Judicial provision intelligent matching method and system based on natural language processing

Abstract

The invention relates to the technical field of data processing, in particular to an intelligent judicial provision matching method and system based on natural language processing, and the method comprises the steps: extracting effective words and sentences in any judgment text, calculating the relevance between each effective word and sentence and a criminal name according to the occurrence frequency of the effective words and sentences in the judgment text, and calculating the relevance between each effective word and sentence and the criminal name; and constructing a text library, wherein the text library comprises referee texts, effective words and sentences and criminal names. According to the method, the updating coefficient is calculated, the updating node is determined, the outdated judgment text is deleted in time, and the latest text library is constructed, so that the method can adapt to legal changes (such as amendment method issuing or new judgment occurrence), the problem of inaccurate crime name recommendation caused by legal basis change is avoided, the timeliness of a recommendation result is ensured, and the recommendation efficiency is improved. And obtaining criminal name recommendation indexes of all criminal names in the referee text according to the latest text library to adapt to a rapidly changing legal environment and realize accurate matching of fact contents in the referee file and the corresponding criminal names.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing. More specifically, the present invention relates to an intelligent matching method and system for judicial provisions based on natural language processing. Background Art

[0002] With the development of society and the continuous improvement of the legal system, the complexity and quantity of judicial provisions have increased significantly. The fields covered by the law are becoming increasingly extensive, involving multiple aspects such as civil, criminal, and administrative, and the content of the provisions is gradually refined and intertwined. At the same time, the continuous introduction of case laws, judicial interpretations, and amended laws has further enriched the connotation of the legal system. In this context, judicial staff face greater challenges in case handling and need to quickly and accurately match applicable legal bases from a vast number of provisions to support the legality and rationality of judgments. Traditional matching methods are no longer fully adaptable to such complexity and dynamic requirements. To improve judicial efficiency and ensure the quality of judgments, intelligent matching methods based on natural language processing technology have gradually attracted attention. Such methods can more efficiently assist judicial staff in completing the provision matching task by deeply analyzing text semantics and combining the professional characteristics of the legal field, promoting the modernization and intelligent development of judicial work.

[0003] As the legal provisions and case laws are updated, the relevance between words and charges changes, and existing methods lack the ability to dynamically adjust. With the promulgation of new laws, amendments, or judicial interpretations, the definitions and applicable conditions of charges will be continuously adjusted. For example, after a certain provision is updated, "malice" may be added as a new element of the "crime of fraud", but matching methods based on historical data or fixed dictionaries cannot reflect this change in a timely manner. This may cause the recommended provisions to be out of touch with current judicial practice, reducing the timeliness and reliability of the matching results. Summary of the Invention

[0004] The present invention provides an intelligent matching method and system for judicial provisions based on natural language processing, aiming to solve the problem that matching methods based on historical data or fixed dictionaries in related technologies cannot reflect this change in a timely manner. This may cause the recommended provisions to be out of touch with current judicial practice, reducing the timeliness and reliability of the matching results.

[0005] In a first aspect, the present invention provides an intelligent matching method for judicial provisions based on natural language processing, including: extracting effective words and sentences from any judgment text, calculating the relevance between each effective word and sentence and the crime name according to the frequency of occurrence of the effective word and sentence in the judgment text, and constructing a text library, where the text library includes judgment texts, effective words and sentences, and crime names; sorting in descending order according to the level of relevance between each effective word and sentence and the crime name to obtain a word and sentence sequence of the crime name, dichotomizing the relevance between all effective words and sentences and the crime name in the word and sentence sequence of the crime name, and taking the category with a larger average relevance as the effective word and sentence sequence of the crime name; calculating an update coefficient of the effective word and sentence sequence of the crime name, where the update coefficient is positively correlated with the difference value between the effective word and sentence sequence of the current node crime name and the effective word and sentence sequences of the crime names of each historical node; if the update coefficient is greater than an update threshold, determining an update node, deleting the judgment text corresponding to the crime name before the update node in the text library, and constructing a latest text library; calculating a recommendation index for each crime name corresponding to the judgment text according to the latest text library to complete the intelligent matching of the judgment text and the crime name, where the recommendation index is positively correlated with the relevance between all effective words and sentences in the judgment text and the crime name. Through this method, the relevant degree between words and sentences and the crime name can be accurately captured, and the interference of irrelevant words and sentences can be avoided, thereby significantly improving the accuracy of crime name recommendation. In addition, through a dynamic update mechanism, an update coefficient is calculated to monitor the change of the effective word and sentence sequence, and when the update coefficient exceeds the threshold, the outdated judgment text is deleted to ensure that the recommendation model is always based on the latest legal basis and sentencing criteria, further reducing the possibility of misjudgment.

[0006] Further, calculating the recommendation index for each crime name corresponding to the judgment text includes: the recommendation index is also positively correlated with the probability that the judgment text is the crime name when all effective words and sentences of the crime name appear. Through this method, the actual application frequency of the crime name in similar situations is reflected. The higher the probability, the more frequently the crime name is adopted, and the higher the recommendation index of the crime name.

[0007] Further, obtaining the effective word and sentence sequence of the crime name includes: dichotomizing the relevance between all effective words and sentences and the crime name in the word and sentence sequence of the crime name, taking the group with a larger average relevance as the effective correlation words, and taking the group with a smaller average relevance as the ineffective correlation words; deleting the ineffective correlation words, retaining the effective correlation words, and constructing the word and sentence sequence corresponding to the effective correlation words as the effective word and sentence sequence of the crime name. By dichotomizing the word and sentence sequence and distinguishing effective correlation words and ineffective correlation words based on the average value, this method can screen out the words and sentences with stronger relevance to the crime name.

[0008] Further, calculating the update coefficient of the effective phrase sequence of the crime name includes: determining the weight coefficient corresponding to each historical node, and weighting the difference value between the effective phrase sequence of the crime name at the current node and the effective phrase sequence of the crime name at each historical node by using the weight coefficient corresponding to each historical node, so as to obtain the update coefficient of the effective phrase sequence of the crime name. The calculation of the update coefficient synthesizes the differences between the current node and multiple historical nodes, and performs weighted adjustment through the weight coefficient, so that the update coefficient can more accurately reflect the change degree of the effective phrase sequence of the crime name.

[0009] Further, the determination method of each historical node includes: setting the current day as the current node, and taking each month before the current node as a historical node to obtain multiple historical nodes.

[0010] Further, determining the update node, where the update node is the Nth day before the current node.

[0011] Further, determining the weight coefficient corresponding to each historical node includes: calculating the weight coefficient corresponding to each historical node based on the time distance between the current node and each historical node, where the weight coefficient is inversely correlated with the time distance of the historical node. The inverse correlation between the weight coefficient and the time distance means that the weight of the farther historical node is smaller, thereby reducing the interference of outdated data on the overall calculation.

[0012] Further, binary-classifying the phrase sequence of the crime name includes: using the Otsu threshold method to perform binary classification on the relevance between all effective phrases in the phrase sequence of the crime name and the crime name.

[0013] Further, the empirical value of the update threshold is 0.7. In the second aspect of the present invention, there is also provided an intelligent matching system for judicial provisions based on natural language processing, including a processor and a memory, where the memory stores a computer program, and the processor executes the computer program to implement the intelligent matching method for judicial provisions based on natural language processing described in any one of the above.

[0014] Beneficial effects: By calculating the update coefficient and determining the update node, outdated judgment texts are deleted in a timely manner and the latest text library is constructed, which can adapt to legal changes (such as the promulgation of amended laws or the emergence of new precedents), avoid the problem of inaccurate crime name recommendations caused by changes in legal bases, thereby ensuring the timeliness of the recommendation results, and obtaining the crime name recommendation index of all crime names in the judgment text according to the latest text library, adapting to the rapidly changing legal environment, and realizing the accurate matching of the factual content in the judgment document with the corresponding crime names. Description of the Drawings

[0015] Figure 1 It is a flowchart schematically showing the calculation of the crime name recommendation index according to an embodiment of the present invention. Detailed implementation manners

[0016] The following will describe in detail the specific implementation manners of the present invention with reference to the accompanying drawings.

[0017] S101: Obtain the judgment text.

[0018] In one embodiment, the judgment text can be obtained from an existing judgment text library, and the text of the factual content and the corresponding charges in the judgment text are collected. And the text of the factual content and the corresponding charges in the judgment text will change as the judgment text library is updated.

[0019] S102: Extract the effective words and sentences from the judgment text.

[0020] In one embodiment, for any judgment text, the effective words and sentences in the judgment text need to be extracted. Specifically, for each sentence in the judgment text, auxiliary words, conjunctions, and prepositions need to be deleted. When deleting words and sentences, the stop word list can be referred to or word segmentation can be performed using a special dictionary in the judicial field (THULAC + custom legal dictionary), and the effective words and sentences of all judgment texts are extracted to build a text library, where the text library contains judgment texts, effective words and sentences, and charges.

[0021] Exemplarily, the charge of the judgment text is the crime of violent injury, and the judgment text shows that "the defendant Zhang San intentionally used violence to injure another person's body in 2023, causing minor injuries to the victim". After removing auxiliary words, conjunctions, and prepositions, the effective words and sentences obtained are: "intentionally, violence, injury, body, minor injury".

[0022] S103: Calculate the relevance between each effective word and sentence and the charge.

[0023] In one embodiment, since a judgment text may correspond to multiple charges, and using a single effective word and sentence in the judgment text cannot accurately determine a single charge, multiple effective words and sentences are needed to determine a charge. Therefore, it is necessary to calculate the relevance between each effective word and sentence and the charge.

[0024] In one embodiment, to calculate the relevance between each effective word and sentence and the charge, the calculation formula is: , where is the relevance between the th effective word and sentence in the judgment text and the charge , is the probability that the charge in the th judgment text corresponds to the th effective word and sentence, is the number of judgment texts. is the The overall frequency of occurrence of a valid phrase in all judgment texts, representing the generalization characteristics of the th valid phrase. The larger this value is, the stronger the independent connection and the greater the relevance of the th valid phrase to this crime. It should be noted that to obtain the probability of the th crime corresponding to the th valid phrase occurring , for example, if the th valid phrases are "violence" and "finance", in a judgment text, for the extracted valid phrases, there can be multiple and some valid phrases may appear repeatedly. For example, for a valid phrase of "violence" and "finance", after extracting valid phrases from each paragraph in the judgment text, there may be multiple repeated valid phrases of "violence" and "finance". At this time, the probability of the valid phrases of "violence" and "finance" in the judgment text appearing among all valid phrases is statistically calculated. The greater the probability, the stronger the independent connection of the valid phrase to this crime.

[0025] In another embodiment, another calculation formula for the relevance between each valid phrase and the crime is provided. This calculation formula not only considers the overall frequency of occurrence of the valid phrase in a single judgment text but also considers the degree of concentration of the probability of the valid phrase and the crime appearing in all judgment texts. The greater the degree of concentration, the stronger the relevance. Therefore, the accuracy of calculating the relevance between the valid phrase and the crime can be further improved. The calculation formula is: . In the formula, is the relevance between the th valid phrase in the judgment text and the crime , is the probability of the th judgment text corresponding to the th crime and the th valid phrase appearing, is the number of judgment texts, represents the number of valid phrases in the judgment text, represents the median function. is the degree of concentration of the probability of the th valid phrase and the crime in the th text. The larger this value is, the stronger the relevance of the th valid phrase to the crime .

[0026] S104: Obtain the sequence of valid phrases for any crime.

[0027] In one embodiment, one charge corresponds to multiple valid phrases, and the relevance between each of the multiple valid phrases and the charge is obtained. Then, a descending order is performed according to the level of relevance between each valid phrase and the charge to obtain a phrase sequence of the charge. Binary classification is performed based on the relevance between all valid phrases in the phrase sequence of the charge and the charge. The group with a larger average relevance is used as the valid associated words, and the group with a smaller average relevance is used as the invalid associated words. Then, the invalid associated words are deleted, and the valid associated words are retained. Finally, the phrase sequence corresponding to the valid associated words is constructed into the valid phrase sequence of the charge. Among them, the Otsu threshold method is used for binary classification of the phrase sequence of the charge. Thus, for any charge, the valid phrase sequence of the charge can be obtained.

[0028] S105: Determine whether to update the valid phrase sequence of the charge.

[0029] In one embodiment, as time goes by, the sentencing basis of the law may change. For example, the emergence of new supreme court judgments and the promulgation of amended laws will lead to changes in the existing sentencing basis, making the phrase recommendation based on previous judgment texts inaccurate and greatly reducing timeliness. Therefore, it is necessary to update the valid phrase sequence of the charge.

[0030] Specifically, an update coefficient is constructed to delete the previous judgment texts. That is, when there are large changes in the valid phrase sequence corresponding to the charge in a short period of time, it indicates that there are significant changes between the existing judgment texts and the previous judgment texts. At this time, the original judgment texts of the charge should be deleted and the new texts should be retained to obtain a new valid phrase sequence.

[0031] In one embodiment, the update coefficient of the valid phrase sequence of the charge is calculated. The update coefficient is positively correlated with the difference value between the valid phrase sequence of the current node charge and the valid phrase sequences of the charges at each historical node. The determination method of each historical node includes: setting the current day as the current node, taking each month before the current node as a historical node to obtain multiple historical nodes, and the number of historical nodes can be set artificially. For example, the number of historical nodes is 4. Then, the update coefficient is calculated, and the calculation formula is: , where is the update coefficient, is the relevance between all valid phrases in the valid phrase sequence of the current node charge and the charge, represents the relevance between all valid phrases in the valid phrase sequence of the charge at the m-th historical node and the charge, represents the number of historical nodes, represents the hyperbolic tangent function, represents the standard deviation function.

[0032] In another embodiment, another method is provided to calculate the update coefficient of the effective phrase sequence of the crime name. Specifically, when calculating the change differences between the effective phrase sequences of the crime names of each historical node and the effective phrase sequence of the current node crime name respectively, since the time distances between each historical node and the current node are different, weight coefficients corresponding to each historical node should be assigned according to the time distance between each historical node and the current node. The farther the historical node is from the current node, the smaller the weight parameter of that historical node, and the sum of the weight coefficients of each historical node is 1. The calculation formula is: , where is the update coefficient, is the relevance between all effective phrases in the effective phrase sequence of the current node crime name and the crime name, represents the relevance between all effective phrases in the effective phrase sequence of the crime name of the m-th historical node and the crime name, represents the number of historical nodes, represents the weight parameter of the m-th historical node, represents the hyperbolic tangent function. It should be noted that determining the weight parameter of the historical node based on the distance between the historical node and the current node reflects the characteristics that recent data is more timely and relevant in judicial practice. For example, the judgment texts in the recent month may better reflect the current legal trends and sentencing bases and require a larger weight, while the data from a year ago may lose some reference value due to legal amendments or changes in case laws and require a smaller weight. In one embodiment, the number of historical nodes is 4, and the weights of the 4 nodes from the current node in order of proximity are 0.4, 0.3, 0.2, and 0.1 respectively. This way of weight assignment ensures that the system pays more attention to the latest data, thus improving the timeliness and practicality of the analysis results.

[0033] According to the above steps, because the legal environment is dynamic, for example, the introduction of new case laws, amended laws, or judicial interpretations may lead to changes in the phrases related to the crime name. In the embodiment, through the weighted calculation of the difference values, these changes can be detected sensitively. For example, if the effective phrases of "intentional injury crime" change from "injury" and "body" to "violence" and "minor injury", the difference value will increase significantly, and the update coefficient will increase accordingly, prompting the system to update. This adaptability enables the system to cope with the evolution of legal practice and maintain long-term effectiveness.

[0034] It should be noted that according to the above two embodiments, the update coefficient of the effective phrase sequence of the crime name can be calculated. Select any one of the embodiments to calculate the update coefficient of the effective phrase sequence of the crime name. When the update coefficient is greater than the update threshold, it is determined that the effective phrase sequence of the crime name needs to be updated. Among them, the empirical value of the update threshold is 0.6. In other embodiments, the empirical value of the update threshold can be 0.7 or 0.62, etc., and it can be adjusted according to the specific implementation situation.

[0035] S106: Determine the update node and construct the latest text library.

[0036] In one embodiment, the update node can be set artificially. For example, when the update coefficient of the effective sentence sequence of the current node's crime name is greater than the update threshold, it is determined that the effective sentence sequence of the crime name needs to be updated. The Nth day before the current node is used as the update node, and the judgment texts corresponding to this crime name before the update node in the text library are deleted, and the latest text library is constructed. Among them, the Nth day can be the day when the amended law is promulgated.

[0037] In one embodiment, since after the promulgation of the amended law, each region may not fully implement the amended law after promulgation, and the implementation standards of each region are also different. Therefore, it is necessary to obtain the update node more accurately. Specifically, calculate the update degree of each day. Taking the Bth day as an example, first calculate the standard deviation of the difference between the relevance of all effective sentences in the effective sentence sequence of the crime name on the Bth day in the text library and the relevance of all effective sentences in the effective sentence sequence of the crime name on the (B - 1)th day. The value obtained by normalizing the standard deviation is used as the update degree of the Bth day. Then, traverse the update degree of each day before the current node in chronological order. If there is a day when the update degree is greater than the update threshold, that day is used as the update node, and the judgment texts corresponding to this crime name before the update node in the text library are deleted, and the remaining judgment texts are constructed into the latest text library.

[0038] By adopting the above steps, by calculating the update coefficient and determining the update node, the outdated judgment texts are timely deleted and the latest text library is constructed, which can adapt to legal changes (such as the promulgation of the amended law or the emergence of new precedents), avoid the problem of inaccurate crime name recommendation caused by the change of legal basis, and thus ensure the timeliness of the recommendation result.

[0039] S107: Calculate the recommendation index of each crime name corresponding to the judgment text according to the latest text library.

[0040] In one embodiment, the recommendation index of each crime name is positively correlated with the relevance of all effective sentences in the judgment text to this crime name. Taking any crime name in any judgment text as an example, calculate the recommendation index of this crime name. The calculation formula is: , is the relevance of the th effective sentence in the judgment text to the crime name , is the recommendation index of the crime name, is the number of effective sentences in the judgment text.

[0041] In another embodiment, a calculation method for calculating the recommendation index of a crime name is also provided. The calculation formula is: , For the relevance between the th valid sentence and the crime name in the judgment text, the recommended index of the crime name, the number of valid sentences in the judgment text, and the probability that other judgment texts are also of the same crime name when all valid sentences of this crime name appear.

[0042] So far, the crime name recommendation coefficient for any crime name corresponding to each judgment text can be obtained. By traversing all the crime names in the judgment text, the crime name recommendation coefficients for all the crime names corresponding to the judgment text can be obtained. Thus, by inputting a text, the crime name recommendation coefficient of this text can be obtained, and the crime name with the highest crime name recommendation coefficient is used as the reference crime name for the staff for this text to assist the staff in judicial article matching.

[0043] The present invention also provides a judicial article intelligent matching system based on natural language processing. The system includes a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, it implements a judicial article intelligent matching method according to the first aspect of the present invention.

[0044] The system also includes other components well-known to those skilled in the art such as a communication bus and a communication interface, and their settings and functions are known in the art, so they will not be described in detail here.

[0045] In the present invention, the foregoing memory may be any tangible medium that contains or stores a program, which can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium may be any suitable magnetic storage medium or magneto-optical storage medium, such as a resistive random access memory (RRAM), a dynamic random access memory (DRAM), a static random access memory (SRAM), an enhanced dynamic random access memory (EDRAM), a high-bandwidth memory (HBM), a hybrid memory cube (HMC), etc., or any other medium that can be used to store the required information and can be accessed by an application program, a module, or both. Any such computer storage medium may be part of the device or accessible or connectable to the device. Any application or module described in the present invention may be implemented using computer-readable / executable instructions stored or otherwise maintained on such a computer-readable medium.

[0046] The above-described embodiments merely represent several implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the scope of the patent application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention.

Claims

1. A judicial clause intelligent matching method based on natural language processing, characterized in that: include: Extracting valid words and sentences from any judgment text, calculating the correlation between each valid word and sentence and the crime according to the frequency of occurrence of the valid word and sentence in the judgment text, and constructing a text library, wherein the text library includes the judgment text, valid words and sentences, and the crime; According to the correlation between each valid phrase and the crime, the phrases are sorted in descending order to obtain a phrase sequence of the crime, and the correlation between all valid phrases and the crime in the phrase sequence of the crime is classified into two categories, and the category with the larger correlation mean is taken as the valid phrase sequence of the crime; Calculate the update coefficient of the effective word sequence of the crime, where the update coefficient is positively correlated with the difference between the effective word sequence of the crime at the current node and the effective word sequence of the crime at each historical node; if the update coefficient is greater than the update threshold, determine the update node, delete the judgment text corresponding to the crime before the update node in the text library, and construct the latest text library; The recommendation index of each crime corresponding to the judgment text is calculated based on the latest text library to complete the intelligent matching of the judgment text and the crime. The recommendation index is positively correlated with the relevance of all valid words and sentences in the judgment text to the crime.

2. The intelligent matching method of judicial provisions based on natural language processing according to claim 1 is characterized in that: Calculate the recommended index for each crime corresponding to the judgment text, including: The recommendation index is also positively correlated with the probability that the judgment text contains the crime when all valid words and phrases of the crime appear.

3. The intelligent matching method of judicial provisions based on natural language processing according to claim 1 is characterized in that: Valid word sequences for obtaining charges include: The correlation between all valid words and sentences in the word sequence of the crime and the crime is classified into two categories, and the group with a larger correlation mean is regarded as a valid correlation word, and the group with a smaller correlation mean is regarded as an invalid correlation word; The invalid associated words are deleted, the valid associated words are retained, and the word sequence corresponding to the valid associated words is constructed as a valid word sequence for the crime.

4. The intelligent matching method of judicial provisions based on natural language processing according to claim 1 is characterized in that: Calculate the update coefficient of the effective word sequence of the crime, including: Determine the weight coefficient corresponding to each historical node, use the corresponding weight coefficient of each historical node to weight the difference between the effective word sequence of the crime at the current node and the effective word sequence of the crime at each historical node, and obtain the update coefficient of the effective word sequence of the crime.

5. The method for intelligent matching of judicial provisions based on natural language processing according to claim 2 is characterized in that: The method for determining each historical node includes: Set today as the current node, and take each month before the current node as a historical node, to obtain multiple historical nodes.

6. The intelligent matching method of judicial provisions based on natural language processing according to claim 1 is characterized in that: An update node is determined, wherein the update node is N days before the current node.

7. The intelligent matching method of judicial provisions based on natural language processing according to claim 1 is characterized in that: Determine the weight coefficient corresponding to each historical node, including: Based on the time distance between the current node and each historical node, a weight coefficient corresponding to each historical node is calculated, wherein the weight coefficient is inversely correlated with the time distance between the historical nodes.

8. The intelligent matching method of judicial provisions based on natural language processing according to claim 1 is characterized in that: The word sequence of the crime is classified into two categories, including: The Otsu threshold method is used to perform binary classification on the correlation between all valid words and sentences in the word sequence of the crime and the crime.

9. The intelligent matching method of judicial provisions based on natural language processing according to claim 1 is characterized in that: The empirical value of the update threshold is 0.

7.

10. A judicial document intelligent matching system based on natural language processing, comprising a processor and a memory, characterized in that: The memory stores a computer program, and the processor executes the computer program to implement the intelligent matching method of judicial provisions based on natural language processing as described in any one of claims 1-9.

Images