empty string or sound information. In this regard, the embodiments of the present application do not make specific limitations.

[0051] Optionally, in one embodiment, the parameter information of the extracted target parameter is registered in a predetermined contract registration system. In this embodiment, a data set constructed for the target contract in the predetermined contract registration system is determined, and the parameter information of each item set in the data set corresponds to a target parameter, that is, the corresponding relationship between the item name of the target parameter set in the data set and the parameter information of the target parameter is set. In this way, based on the above corresponding relationship, the parameter information of the extracted target parameter is registered in the predetermined contract registration system to improve the accuracy of parameter information entry. It should be noted that the operation process of entering parameter information into the contract registration system is executed through a robot automation process. Therefore, when entering parameter information, first, the elements of the contract registration system interface are identified, and based on the above corresponding relationship, the parameter information is sequentially entered into the corresponding positions to improve the efficiency and accuracy of parameter information entry.

[0052] As can be seen from the above, in the solution provided by the embodiments of the present application, corresponding positioning information is set for each contract category, and the positioning information is used to indicate the parameters for which parameter information extraction is to be performed in the contracts of the contract category, and the text positions of the parameter information of the parameters in any contract. In this way, after obtaining the target contract, the target positioning information corresponding to the target contract category to which the target contract belongs is determined. Thus, through the determined target positioning information, the target paragraph corresponding to the target parameter is determined in the target contract, and it is analyzed whether there is a target single sentence that matches any of the parameter feature sentences corresponding to the target parameter in each single sentence of the target paragraph; when there is a target single sentence, the parameter information of the target parameter in the target contract is extracted based on the target single sentence. Among them, a parameter in a contract category corresponds to multiple parameter feature sentences, and each parameter feature sentence corresponding to the parameter is a candidate sentence for describing the parameter information of the parameter in the contract of the contract category, and among the multiple parameter feature sentences corresponding to the parameter, different parameter feature sentences are different. That is to say, there is a one-to-many correspondence between the parameter and the corresponding parameter feature sentence. When extracting parameter information, based on this corresponding relationship for target single sentence matching can improve the extraction accuracy of parameter information and the extraction precision when extracting parameter information in the contract.

[0053] Optionally, in one implementation, the above step S105, analyzing whether there is a target sentence matching any parameter feature sentence corresponding to the target parameter in each single sentence of the target paragraph corresponding to the target parameter, may include the following steps:

[0054] Step A1: Segment the target paragraph corresponding to the target parameter to obtain multiple single sentences of the target paragraph;

[0055] Step A2: traverse multiple single sentences, and when traversing to each current single sentence, perform the following process:

[0056] Step B1: Perform string matching on the single sentence and each parameter feature sentence in the multiple parameter feature sentences corresponding to the target parameter. Step B2: If the single sentence matches any parameter feature sentence, the single sentence is determined to be the target single sentence and the traversal ends; otherwise, execute step B3. Step B3: Perform a predetermined vectorization process on the single sentence to obtain the first vector feature of the single sentence, and calculate the similarity between the first vector feature and the second vector feature of each parameter feature sentence corresponding to the target parameter;

[0057] The second vector feature of each parameter feature sentence corresponding to the target parameter is obtained by performing a predetermined vectorization process on the parameter feature sentence. Step B4: If the similarity between the first vector feature of the single sentence and the second vector feature of any parameter feature sentence is not less than the similarity threshold, the single sentence is determined to be the target single sentence and the traversal ends; otherwise, the traversal proceeds to the next single sentence.

[0058] In this implementation, the target paragraph corresponding to the target parameter is processed by sentence segmentation to obtain multiple single sentences of the target paragraph. Then, multiple single sentences are traversed, and when traversing to each current single sentence, the above steps B1-B3 are executed, that is, for each current single sentence, the single sentence is string matched with each parameter feature sentence in the multiple parameter feature sentences corresponding to the target parameter. If the single sentence matches any parameter feature sentence, it is determined that the single sentence contains the target parameter and the parameter information of the target parameter. At this time, it can be determined that the single sentence is the target single sentence, and the traversal ends. If the single sentence does not match any parameter feature sentence, a predetermined vectorization process is performed on the single sentence to obtain the first vector feature of the single sentence, and then the similarity between the first vector feature and the second vector feature of each parameter feature sentence corresponding to the target parameter is calculated to determine whether there is text similarity between the single sentence and any parameter feature sentence.

[0059] If the similarity between the first vector feature of the single sentence and the second vector feature of any parameter feature sentence is not lower than the similarity threshold, it indicates that the text similarity between the single sentence and any parameter feature sentence is relatively high. At this time, determine the single sentence as the target single sentence and end the traversal. If the similarity between the first vector feature of the single sentence and the second vector feature of any parameter feature sentence is lower than the similarity threshold, it indicates that the single sentence does not contain the target parameter and the possibility of containing the parameter information of the target parameter is also relatively low. At this time, abandon the single sentence and traverse to the next single sentence. Among them, the second vector feature of each parameter feature sentence corresponding to the target parameter is obtained by performing a predetermined vectorization process on the parameter feature sentence. For the specific implementation manner of the predetermined vectorization process, see steps E1 - E3 below and will not be elaborated here.

[0060] Optionally, perform sentence splitting on the target paragraph corresponding to the target parameter, that is, perform text segmentation on the target paragraph according to the specific characters included in the target paragraph. Among them, the above specific characters may include punctuation marks such as full stops and exclamation marks, or may also be words with a conjunction part of speech in the text. In this regard, the embodiments of the present application do not make specific limitations.

[0061] In this implementation manner, when determining whether there is a target single sentence in the target paragraph that matches any parameter feature sentence corresponding to the target parameter, through two judgment processes, that is, whether the string matches and whether the similarity is not lower than the similarity threshold, to reduce the possibility of missing or misextracting parameter information, thereby improving the extraction accuracy of parameter information and the extraction precision when extracting parameter information in the contract.

[0062] Among them, it can be determined whether there is a target single sentence in each single sentence of the target paragraph corresponding to the target parameter that matches any parameter feature sentence corresponding to the target parameter through keyword matching, or it can also be determined through regular matching. In this regard, the embodiments of the present application do not make specific limitations.

[0063] Optionally, in one implementation manner, when the target parameter is a parameter of the first category, in the above step S106, based on the target single sentence, extracting the parameter information of the target parameter in the target contract may include the following steps:

[0064] Step C: Analyze the target semantics represented by the target single sentence; if the target semantics is a negative semantics, use the first field value as the parameter information of the target parameter; otherwise, use the second field value as the parameter information of the target parameter.

[0065] In this implementation manner, as described above, since the parameter categories to which the target parameters belong are different, the presentation manners of the corresponding extracted parameter information are different. Therefore, the parameter categories involved in parameter information extraction can be classified to obtain a first category and a second category, and corresponding parameter information extraction manners can be set for different categories of parameters.

[0066] When the target parameter is a parameter of the first category, the parameter information extraction manner corresponding to this category is used for information extraction. Specifically: analyze the target semantics represented by the target single sentence. If the analyzed target semantics is a negative semantics, then use the first field value as the parameter information of the target parameter, that is, the set first field value is the field value used to represent the negative semantics. Exemplarily, the first field value is 0, and the target single sentence is "information is not reported", then the semantics represented by this target single sentence is a negative semantics. At this time, use 0 as the parameter information of the target parameter. That is to say, when presenting the numerical value 0 to the user, the user can clearly know that the meaning represented by this first field value is "information is not reported". If the analyzed target semantics is an affirmative semantics, then use the second field value as the parameter information of the target parameter, that is, the set second field value is the field value used to represent the affirmative semantics. Exemplarily, the second field value is 1, and the target single sentence is "information is reported", then the semantics represented by this target single sentence is an affirmative semantics. At this time, use 1 as the parameter information of the target parameter. That is to say, when presenting the numerical value 1 to the user, the user can clearly know that the meaning represented by this second field value is "information is reported".

[0067] That is to say, when the target parameter is a parameter of the first category, the extracted parameter information is not the text content existing in the target single sentence, but a judgment result obtained by judging whether the target parameter exists in the target single sentence. Among them, the above-mentioned parameters of the first category can be Boolean values, and the embodiments of the present application do not make specific limitations on this. In addition, the above-mentioned first field value and second field value can be specific numerical values, such as 1 and 0, or can be specific words, such as "yes" and "no", or can be English letters Y and N, etc., and the embodiments of the present application do not make specific limitations on this.

[0068] In this implementation manner, the target parameters whose extracted parameter information is not the text content existing in the target single sentence are classified into the parameters of the first category. Thus, a corresponding parameter extraction manner is set for the parameters of this first category to reduce the possibility of missing or mis-extracting the parameter information of this category of parameters, and thus the extraction scope when extracting the parameter information in the contract.

[0069] Optionally, in an implementation manner, when the target parameter is a parameter of the second category, in the above step S106, extracting the parameter information of the target parameter in the target contract based on the target single sentence may include the following steps:

[0070] Step D: Extract the content of the target single sentence according to the extraction position set for the parameter feature sentence matched for the target single sentence, so as to obtain the parameter information of the target parameter in the target contract.

[0071] In this implementation manner, when the target parameter is a parameter of the second category, the extracted parameter information is the text content existing in the target single sentence. At this time, it is necessary to clarify the position corresponding to the parameter information of the target parameter in the target single sentence. Therefore, the extraction position corresponding to the parameter information can be set in advance for each parameter feature sentence corresponding to the target parameter. Thus, when extracting the parameter information of the target parameter, the content of the target single sentence can be extracted according to the extraction position set for the parameter feature sentence matched for the target single sentence, so as to obtain the parameter information of the target parameter in the target contract.

[0072] Among them, the parameter of the second category above can be a string, and in this regard, the embodiments of the present application do not make specific limitations.

[0073] Optionally, in each parameter feature sentence, a corresponding string prefix and a string suffix are set for the extraction position corresponding to the parameter information. Thus, when extracting the parameter information, the string between the string prefix and the string suffix in the parameter feature sentence is directly extracted as the parameter information of the target parameter.

[0074] In this implementation manner, the target parameter whose extracted parameter information is the text content existing in the target single sentence is classified as a parameter of the second category. Thus, a corresponding parameter extraction method is set for the parameter of the second category, so as to reduce the possibility of missing extraction and mis-extraction of the parameter information of this category of parameters. Thus, the extraction accuracy of the parameter information is improved, and the extraction accuracy when extracting the parameter information in the contract is improved.

[0075] Optionally, in one embodiment, the predetermined vectorization process may include the following steps:

[0076] Step E1: Extract at least one keyword from the statement to be processed; wherein, the statement to be processed is any parameter feature sentence or any single sentence. Step E2: For each keyword, calculate a first value representing the occurrence frequency of the keyword in the target document, and calculate a second value representing the distribution of the keyword in the target document set. Based on the product value of the first value and the second value of the keyword, construct a vector feature corresponding to the keyword; wherein, when the statement to be processed is any single sentence, the target document is the target contract, and when the statement to be processed is any parameter feature sentence, the target document is a sample contract containing any parameter feature sentence; the target document set includes other documents belonging to the business field corresponding to the target document except the target document. Step E3: Take the vector matrix obtained by concatenating the vector features corresponding to each keyword as the vector feature corresponding to the statement to be processed; wherein, each matrix element in the vector matrix is used to represent the vector feature corresponding to a keyword.

[0077] In this embodiment, when performing the predetermined vectorization process, first extract at least one keyword from the statement to be processed. Among them, considering that any single sentence and any parameter feature sentence are both involved in performing the predetermined vectorization process to obtain the corresponding vector features, therefore, the above-mentioned statement to be processed is any parameter feature sentence or any single sentence.

[0078] Then, for each extracted keyword, calculate a first value representing the occurrence frequency of the keyword in the target document.

[0079] Optionally, in one implementation, in the above step E2, calculating the first value representing the occurrence frequency of the keyword in the target document may include the following steps:

[0080] Step F1: When the keyword belongs to the vocabulary of the business field corresponding to the target contract category, determine the word frequency weight corresponding to the business field of the target contract category, and calculate the occurrence frequency of the keyword in the target document, and take the product value of the calculated occurrence frequency and the word frequency weight as the first value representing the occurrence frequency of the keyword in the target document; and / or, Step F2: When the keyword does not belong to the vocabulary of the business field corresponding to the target contract category, calculate the occurrence frequency of the keyword in the target document, and take the calculated occurrence frequency as the first value representing the occurrence frequency of the keyword in the target document.

[0081] In this implementation, in order to reduce the impact of words with low relevance to the business field corresponding to the target contract category on the first value of words with high relevance to the business field corresponding to the target contract category, corresponding word frequency weights can be set in advance for the words in the business field corresponding to the target contract category. When calculating the first value of the keyword, it is first determined whether the keyword belongs to the words in the business field corresponding to the target contract category. Thus, based on the judgment result, the first value of the keyword is calculated using the corresponding calculation method of the first value.

[0082] When the keyword belongs to the words in the business field corresponding to the target contract category, determine the word frequency weight corresponding to the business field corresponding to the target contract category, and calculate the frequency of occurrence of the keyword in the target document. Then, take the product value of the calculated frequency of occurrence and the word frequency weight as the first value used to represent the frequency of occurrence of the keyword in the target document. Exemplarily, use the first formula to calculate the first value used to represent the frequency of occurrence of the keyword in the target document, where the first formula is:

[0083] ;

[0084] Wherein, represents the word frequency of the th keyword in the target document , that is, the first value used to represent the frequency of occurrence of the th keyword in the target document , represents the frequency of occurrence of the th keyword in the target document , represents the total number of all words in the target document , represents any word in the target document ; represents the word frequency weight corresponding to the th keyword.

[0085] When the keyword does not belong to the words in the business field corresponding to the target contract category, calculate the frequency of occurrence of the keyword in the target document, and take the calculated frequency of occurrence as the first value used to represent the frequency of occurrence of the keyword in the target document. Exemplarily, use the second formula to calculate the first value used to represent the frequency of occurrence of the keyword in the target document, where the second formula is:

[0086] .

[0087] Then, calculate the second value used to represent the distribution of the keyword in the target document.

[0088] Exemplarily, using the third formula, a second value for characterizing the distribution of the keyword in the target document set is calculated, where the third formula is:

[0089] ;

[0090] where represents the inverse document frequency of the th keyword, that is, the second value for characterizing the distribution of the th keyword in the target document set , represents the total number of documents included in the target document set , represents the total number of documents in the target document set that contain the th keyword; where the logarithmic function uses the natural logarithm.

[0091] After that, based on the product value of the first value and the second value of the keyword, a vector feature corresponding to the keyword is constructed. Wherein, when the to-be-processed statement is any single sentence, the above-mentioned target document is the target contract for which parameter information is to be extracted; when the to-be-processed statement is any parameter feature sentence, the above-mentioned target document is the sample contract containing any parameter feature sentence. Wherein, the above-mentioned target document set includes other documents belonging to the business field corresponding to the target document except the target document. Exemplarily, the above-mentioned other documents may be at least one document such as a book or a report containing vocabulary related to the business field corresponding to the target document for expanding domain-related professional vocabulary, and in this regard, the embodiments of the present application do not make specific limitations.

[0092] In this way, after obtaining the vector feature corresponding to each keyword, the vector matrix obtained by splicing the vector features corresponding to each keyword is used as the vector feature corresponding to the to-be-processed statement. In this vector matrix, each matrix element is used to represent the vector feature corresponding to a keyword, and each row of matrix elements is used to represent the vector feature corresponding to a to-be-processed statement. It can be understood that a row of matrix elements is a set of vector features corresponding to each keyword included in a to-be-processed statement.

[0093] Optionally, for the same contract category, the vector features of the parameter feature sentences corresponding to the parameters of the contract category are stored in a vector set for subsequent similarity calculation.

[0094] In this embodiment, based on the judgment result of whether a keyword belongs to the vocabulary of the business field corresponding to the target contract category, the corresponding calculation method of the first value is called for targeted calculation. Moreover, by setting the word frequency weight corresponding to the business field of the target contract category, when calculating the first value representing the frequency of occurrence of the keyword in the target document, the above word frequency weight is used to weight the vocabulary belonging to the business field corresponding to the target contract category, so as to reduce the influence of the vocabulary with low relevance to the business field corresponding to the target contract category on the vocabulary with high relevance to the business field corresponding to the target contract category obtained by calculation.

[0095] Optionally, in an embodiment, in a parameter information extraction method provided by an embodiment of the present application, the following steps may further be included:

[0096] Step H: In response to meeting the update condition, perform an update process on the word frequency weight and the similarity threshold;

[0097] Among them, the update process includes the following steps:

[0098] Step G1: Obtain each candidate combination; each candidate combination includes a candidate weight and a candidate threshold. Step G2: Determine the standby single sentences that meet the screening conditions from each standby single sentence to obtain at least one historical single sentence to be utilized; that any standby single sentence meets the screening conditions includes: when the standby single sentence performs string matching with the corresponding historical parameter feature sentences, all matches fail; the corresponding historical parameter feature sentence of any standby single sentence is: the parameter feature sentence corresponding to the parameter to which the parameter information extracted from the standby single sentence belongs. Step G3: For each candidate combination, based on the candidate combination, determine the target sentence attribute of each historical single sentence; where the target sentence attribute includes the first attribute, the second attribute, or the third attribute. Step G4: For each candidate combination, based on the first quantity of the historical single sentences belonging to the first attribute, the second quantity of the historical single sentences belonging to the second attribute, and the third quantity of the historical single sentences belonging to the third attribute, calculate the harmonic mean of the candidate combination for representing the accuracy rate and recall rate of parameter information extraction. Step G5: Determine the candidate combination corresponding to the calculated maximum harmonic mean as the target combination, and use the candidate weight and candidate threshold in the target combination as the updated word frequency weight and the updated similarity threshold respectively.

[0099] In this embodiment, by setting the update condition, the word frequency weight and the similarity threshold are updated to obtain the updated word frequency weight and the similarity threshold, so that the accuracy of parameter information extraction is improved based on the updated word frequency weight and the similarity threshold. The above-mentioned update condition may include: the current time point reaches the predetermined update time point, the update instruction sent by the user is received, and the parameter feature sentence constructed for the target parameter in the target contract category is detected. At least one of the conditions has data changes, and the embodiment of the present application does not make specific limitations on this.

[0100] The above-mentioned updating process is specifically as follows: obtaining each candidate combination, each candidate combination includes a candidate weight and a candidate threshold, then using each target sentence determined in the contract from which parameter information has been extracted as each backup sentence, then determining the backup sentence that meets the screening conditions from the acquired backup sentences, and obtaining at least one historical sentence to be used. It should be noted that the word frequency weight and similarity threshold only involve the process of vectorization and similarity matching after the string matching between the single sentence and the corresponding parameter feature sentence fails, therefore, each historical sentence obtained is a backup sentence that fails to match the string matching between the corresponding historical parameter feature sentences.

[0101] Then, for each candidate combination, based on the candidate combination, the target sentence attribute of each historical sentence is determined, that is, the target sentence attribute of the accuracy of the parameter information extracted from each historical sentence under the candidate combination is determined. Wherein, the above-mentioned target sentence attribute includes a first attribute, a second attribute or a third attribute, and the first attribute, the second attribute and the third attribute are all used to characterize the accuracy of the parameter information extracted from the historical sentence, and the accuracy they characterize is different. In this way, for each candidate combination, based on the first number of historical sentences belonging to the first attribute, the second number of historical sentences belonging to the second attribute, and the third number of historical sentences belonging to the third attribute, the harmonic mean of the candidate combination for characterizing the accuracy and recall rate of parameter information extraction is calculated. Thus, based on each harmonic mean calculated, the candidate combination corresponding to the harmonic mean with the largest value calculated is used as the target combination, and the candidate weight in the target combination is used as the updated word frequency weight, and the candidate threshold in the target combination is used as the updated similarity threshold, so that the word frequency weight and similarity threshold are updated.

[0102] Exemplarily, the historical sentence with the first attribute has a true positive attribute in the candidate combination, the historical sentence with the second attribute has a false positive attribute in the candidate combination, and the historical sentence with the third attribute has a false negative attribute in the candidate combination.

[0103] Optionally, for each candidate combination, the historical single sentence of the first attribute determined under this candidate combination can be added to each historical parameter feature sentence corresponding to this historical single sentence, so as to achieve the expansion of the parameter feature sentence of the parameter corresponding to this historical single sentence when this candidate combination is the target combination.

[0104] Optionally, for each candidate combination, the historical single sentence of the second attribute determined under this candidate combination can be added to the blacklist expansion of the parameter feature sentence corresponding to the parameter corresponding to this historical single sentence.

[0105] Optionally, for each candidate combination, the historical single sentence of the third attribute determined under this candidate combination can be added to each historical parameter feature sentence corresponding to this historical single sentence, so as to achieve the expansion of the parameter feature sentence of the parameter corresponding to this historical single sentence when this candidate combination is the target combination.

[0106] Optionally, within the numerical range corresponding to the candidate threshold, multiple candidate thresholds are obtained according to the first step size; within the numerical range corresponding to the candidate weight, multiple candidate weights are obtained according to the second step size; the multiple obtained candidate thresholds and multiple candidate weights are combined pairwise to obtain each candidate combination. Among them, each candidate combination is different, and in any two candidate combinations, at least one value is different. For example, at least one of the candidate weights and candidate thresholds in different candidate combinations is different.

[0107] Optionally, the word frequency weights corresponding to different target parameters can be different, and each word frequency weight is adjusted according to the corresponding predetermined ratio based on the updated word frequency weight; the similarity thresholds corresponding to different target parameters can also be different, and each similarity threshold is adjusted according to the corresponding predetermined ratio based on the updated similarity threshold.

[0108] In this embodiment, by continuously updating the word frequency weight and the similarity threshold, the adaptability of the word frequency weight and the similarity threshold is improved, and further, the accuracy of parameter information extraction is improved.

[0109] Optionally, in one embodiment, the determination method of the target statement attribute of each historical single sentence may include the following steps:

[0110] Step J1: Using the candidate weights in the candidate combination as the word frequency weights utilized in the predetermined vectorization process, perform a predetermined vectorization process on the historical single sentence to obtain the third vector feature of the historical single sentence. Step J2: Calculate the similarity between the third vector feature of the historical single sentence and the fourth vector features of the respective historical parameter feature sentences. Step J3: Determine whether there is a target similarity among the calculated similarities that meets the candidate threshold in the candidate combination. Step J4: If there is, output the predetermined parameter verification information so that the user can verify whether the historical single sentence is similar to the historical parameter feature sentence corresponding to the target similarity; in response to the confirmation instruction sent by the user to indicate similarity, determine that the target sentence attribute of the historical single sentence under the candidate combination is the first attribute; in response to the confirmation instruction sent by the user to indicate dissimilarity, determine that the target sentence attribute of the historical single sentence under the candidate combination is the second attribute. Step J5: If not, output the information indicating that the parameter extraction fails so that the user can confirm whether the candidate threshold in the candidate combination is too high; in response to the confirmation instruction sent by the user to indicate that the threshold is too high, determine that the target sentence attribute of the historical single sentence under the candidate combination is the third attribute.

[0111] In this implementation manner, when determining the target statement attribute under a candidate combination for each historical single sentence, first, the candidate weight in the candidate combination is used as the word frequency weight utilized in the predetermined vectorization process, and the predetermined vectorization process is performed on the historical single sentence to obtain the third vector feature of the historical single sentence; then, the similarity between the third vector feature of the historical single sentence and the fourth vector features of the corresponding historical parameter feature sentences is calculated. Thus, based on the calculated similarity, it is determined whether there is a target similarity among the similarities that meets the candidate threshold in the candidate combination. If there is, the predetermined parameter verification information is output so that the user can verify whether the historical single sentence is similar to the historical parameter feature sentence corresponding to the target similarity, that is, manual intervention is required here to verify whether the historical single sentence is similar to the historical parameter feature sentence corresponding to the target similarity. Thus, based on the user's own judgment, a confirmation instruction indicating whether it is similar is sent, and based on the received confirmation instruction sent by the user, the target statement attribute of the historical single sentence under the candidate combination is determined. Specifically: in response to the confirmation instruction sent by the user indicating similarity, it is determined that the target statement attribute of the historical single sentence under the candidate combination is the first attribute; or, in response to the confirmation instruction sent by the user indicating dissimilarity, it is determined that the target statement attribute of the historical single sentence under the candidate combination is the second attribute. If not, it indicates that the string matching between the historical single sentence and the corresponding historical parameter feature sentences fails, and the similarity matching also fails. At this time, the information indicating that the parameter information extraction fails is output so that it can be used to confirm whether the candidate threshold in the candidate combination is too high, that is, there is actually a target similarity, but due to the candidate threshold being set too high, the final determination result is that there is none. Thus, in response to the confirmation instruction sent by the user indicating that the threshold is too high, it is determined that the target statement attribute of the historical single sentence under the candidate combination is the third attribute.

[0112] Optionally, when there is none, the user can also send a confirmation instruction indicating that the threshold is reasonable, that is to say, the historical single sentence does not match the corresponding historical parameter feature sentences. At this time, it can be determined that the target statement attribute of the historical single sentence under the candidate combination is the fourth attribute. That is to say, compared with the above three attributes, the accuracy represented by the fourth attribute is the lowest. For example, the historical single sentence with the fourth attribute is the statement attribute being a true negative under a candidate combination.

[0113] In this embodiment, the determination of the statement attributes of the same historical single sentence under different candidate combinations provides a data basis for calculating the harmonic mean under each candidate combination, so as to improve the determination efficiency of the target combination.

[0114] Next, a specific embodiment is used to exemplarily introduce the solution provided by the embodiment of the present application.

[0115] To address the deficiencies in existing general contract information extraction methods, in an automated method and system for full-process fund contract element extraction and registration provided in this specific embodiment (i.e., a parameter information extraction method provided in the embodiments of this application), first, a hybrid method based on rules (i.e., string matching in the embodiments of this application) and text similarity (i.e., similarity in the embodiments of this application) is used to extract key parameters (parameter information in the embodiments of this application) in the fund contract, so as to solve the problems of narrow extraction scope and low extraction accuracy of parameter information. In addition, through Robotic Process Automation (RPA), the key parameters extracted are entered into the fund registration system (contract registration system in the embodiments of this application) by means of input, selection, etc. That is to say, RPA is used as a bridge between the contract system and the fund registration system to break the information silo problem existing between different systems, and at the same time improve the input accuracy when key parameters are entered into the fund registration system and save labor costs.

[0116] The implementation process of this specific embodiment involves RPA, that is, RPA is used to extract parameter information. According to the functions involved in the parameter information extraction process, two functional modules are set: a contract information extraction module, which correspondingly includes a contract download sub-module, a contract text preprocessing sub-module, a parameter matching and extraction sub-module, etc.; a parameter entry module (which can also be called a parameter maintenance module). Through this parameter entry module, the parameter information extracted by the contract information extraction module can be automatically entered into at least one registration page in the fund registration system.

[0117] The following combines Figure 2 the schematic flowchart of a specific embodiment provided in the embodiments of this application shown below to specifically introduce the method flow involved in this specific embodiment. The method flow may include the following steps:

[0118] Before executing this method flow, the following steps can be pre-executed: offline build a custom corpus. The custom corpus includes a chapter marker table (location information in the embodiments of this application), the characteristic sentences of each parameter (parameter characteristic sentences corresponding to the parameters in the embodiments of this application), and the TF-IDF (term frequency-inverse document frequency) vector matrix of each characteristic sentence (vector characteristics of the parameter characteristic sentences in the embodiments of this application), etc.

[0119] Exemplarily, taking the fund contract as an example, the data values of the key parameters to be extracted from this type of contract (which can also be referred to as the data values of the parameters to be registered, that is, the parameter information of the target parameters in the embodiments of the present application) are approximately more than 130 items, and each key parameter corresponds to dozens of parameter feature sentences (which can also be referred to as feature statements). Therefore, the contract information extraction module can be called in advance to offline construct a feature sentence corpus (which can also be referred to as a custom corpus) according to the classification of fund contracts. This corpus includes various corpus tables:

[0120] (1) Chapter table (which can also be referred to as the chapter marking table, or chapter feature table; that is, the location information in the embodiments of the present application): Since the chapter positions where the key parameters are located are fixed, through this chapter table, the positions where the key parameters are located (the text positions of the parameter information of the target parameters in the present application in the contract) can be quickly located to improve the extraction efficiency. Moreover, this chapter table also includes fields such as chapter start and end feature sentences for paragraph segmentation.

[0121] (2) Parameter feature sentence table (that is, the parameter feature sentence set in the present application): It is used to record the feature sentences of each parameter. In this parameter feature sentence table, according to the extracted parameter types, the corresponding parameter feature sentences can be divided into three types: single sentence, prefix, and suffix. Exemplarily, if only the existence of the key parameter needs to be judged, a single sentence can be constructed; if strings such as names and numbers need to be extracted from the text, prefixes and suffixes are constructed for the parameter feature sentences corresponding to the parameters.

[0122] (3) Parameter feature sentence blacklist table (that is, the parameter feature sentence blacklist set in the present application): It is used to record the contract single sentences that meet the following conditions. That is, according to the similarity result value between the contract single sentence and the parameter feature sentence, the contract single sentence that meets the similarity threshold but is not a parameter feature sentence, that is, the calculated cosine similarity exceeds the similarity threshold, but the meaning represented by this contract single sentence is not the meaning represented by the corresponding parameter feature sentence.

[0123] (4) Parameter feature sentence vector table (that is, the vector set in the present application): It is used to record the TF-IDF vector matrix of each feature sentence; the TF-IDF vector matrix of each feature sentence is obtained by vectorizing the existing parameter feature sentences in the parameter feature sentence table through the TF-IDF algorithm. Specifically:

[0124] Step a: Preprocess the parameter feature sentences. This preprocessing includes word segmentation and stop word removal; this step can remove data noise to improve the accuracy of subsequent data processing;

[0125] Step b: For each word in the parameter feature sentence, calculate the word frequency of the word (i.e., the first value in the embodiments of the present application), that is, calculate the occurrence frequency of the word in the current document, and perform normalization processing on the calculated occurrence frequency. Among them, use the above second formula for normalization processing.

[0126] It should be noted that considering that the fund contract belongs to a specific financial field, a domain knowledge base of all parameters related to this field can be constructed to strengthen the word frequency weight of specific words (i.e., the vocabulary in the business field corresponding to the target contract category in the embodiments of the present application), thereby reducing the influence of words with lower relevance to this field on the word frequency of each word in the parameter feature sentence. Among them, use the above first formula for normalization processing. Among them, is a custom parameter greater than 1 (i.e., the word frequency weight in the embodiments of the present application), which is used to strengthen the words belonging to the domain knowledge base.

[0127] Step c: For each word in the parameter feature sentence, calculate the inverse document frequency of the word (i.e., the second value in the embodiments of the present application). Among them, use the above third formula to calculate the inverse document frequency of the word. It should be noted that in the document set involved in calculating the inverse document frequency of the word, in addition to the current document where the word is located, it includes other documents belonging to the business field corresponding to the current document. Exemplarily, the above other documents can be at least one document such as historical private equity fund contracts, the Fund Law, fund reports, industry research reports, etc. for expanding professional vocabulary related to the field. In this regard, the embodiments of the present application do not make specific limitations.

[0128] Step d: For each word in the parameter feature sentence, calculate the TF-IDF value of the word (i.e., the product value of the first value and the second value in the embodiments of the present application). Among them, use the following formula to calculate the TF-IDF value of the word;

[0129] .

[0130] Step e: For the TF-IDF value of each word in the parameter feature sentence, construct a TF-IDF vector matrix of the parameter feature sentence.

[0131] The construction of the above custom corpus provides data preparation for subsequent extraction of key parameters from contracts.

[0132] Then, execute the process of the method flow as Figure 2 shown. The method flow may include the following steps:

[0133] S201: Automatically log in to the contract system and the registration system. Specifically: The robot applying RPA technology can automatically log in to the contract system and the registration system (i.e., the fund registration system in the embodiments of the present application) at a preset specific time.

[0134] S202: The contract system screens the fund contracts to be processed that meet the conditions and downloads them. Specifically: The robot screens the fund products to be registered from the contract system and downloads the fund contracts corresponding to the fund products to be registered.

[0135] S203: Read the contract text and perform preprocessing for sentence splitting. Specifically: Obtain the full text data of the downloaded fund contract; since the parameters to be extracted in the fund contract are distributed in fixed chapters, the full text data of the fund contract can be segmented into several paragraphs according to the chapter table in the custom corpus. Then, for each paragraph, perform text preprocessing according to full stops, exclamation marks, question marks, etc., and divide it into several short sentences of the paragraph.

[0136] S204: Obtain the key parameters one by one. Specifically, obtain the parameter feature sentence corresponding to each key parameter from the custom corpus, and then, in the paragraph corresponding to each key parameter, use the parameter feature sentence corresponding to the key parameter to match each short sentence obtained by the division, and determine the key sentence that matches successfully (i.e., the target single sentence in the embodiments of the present application).

[0137] For each short sentence obtained by the division, determine whether there is a corresponding parameter feature sentence in the short sentence. If so, perform parameter extraction on the short sentence and execute step S205; if not, perform word segmentation on the single sentence and vectorize it to obtain the vector of the short sentence. Then, calculate the cosine similarity between the short sentence and the corresponding parameter feature sentence. If the obtained cosine similarity is greater than the similarity threshold, perform parameter extraction on the short sentence and execute step S205; if not, execute step S206 to feedback to the business personnel that the business personnel have not extracted the key parameters. That is to say, if there is no corresponding parameter feature sentence in the short sentence, perform word segmentation and vectorization processing on the short sentence, calculate the similarity between the vector of the short sentence and the vector of the corresponding parameter feature sentence, and when the obtained similarity is greater than the similarity threshold, determine the short sentence as the key sentence for parameter extraction.

[0138] Among them, the data value types of the parameters to be extracted (key parameters) are boolean values and strings. Among them, if it is necessary to judge whether there is a parameter feature sentence in the contract, a boolean value is output; if it is necessary to extract specific parameters from the contract, a string is output. Exemplarily, Figure 3 is a schematic flowchart of a parameter extraction process using a boolean value type provided by the embodiments of the present application, Figure 4A schematic flowchart of a process for extracting parameters of string type provided by an embodiment of the present application. For the sake of easy reading, Figure 3 and Figure 4 The specific implementation manners shown will be described in detail in the following steps S301 - S307 and steps S401 - S408.

[0139] S205: Register one by one into the registration system. Specifically, after extracting the key parameters one by one, RPA can automatically log in to the registration system to determine whether the fund product corresponding to the fund contract needs to be maintained, that is, to determine whether the fund contract has been logged in to the registration system. If the fund product corresponding to the fund contract has not been maintained, a new product page will be created and the extracted key parameters will be registered one by one into the new product page. Among them, the new product page may include: parameter pages such as basic information, product raising, product storage, etc.

[0140] Exemplarily, Figure 5 A schematic diagram of a product interface provided by an embodiment of the present application. As Figure 5 shown, the product information that can be set in this new product page (which can also be called the interface for product information maintenance) includes: basic information, product raising, product existence, product trading restrictions, net value liquidation rules, product agency relationship, etc. Taking the parameter page corresponding to the basic information as an example, this parameter page may include: product code, product short name, product full name, custodian, investment trading system, payment channel, net value type, valuation system, product type, fee category, reporting method, etc. And, a note box can be set in each parameter page. In this way, the extracted key parameters are registered one by one into the value input boxes of the corresponding items in this new product page to complete the registration of the key parameters. Among them, the check relationship between the parameter categories and parameter values is preset, that is, the corresponding relationship between the key parameters and the corresponding items of the key parameters. Thus, through the page element recognition of the RPA technology, each extracted key parameter is entered into the corresponding position in turn. It should be noted that the input operation includes computer operations such as input, selection, shortcut keys, text setting, etc., which are not specifically limited in this application.

[0141] S206: Feedback to the business personnel.

[0142] S207: The business personnel trigger the process of optimizing the weights of domain words and the similarity threshold according to historical data. Among them, the above process of optimizing the weights of domain words and the similarity threshold according to historical data is the process of updating the word frequency weights and the similarity threshold in the embodiments of the present application in response to meeting the update conditions, and obtaining the updated word frequency weights and similarity threshold. Specifically: In this specific embodiment, a robot can be used to monitor the parameter extraction process in real time, and analyze the similarity calculation results according to the basic analysis method of the confusion matrix. Thus, a parameter optimization feedback mechanism based on historical data is established. For example, for a certain key parameter y, in the process of matching and calculating the similarity between the contract short sentences (i.e., the historical single sentences in the embodiments of the present application) in the historical data and the corresponding parameter feature sentences one by one through the above mechanism, the following three situations will occur: (1) The rule matching is successful, and the extraction of the key parameter y is completed; among them, the above rule matching is the string matching in the embodiments of the present application; (2) The rule matching fails, but the calculated similarity is not lower than the similarity threshold. At this time, a notice is sent to the business personnel to verify the data accuracy of the key parameter y. If the data accuracy of the key parameter y is verified to be correct, that is, the statement attribute of the contract short sentence containing the key parameter y is true positive, the contract short sentence corresponding to the key parameter y is added to the parameter feature sentence table; if the data accuracy of the key parameter y is verified to be incorrect, at this time, it can be determined that the set similarity threshold is too low, and the statement attribute of the contract short sentence containing the key parameter y under this similarity threshold is false positive, and the contract short sentence corresponding to the key parameter y is added to the parameter feature sentence blacklist table; (3) The rule matching fails, and the calculated similarity is lower than the similarity threshold. At this time, a notice is sent to the business personnel to remind that the extraction of the key parameter y fails. If the matching fails due to the set similarity threshold being too high, the statement attribute of the contract short sentence containing the key parameter y under this similarity threshold is false negative, and the contract short sentence corresponding to the key parameter y is added to the parameter feature sentence table; otherwise, it is determined that the contract short sentence does not contain the key parameter y, and the statement attribute of the contract short sentence containing the key parameter y under this similarity threshold is true negative.

[0143] It should be noted that for the custom parameters, within the corresponding value range of the custom parameters, iterative adjustment is performed with the first step size to ensure that the specific words included in each parameter feature sentence have better weight values, so as to prevent the occurrence of situations where the adjustment effect is not obvious or overfitting. Among them, the above first step size can be values such as 0.1, 0.2, 0.3, 0.05, 0.55, etc. For the similarity threshold, within the corresponding value range of the similarity threshold, iterative adjustment is performed with the second step size to balance the accuracy and recall rate of the sample classification model, that is, the accuracy and recall rate of parameter information extraction in the embodiments of the present application. Among them, the above second step size can be values such as 0.01, 0.02, 0.03, 0.1, 0.2, etc.

[0144] The above accuracy rate refers to the first ratio of the contract short sentences of historical data predicted as positive examples (the word attribute is true positive) that are actually positive examples. The formula is as follows:

[0145] ;

[0146] Among them, represents the first ratio, represents the number of contract short sentences with the word attribute of true positive in the contract short sentences of historical data, represents the number of contract short sentences with the word attribute of false positive in the contract short sentences of historical data.

[0147] The recall rate refers to the second ratio of the contract short sentences of historical data that are actually positive examples and are correctly predicted as positive examples. The formula is as follows:

[0148] ;

[0149] Among them, represents the second ratio, represents the number of contract short sentences with the word attribute of false negative in the contract short sentences of historical data.

[0150] Considering that the values of the custom parameters and the similarity threshold affect each other, therefore, in this specific embodiment, by means of the concept of performance parameters for measuring the sample classification model, through cross-validation, among different combinations of custom parameters and similarity thresholds, the combination corresponding to the highest value of the harmonic mean of the calculated accuracy rate and recall rate is determined as the best combination. The specific formula is as follows:

[0151] ;

[0152] Among them, represents the harmonic mean calculated under the th combination of custom parameters and similarity thresholds, represents the The first ratio calculated under a combination of custom parameters and a similarity threshold; Characterizing the The second ratio calculated under a combination of custom parameters and a similarity threshold.

[0153] The following combines Figure 3 to specifically describe the parameter extraction process using boolean-type parameters. As Figure 3 shown, the specific process when it is necessary to determine whether there is a parameter feature sentence in the contract includes:

[0154] S301: Locate the section where the parameter is located and obtain the segmented paragraph short sentences. Specifically, after locating the section where the parameter is located, obtain the paragraph text where the parameter is located (i.e., the step of determining the target paragraph corresponding to the target parameter in the embodiment of the present application). Then, preprocess the paragraph text and segment the paragraph text into several short sentences according to punctuation marks such as full stops and exclamation marks (i.e., each single sentence of the target paragraph in the embodiment of the present application).

[0155] S302: Match each short sentence with the corresponding parameter feature sentence one by one and determine whether the match is successful; if so, execute step S303; if not, execute step S304. Specifically: For each short sentence, perform string matching using the short sentence and the corresponding parameter feature sentence. If the match is successful, it indicates that the short sentence contains the corresponding parameter feature sentence. At this time, execute step S303 and output a field for characterizing correctness, such as the value 1, or the character ture (correct), etc.; if the match fails, it indicates that based on the string, it cannot be determined that the short sentence contains the corresponding parameter feature sentence. At this time, execute step S304 and perform a similarity matching process on the short sentence and the corresponding parameter feature sentence.

[0156] S303: Output a field for characterizing correctness. Among them, the field for characterizing correctness is the second field value in the embodiment of the present application.

[0157] S304: Perform vectorization processing on the short sentence. Specifically: Perform the above-mentioned vectorization processing on the short sentence using the TF-IDF algorithm to obtain the vector of the short sentence. Among them, the process of performing vectorization processing on the short sentence is similar to the above steps a - e and will not be elaborated here.

[0158] S305: Calculate the similarity between the vector of the short sentence and the vectors of the corresponding parameter feature sentences one by one. Specifically: calculate the cosine value of the angle between the vector of the short sentence and the vector of the corresponding parameter feature sentence to determine the similarity between the two vectors. Among them, the obtained cosine value approaching 1 indicates that the two vectors are closer, the angle approaches 0, and the similarity between the short sentence and the corresponding parameter feature sentence is higher; the obtained cosine value approaching 0 indicates that the two vectors are less similar, the angle approaches 90 degrees, and the similarity between the short sentence and the corresponding parameter feature sentence is lower. Exemplarily, the similarity between the two vectors is calculated using the following formula,

[0159] ;

[0160] Among them, represents the similarity between the th short sentence and the th parameter feature sentence of the corresponding parameter feature sentence, represents the vector of the th short sentence, represents the vector of the th parameter feature sentence, represents the modulus length of the th short sentence, represents the modulus length of the th parameter feature sentence. The numerator in the above formula is the dot product of the two vectors, and the denominator is the product of the modulus lengths of the two vectors.

[0161] S306: Determine whether the calculated similarity exceeds the similarity threshold; if so, execute step S303; if not, execute step S307. Specifically: when the calculated similarity reaches the similarity threshold θ, execute step S303. If not, it indicates that there is no parameter feature sentence in the parameter feature sentences corresponding to the short sentence that matches the short sentence. At this time, execute step S307.

[0162] S307: Output the field used to represent the error. Among them, the field used to represent the error is the first field value in the embodiments of the present application. This ends, that is, the process of extracting parameters using the boolean value type ends.

[0163] The following specifically describes the process of extracting parameters using the string type in combination with Figure 4 , as shown in Figure 4 , the specific process of extracting specific parameters from the contract includes:

[0164] S401: Locate the section where the parameter is located and obtain the segmented paragraph short sentences. The specific implementation method of this step S401 is similar to the above step S301 and will not be elaborated.

[0165] S402: Match each short sentence with the corresponding parameter feature sentence one by one and determine whether the match is successful. If so, execute step S403; if the match fails, execute step S404. Specifically: for each short sentence, perform string matching using the short sentence and the corresponding parameter feature sentence. If the match is successful, it indicates that the short sentence contains the corresponding parameter feature sentence. At this time, execute step S403 and extract the string from the short sentence using regular extraction. If the match fails, it indicates that based on string matching, it is impossible to determine that the short sentence contains the corresponding parameter feature sentence. At this time, execute step S404 and perform a similarity matching process on the short sentence and the corresponding parameter feature sentence.

[0166] S403: Extract the string using regular extraction. The extracted string is the parameter information of the target parameter in the embodiment of the present application.

[0167] S404: Vectorize the short sentence. The specific implementation method of this step S404 is similar to the above step S304 and will not be elaborated.

[0168] S405: Calculate the similarity between the vector of the short sentence and the vector of the corresponding parameter feature sentence one by one. The specific implementation method of this step S405 is similar to the above step S305 and will not be elaborated.

[0169] S406: Determine whether the calculated similarity exceeds the similarity threshold. If so, execute step S407; if not, execute step S408. Specifically: when the calculated similarity reaches the similarity threshold θ, execute step S407. If not, it indicates that there is no parameter feature sentence in the parameter feature sentence corresponding to the short sentence that matches the short sentence. At this time, execute step S408.

[0170] S407: Determine that the short sentence is an approximate feature sentence similar to the corresponding parameter feature sentence. The determined approximate feature sentence is the target single sentence in the embodiment of the present application.

[0171] S408: Output the parameter as an empty string for warning. This ends the process of extracting parameters using string type.

[0172] It should be noted that based on the above steps S301 - S307 and steps S401 - S408, it can be seen that vectorizing the short sentence is repeatedly used for multiple parameter extractions. Optionally, before parameter extraction, preprocess the paragraph text, divide it into multiple short sentences, and vectorize each divided short sentence to obtain the vector of each short sentence. And the time complexity of the parameter extraction algorithm is only O （n）, that is, the entire parameter extraction process only needs to loop once, and the vectors of the feature sentences can be pre-generated offline without having to be repeatedly constructed during the loop, so as to improve the real-time performance of parameter extraction.

[0173] In this specific embodiment, based on the hybrid method of rules and text similarity, the advantages of rule extraction and text similarity calculation are fully combined. The first is the efficiency of parameter extraction for contracts, that is, rule extraction can usually quickly identify and extract information in specific patterns, while text similarity calculation can more accurately capture semantic similarities and handle cases that do not conform to the rules but are semantically similar. Combining the two can improve the accuracy of extraction while ensuring efficiency. The second is the adaptability of parameter extraction for contracts, that is, the complexity and diversity of fund contracts pose greater challenges to information extraction. The above hybrid method can flexibly adjust the rules and similarity calculation weights according to specific parameters. For certain specific extraction requirements, more reliance can be placed on rule extraction; for extraction tasks that require a deeper understanding of text semantics, more reliance can be placed on text similarity extraction. The combination of the two has better adaptability and can more comprehensively and accurately extract information that has both fixed rules and semantic changes. Moreover, the maintenance of fund parameters (that is, entering the key parameters extracted into the fund registration system) is completed by the robot, which is not affected by objective factors such as personnel flow and professional capabilities, fully releasing the operation human resources. At the same time, the high efficiency and accuracy of RPA also help to reduce losses and correction costs caused by human errors, greatly improving the accuracy of fund parameter entry.

[0174] In addition, the robot can also achieve real-time monitoring and process optimization of the parameter extraction process. Through real-time monitoring of the contract information extraction process, the robot can timely discover and solve problems to ensure the stability and reliability of the business process. At the same time, the robot can also continuously optimize the extraction rules and algorithms based on historical data and real-time feedback to improve the processing effect and efficiency.

[0175] Corresponding to the above method embodiment, the embodiment of the present application also provides a parameter information extraction device, as Figure 6 shown. The device includes:

[0176] A contract acquisition module 610, configured to acquire a target contract for which parameter information extraction is to be performed;

[0177] A positioning information determination module 620, configured to determine target positioning information corresponding to the target contract category to which the target contract belongs; wherein, the target positioning information is used to indicate target parameters to be extracted for contracts of the target contract category, and the text positions of the parameter information of the target parameters in any contract;

[0178] A paragraph determination module 630, configured to locate, in the target contract, a paragraph containing parameter information of the target parameter based on the target location information, so as to obtain a target paragraph corresponding to the target parameter;

[0179] A feature sentence acquisition module 640, configured to acquire a plurality of parameter feature sentences corresponding to the target parameter; wherein, each parameter feature sentence corresponding to the target parameter is: a candidate sentence for describing the parameter information of the target parameter in the contract of the target contract category; among the plurality of parameter feature sentences corresponding to the target parameter, different parameter feature sentences are different;

[0180] A single sentence analysis module 650, configured to analyze whether there is a target single sentence in each single sentence of the target paragraph corresponding to the target parameter that matches any one of the parameter feature sentences corresponding to the target parameter;

[0181] An information extraction module 660, configured to, in response to the existence of a matching target single sentence, extract the parameter information of the target parameter in the target contract based on the target single sentence.

[0182] Optionally, in an implementation manner, the single sentence analysis module 650 is specifically configured to: perform a clause splitting process on the target paragraph corresponding to the target parameter to obtain a plurality of single sentences of the target paragraph; traverse the plurality of single sentences, and when traversing to each current single sentence, perform the following process: perform a string matching between the single sentence and each parameter feature sentence in the plurality of parameter feature sentences corresponding to the target parameter; if the single sentence matches any one of the parameter feature sentences, determine that the single sentence is a target single sentence and end the traversal; otherwise, perform a predetermined vectorization process on the single sentence to obtain a first vector feature of the single sentence, and calculate a similarity between the first vector feature and a second vector feature of each parameter feature sentence corresponding to the target parameter; wherein, the second vector feature of each parameter feature sentence corresponding to the target parameter is obtained by performing the predetermined vectorization process on the parameter feature sentence; if the similarity between the first vector feature of the single sentence and the second vector feature of any one of the parameter feature sentences is not lower than a similarity threshold, determine that the single sentence is a target single sentence and end the traversal, otherwise, traverse to the next single sentence.

[0183] Optionally, in one implementation, when the target parameter is a parameter of the first category, the information extraction module 660 is specifically configured to: analyze the target semantics represented by the target single sentence; if the target semantics is a negative semantics, use the first field value as the parameter information of the target parameter; otherwise, use the second field value as the parameter information of the target parameter; and / or, when the target parameter is a parameter of the second category, the information extraction module 660 is specifically configured to: extract the content of the target single sentence according to the extraction position set for the parameter feature sentence matched by the target single sentence, so as to obtain the parameter information of the target parameter in the target contract.

[0184] Optionally, in one implementation, the predetermined vectorization process includes: extracting at least one keyword from the sentence to be processed; where the sentence to be processed is any parameter feature sentence or any single sentence; for each keyword, calculating a first value used to represent the occurrence frequency of the keyword in the target document, and calculating a second value used to represent the distribution of the keyword in the target document set, and constructing a vector feature corresponding to the keyword based on the product value of the first value and the second value of the keyword; where when the sentence to be processed is any single sentence, the target document is the target contract, and when the sentence to be processed is any parameter feature sentence, the target document is a sample contract containing the any parameter feature sentence; the target document set includes other documents belonging to the business field corresponding to the target document except the target document; taking the vector matrix obtained by splicing the vector features corresponding to each keyword as the vector feature corresponding to the sentence to be processed; where each matrix element in the vector matrix is used to represent the vector feature corresponding to a keyword.

[0185] Optionally, in one implementation, calculating the first value used to represent the occurrence frequency of the keyword in the target document includes: when the keyword belongs to the vocabulary of the business field corresponding to the target contract category, determining the word frequency weight corresponding to the business field of the target contract category, and calculating the occurrence frequency of the keyword in the target document, and taking the product value of the calculated occurrence frequency and the word frequency weight as the first value used to represent the occurrence frequency of the keyword in the target document; and / or, when the keyword does not belong to the vocabulary of the business field corresponding to the target contract category, calculating the occurrence frequency of the keyword in the target document, and taking the calculated occurrence frequency as the first value used to represent the occurrence frequency of the keyword in the target document.

[0186] Optionally, in one implementation, the apparatus further includes an update module, configured to perform an update process on the word frequency weight and the similarity threshold in response to meeting an update condition; wherein the update process includes: obtaining each candidate combination; each candidate combination includes a candidate weight and a candidate threshold; determining, from each spare single sentence, the spare single sentences that meet the screening condition, to obtain at least one historical single sentence to be utilized; that any spare single sentence meets the screening condition includes: when the spare single sentence performs string matching with the corresponding historical parameter feature sentences, all the matches fail; the corresponding historical parameter feature sentence of any spare single sentence is: the parameter feature sentence corresponding to the parameter to which the parameter information extracted from the spare single sentence belongs; for each candidate combination, based on the candidate combination, determining the target sentence attribute of each historical single sentence; wherein the target sentence attribute includes a first attribute, a second attribute or a third attribute; for each candidate combination, based on the first quantity of the historical single sentences belonging to the first attribute, the second quantity of the historical single sentences belonging to the second attribute, and the third quantity of the historical single sentences belonging to the third attribute, calculating the harmonic mean of the candidate combination for characterizing the accuracy rate and recall rate of parameter information extraction; determining the candidate combination corresponding to the largest calculated harmonic mean as the target combination, and using the candidate weight and the candidate threshold in the target combination as the updated word frequency weight and the updated similarity threshold respectively.

[0187] Optionally, in one implementation, the determining manner of the target sentence attribute of each historical single sentence includes: using the candidate weight in the candidate combination as the word frequency weight utilized in the predetermined vectorization process, performing the predetermined vectorization process on the historical single sentence, to obtain the third vector feature of the historical single sentence; calculating the similarity between the third vector feature of the historical single sentence and the fourth vector features of the corresponding historical parameter feature sentences; determining whether there is a target similarity that meets the candidate threshold in the calculated similarities; if so, outputting predetermined parameter verification information, so that the user verifies whether the historical single sentence is similar to the historical parameter feature sentence corresponding to the target similarity; in response to the confirmation instruction sent by the user for characterizing similarity, determining that the target sentence attribute of the historical single sentence under the candidate combination is the first attribute; in response to the confirmation instruction sent by the user for characterizing dissimilarity, determining that the target sentence attribute of the historical single sentence under the candidate combination is the second attribute; if not, outputting predetermined parameter extraction failure information, so that the user confirms whether the candidate threshold in the candidate combination is too high; in response to the confirmation instruction sent by the user for characterizing that the threshold is too high, determining that the target sentence attribute of the historical single sentence under the candidate combination is the third attribute.

[0188] The embodiments of the present application further provide an electronic device, such asFigure 7 As shown in the figure, it includes a processor 701, a communication interface 702, a memory 703, and a communication bus 704. Among them, the processor 701, the communication interface 702, and the memory 703 complete mutual communication through the communication bus 704.

[0189] The memory 703 is used to store computer programs; the processor 701, when executing the programs stored on the memory 703, implements the parameter information extraction method described in any one of the above embodiments.

[0190] The communication bus mentioned in the above electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, only a thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0191] The communication interface is used for communication between the above electronic device and other devices.

[0192] The memory may include a Random Access Memory (RAM), or may also include a Non-Volatile Memory (NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located far from the aforementioned processor.

[0193] The above-mentioned processor may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0194] In another embodiment provided by the present application, a computer-readable storage medium is also provided. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the parameter information extraction method described in any one of the above are implemented.

[0195] In another embodiment provided by the present application, a computer program product including instructions is further provided. When it runs on a computer, it causes the computer to execute any one of the parameter information extraction methods in the above embodiments.

[0196] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wirelessly (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that the computer can access or a data storage device such as a server or data center that includes one or more integrated available media. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium (for example, a solid state disk (SSD)).

Claims

1. A method for extracting parameter information, characterized in that, Including: Obtain a target contract for which parameter information extraction is to be performed; Determine the target positioning information corresponding to the target contract category to which the target contract belongs; wherein, the target positioning information is used to indicate the target parameters for which parameter information extraction is to be performed in the contracts of the target contract category, and the text positions of the parameter information of the target parameters in any contract; Based on the target positioning information, in the target contract, locate the paragraphs containing the parameter information of the target parameters to obtain the target paragraphs corresponding to the target parameters; Obtain multiple parameter feature sentences corresponding to the target parameters; wherein, each parameter feature sentence corresponding to the target parameter is a candidate sentence for describing the parameter information of the target parameter in the contracts of the target contract category; among the multiple parameter feature sentences corresponding to the target parameter, different parameter feature sentences are different; Analyze whether there is a target single sentence in each single sentence of the target paragraph corresponding to the target parameter that matches any parameter feature sentence corresponding to the target parameter; In response to the existence of a matching target single sentence, based on the target single sentence, extract the parameter information of the target parameter in the target contract.

2. The method according to claim 1, wherein The analyzing whether there is a target single sentence in each single sentence of the target paragraph corresponding to the target parameter that matches any parameter feature sentence corresponding to the target parameter includes: Perform clause splitting on the target paragraph corresponding to the target parameter to obtain multiple single sentences of the target paragraph; Traverse the multiple single sentences, and when traversing to each current single sentence, perform the following process: Perform string matching between this single sentence and each parameter feature sentence among the multiple parameter feature sentences corresponding to the target parameter; If this single sentence matches any parameter feature sentence, determine this single sentence as the target single sentence and end the traversal; Otherwise, perform a predetermined vectorization process on this single sentence to obtain the first vector feature of this single sentence, and calculate the similarity between the first vector feature and the second vector feature of each parameter feature sentence corresponding to the target parameter; wherein, the second vector feature of each parameter feature sentence corresponding to the target parameter is obtained by performing the predetermined vectorization process on this parameter feature sentence; If the similarity between the first vector feature of this single sentence and the second vector feature of any parameter feature sentence is not lower than the similarity threshold, determine this single sentence as the target single sentence and end the traversal, otherwise, traverse to the next single sentence.

3. The method according to claim 1 or 2, characterized in that, When the target parameter is a parameter of the first category, the extracting the parameter information of the target parameter in the target contract based on the target single sentence includes: Analyze the target semantics represented by the target single sentence; If the target semantics is a negative semantics, use the first field value as the parameter information of the target parameter; otherwise, use the second field value as the parameter information of the target parameter; And / or When the target parameter is a parameter of the second category, the extracting the parameter information of the target parameter in the target contract based on the target single sentence includes: Extract the content of the target single sentence according to the extraction position set by the parameter feature sentence matched for the target single sentence, so as to obtain the parameter information of the target parameter in the target contract.

4. The method according to claim 2, wherein The predetermined vectorization processing includes: Extract at least one keyword from the sentence to be processed; wherein, the sentence to be processed is any parameter feature sentence or any single sentence; For each keyword, calculate a first value used to represent the occurrence frequency of the keyword in the target document, and calculate a second value used to represent the distribution of the keyword in the target document set. Based on the product value of the first value and the second value of the keyword, construct a vector feature corresponding to the keyword; wherein, when the sentence to be processed is any single sentence, the target document is the target contract, and when the sentence to be processed is any parameter feature sentence, the target document is a sample contract containing the any parameter feature sentence; the target document set includes other documents belonging to the business field corresponding to the target document except the target document. Use the vector matrix obtained by splicing the vector features corresponding to each keyword as the vector feature corresponding to the sentence to be processed; wherein, each matrix element in the vector matrix is used to represent the vector feature corresponding to a keyword.

5. The method according to claim 4, wherein The calculation of the first value used to represent the occurrence frequency of the keyword in the target document includes: When the keyword belongs to the vocabulary of the business field corresponding to the target contract category, determine the word frequency weight corresponding to the business field of the target contract category, and calculate the occurrence frequency of the keyword in the target document, and use the product value of the calculated occurrence frequency and the word frequency weight as the first value used to represent the occurrence frequency of the keyword in the target document; And / or When the keyword does not belong to the vocabulary of the business field corresponding to the target contract category, calculate the occurrence frequency of the keyword in the target document, and use the calculated occurrence frequency as the first value used to represent the occurrence frequency of the keyword in the target document.

6. The method according to claim 5, characterized in that, The method further includes: In response to meeting the update condition, perform an update process on the word frequency weight and the similarity threshold; wherein, the update process includes: Obtain each candidate combination; each candidate combination includes a candidate weight and a candidate threshold; Determine the spare single sentences that meet the screening conditions from each spare single sentence to obtain at least one historical single sentence to be utilized; that any spare single sentence meets the screening conditions includes: when the spare single sentence performs string matching with the corresponding historical parameter feature sentences, all matches fail; the corresponding historical parameter feature sentence of any spare single sentence is: the parameter feature sentence corresponding to the parameter to which the parameter information extracted from the spare single sentence belongs; For each candidate combination, based on the candidate combination, determine the target sentence attribute of each historical single sentence; wherein, the target sentence attribute includes the first attribute, the second attribute or the third attribute. For each candidate combination, based on the first quantity of historical single sentences belonging to the first attribute, the second quantity of historical single sentences belonging to the second attribute, and the third quantity of historical single sentences belonging to the third attribute, calculate the harmonic mean of this candidate combination for characterizing the accuracy rate and recall rate of parameter information extraction; Determine the candidate combination corresponding to the calculated harmonic mean with the largest value as the target combination, and use the candidate weight and candidate threshold in the target combination as the updated word frequency weight and the updated similarity threshold respectively.

7. The method according to claim 6, wherein The determination method of the target statement attribute of each historical single sentence includes: Use the candidate weight in this candidate combination as the word frequency weight used in the predetermined vectorization process, and perform the predetermined vectorization process on this historical single sentence to obtain the third vector feature of this historical single sentence; Calculate the similarity between the third vector feature of this historical single sentence and the fourth vector features of the corresponding historical parameter feature sentences; Determine whether there is a target similarity that meets the candidate threshold in the calculated similarities; If it exists, output the predetermined parameter verification information so that the user can verify whether the historical single sentence is similar to the historical parameter feature sentence corresponding to the target similarity; in response to the confirmation instruction sent by the user for characterizing similarity, determine that the target statement attribute of this historical single sentence under this candidate combination is the first attribute; in response to the confirmation instruction sent by the user for characterizing dissimilarity, determine that the target statement attribute of this historical single sentence under this candidate combination is the second attribute; If it does not exist, output the information indicating that the parameter extraction fails so that the user can confirm whether the candidate threshold in this candidate combination is too high; in response to the confirmation instruction sent by the user for characterizing that the threshold is too high, determine that the target statement attribute of this historical single sentence under this candidate combination is the third attribute.

8. A parameter information extraction device, characterized in that Includes: A contract acquisition module for acquiring a target contract for which parameter information extraction is to be performed; A location information determination module for determining the target location information corresponding to the target contract category to which the target contract belongs; wherein, the target location information is used to indicate the target parameters for which parameter information extraction is to be performed in the contracts of the target contract category, and the text positions of the parameter information of the target parameters in any contract; A paragraph determination module for locating, based on the target location information, the paragraph containing the parameter information of the target parameter in the target contract to obtain the target paragraph corresponding to the target parameter; A feature sentence acquisition module for acquiring multiple parameter feature sentences corresponding to the target parameter; wherein, each parameter feature sentence corresponding to the target parameter is a candidate sentence for describing the parameter information of the target parameter in the contracts of the target contract category; among the multiple parameter feature sentences corresponding to the target parameter, different parameter feature sentences are different; A single sentence analysis module for analyzing whether there is a target single sentence in each single sentence of the target paragraph corresponding to the target parameter that matches any parameter feature sentence corresponding to the target parameter; An information extraction module, configured to extract parameter information of the target parameter in the target contract based on the target single sentence in response to the existence of a matching target single sentence.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete mutual communication through the communication bus; The memory is used to store computer programs; The processor is configured to implement the method according to any one of claims 1-7 when executing the program stored on the memory.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the method according to any one of claims 1-7 is implemented.

Images