Method and device for matching case facts with legal provisions

Abstract

The present application provides a method and device for matching case facts with legal provisions. The method includes: segmenting the case fact description text to generate a first phrase set; in the first phrase set, screening out a second phrase set with preset attributes; establishing a mapping of the second phrase set in a legal text database to obtain a number of legal provisions that have a mapping relationship with the second phrase set; and determining the matching degree between the case fact description text and the several legal provisions through a pre-trained Bert‑wwm coding matching model. In this way, encoding can be performed in combination with text semantics, so that when matching the case description text, matching of regulations of different granularities can be performed in combination with context semantics, thereby increasing the matching accuracy of the target regulations.

Description

Technical Field

[0001] The present application relates to the field of text recommendation technology, and in particular to a method and device for matching case facts with legal provisions. Background Art

[0002] With the continuous development of natural language technology, especially the development of pre-trained language models in recent years, the application of natural language processing technology in the legal field has made great progress. For example, it is used for case classification and regulation matching. These mainly determine the case or regulation corresponding to the pending case based on the keywords in the pending case and the regulations.

[0003] In the process of implementing the prior art, the inventor discovered that:

[0004] Matching the corresponding regulations based on the case description text and individual keywords in the regulations is not conducive to determining the semantic connection between the relevant keywords in the case description text, which may cause the current pending case to be matched with regulations of other categories, increasing the possibility of mismatching between the case description text and the corresponding regulations.

[0005] Therefore, it is necessary to provide a method and device for accurately matching case facts and legal provisions according to case description text. Summary of the invention

[0006] The embodiments of the present application provide a method and device for matching case facts with legal provisions, so as to solve the technical problem of low matching accuracy between case description text and regulations.

[0007] Specifically, a method for matching case facts with legal provisions includes the following steps:

[0008] Segmenting the case fact description text to generate a first phrase set;

[0009] Filtering out a second set of phrases having a preset attribute from the first set of phrases;

[0010] Establishing a mapping of the second phrase set in a legal text database to obtain a number of legal provisions having a mapping relationship with the second phrase set;

[0011] The matching degree between the case fact description text and the several legal provisions is determined through the pre-trained Bert-wwm coding matching model.

[0012] Furthermore, the word segmentation of the case fact description text to generate the first phrase set specifically includes:

[0013] The case fact description text is segmented using the Unigram word segmentation model to generate a unigram set consisting of several unigrams.

[0014] Furthermore, the method further comprises:

[0015] The standard regulatory text is segmented by the Unigram segmentation model to generate a unigram regulatory text database consisting of several unigrams.

[0016] Segment the standard regulatory texts using the Bigram segmentation model to generate a bigram regulatory text database consisting of several bigrams.

[0017] The standard regulatory text is segmented by the Trigram segmentation model to generate a trigram regulatory text database consisting of several trigrams.

[0018] The unigram legal text database, the bigram legal text database, and the trigram legal text database are merged to generate a legal text database.

[0019] Furthermore, the matching degree between the case fact description text and the several legal provisions is determined by using the pre-trained Bert-wwm coding matching model, specifically including:

[0020] Determine, according to the plurality of legal provisions and through a pre-trained Bert-wwm coding matching model, a plurality of first coding vectors of the plurality of legal provisions;

[0021] Determine a second encoding vector of the case fact description text through a pre-trained Bert-wwm encoding matching model according to the case fact description text;

[0022] Using a preset vector similarity calculation rule, calculate the cosine value of the angle between the first encoding vector and the second encoding vector;

[0023] According to the calculation result of the cosine value, the matching degree between the case fact description text and the several legal provisions is determined.

[0024] Furthermore, the matching degree between the case fact description text and the several legal provisions is determined by using the pre-trained Bert-wwm coding matching model, specifically including:

[0025] Determine, according to the plurality of legal provisions and through a pre-trained Bert-wwm coding matching model, a plurality of first coding vectors of the plurality of legal provisions;

[0026] Determine a second encoding vector of the case fact description text through a pre-trained Bert-wwm encoding matching model according to the case fact description text;

[0027] Obtain a recombined third vector according to the first encoding vector and the second encoding vector and using a vector recombination rule;

[0028] According to the third vector and through the pre-trained neural network fully connected algorithm, the matching degree between the case fact description text and the several legal provisions is determined.

[0029] Furthermore, the training process of the pre-trained Bert-wwm coding matching model includes:

[0030] Several standard regulatory texts and several training case fact description texts are used to perform negative feedback optimization of the Bert-wwm coding matching model.

[0031] The embodiment of the present application also provides a device for matching case facts with legal provisions.

[0032] Specifically, a device for matching case facts with legal provisions includes:

[0033] A word segmentation module, used to segment the case fact description text to generate a first phrase set;

[0034] A screening module, used to screen out a second set of phrases having a preset attribute from the first set of phrases;

[0035] A mapping module, used to establish a mapping of the second phrase set in a legal text database, and obtain a number of legal provisions having a mapping relationship with the second phrase set;

[0036] The matching degree calculation module is used to determine the matching degree between the case fact description text and the several legal provisions through a pre-trained Bert-wwm coding matching model.

[0037] Furthermore, the word segmentation module is used to segment the case fact description text to generate a first phrase set, specifically for:

[0038] The case fact description text is segmented using the Unigram word segmentation model to generate a unigram set consisting of several unigrams.

[0039] Furthermore, the matching degree calculation module is used to determine the matching degree between the case fact description text and the several legal provisions through the pre-trained Bert-wwm coding matching model, specifically for:

[0040] Determine, according to the plurality of legal provisions and through a pre-trained Bert-wwm coding matching model, a plurality of first coding vectors of the plurality of legal provisions;

[0041] Determine a second encoding vector of the case fact description text according to the case fact description text and through a pre-trained Bert-wwm encoding matching model;

[0042] Using a preset vector similarity calculation rule, calculate the cosine value of the angle between the first encoding vector and the second encoding vector;

[0043] According to the calculation result of the cosine value, the matching degree between the case fact description text and the several legal provisions is determined.

[0044] Furthermore, the matching degree calculation module is used to determine the matching degree between the case fact description text and the several legal provisions through the pre-trained Bert-wwm coding matching model, specifically for:

[0045] Determine, according to the plurality of legal provisions and through a pre-trained Bert-wwm coding matching model, a plurality of first coding vectors of the plurality of legal provisions;

[0046] Determine a second encoding vector of the case fact description text through a pre-trained Bert-wwm encoding matching model according to the case fact description text;

[0047] Obtain a recombined third vector according to the first encoding vector and the second encoding vector and using a vector recombination rule;

[0048] According to the third vector and through the pre-trained neural network fully connected algorithm, the matching degree between the case fact description text and the several legal provisions is determined.

[0049] The technical solution provided by the application embodiment has at least the following beneficial effects:

[0050] By segmenting the words of the regulations through a multi-word segmentation model, we obtain phrases with different granularities. In addition, we can encode them in combination with text semantics, so that when matching case description texts, we can match regulations of different granularities in combination with context semantics, thereby increasing the matching accuracy of the target regulations. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0052] Figure 1 A flowchart of a method for matching case facts with legal provisions provided in an embodiment of the present application.

[0053] Figure 2A schematic diagram of the structure of a device for matching case facts with legal provisions provided in an embodiment of the present application.

[0054] 100 Device for matching case facts with legal provisions

[0055] 11 Word segmentation module

[0056] 12 Filter Module

[0057] 13 Mapping Module

[0058] 14 Matching degree calculation module DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. 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 without making creative work are within the scope of protection of the present application.

[0060] The present application embodiment provides a method for matching case facts with legal provisions, which is mainly used for matching case fact description text with legal text in the legal field. That is, according to the case description information, the legal provisions with a certain degree of relevance to the current case are matched, so as to facilitate the subsequent work of relevant personnel. Please refer to Figure 1 , a method for matching case facts with legal provisions provided in an embodiment of the present application, comprising the following steps:

[0061] S100: Segmenting the case fact description text to generate a first phrase set;

[0062] S200: Filtering out a second set of phrases having a preset attribute from the first set of phrases;

[0063] S300: Establishing a mapping of the second phrase set in a legal text database to obtain a number of legal provisions having a mapping relationship with the second phrase set;

[0064] S400: Determine the matching degree between the case fact description text and the plurality of legal provisions through a pre-trained Bert-wwm coding matching model.

[0065] S100: Segment the case fact description text to generate a first phrase set.

[0066] The case fact description text here mainly refers to the important fact description text used in the description of legal cases. In practical applications, it can be understood as a legal document, for example, it can be expressed as at least one of the following legal documents: case acceptance approval form, conversation transcript, power of attorney, indictment, appeal, counterclaim, litigation preservation, evidence preservation, court ruling, judgment, mediation, and case closing report. The case fact description text is segmented, that is, the acquired case description text is segmented. After the case fact description text is segmented, a phrase set consisting of phrases in several case fact description texts is obtained. That is, a first phrase set is generated. The first phrase set here corresponds to a set consisting of several case fact description text phrases.

[0067] In a specific implementation provided in the present application, the text describing the facts of the case is segmented using an Ngram segmentation model. The Ngram segmentation model can perform a sliding window operation of size N on the content of the text according to the words, thereby forming a sequence of word fragments of length N. And during the segmentation process, the appearance of the Nth word is related to the previous N-1 words. That is, it is possible to fully combine the semantics of the above text to perform segmentation and obtain several phrase fragments of a preset length. In actual applications, the length of the word fragments obtained by segmentation depends on the preset value of N, which can be set according to actual conditions. For example, the case fact description text can be segmented by a Unigram segmentation model (i.e., N=1), and a first set of words consisting of several unigrams (i.e., a unigram set) can be obtained; or, the case fact description text can be segmented by a Bigram segmentation model (i.e., N=2), and a first set of words consisting of several bigrams (i.e., a bigram set) can be obtained; or, the case fact description text can be segmented by a Trigram segmentation model (i.e., N=3), and a first set of words consisting of several ternary words (i.e., a ternary set) can be obtained. The specific value of N here obviously does not constitute a limitation on the protection scope of this application.

[0068] Furthermore, in a preferred embodiment provided in the present application, the word segmentation of the case fact description text to generate a first phrase set specifically includes: word segmentation of the case fact description text through a Unigram word segmentation model to generate a unigram phrase set consisting of a plurality of unigram phrases.

[0069] Here, the Unigram word segmentation model is preferred to perform word segmentation on the case fact description text, that is, N=1 in the Ngram word segmentation model. At this time, each phrase fragment obtained by word segmentation consists of one word. For example, the description of "people under the age of eighteen" in the case fact description text is segmented by the Unigram word segmentation model to obtain a unigram phrase set consisting of unigram phrases such as "under", "eighteen", "years old", "of", and "people". Compared with the original case fact description text, the unigram phrase set is more detailed and has a smaller granularity, which can effectively increase the accuracy of subsequent corresponding regulations matching.

[0070] S200: Filtering out a second set of phrases having preset attributes from the first set of phrases.

[0071] The preset attributes here can be understood as preset screening rules. It is understandable that the case fact description text can obtain a first phrase set consisting of several word segmentation fragments after word segmentation processing. However, punctuation marks or stop words in the case fact description text will also exist as relevant elements in the first phrase set. Here, punctuation marks or stop words in the case fact description text are regarded as irrelevant elements in the first phrase set. These irrelevant elements do not increase the accuracy of the case description text matching with the regulatory terms, but will increase the amount of calculation in the process of matching the case description text with the regulatory terms. Therefore, the present application deletes these irrelevant elements so as to reduce the amount of calculation in the matching process without affecting the accuracy of the case description text matching with the regulatory terms, thereby improving the matching efficiency of the case description text and the regulatory terms. For example, when the case fact description text is segmented by the Unigram word segmentation model (i.e., N=1) to obtain a first phrase set (i.e., a unigram phrase set) consisting of several unigram phrases, the corresponding second phrase set is the unigram phrase set after screening. Alternatively, when the case facts description text is segmented by a Bigram segmentation model (i.e., N=2) to obtain a first phrase set (i.e., a bigram phrase set) consisting of a number of bigram phrases, the corresponding second phrase set obtained is a screened bigram phrase set. Alternatively, when the case facts description text is segmented by a Trigram segmentation model (i.e., N=3) to obtain a first phrase set (i.e., a triram phrase set) consisting of a number of triram phrases, the corresponding second phrase set obtained is a screened triram phrase set. The specific value of N here obviously does not constitute a limitation on the scope of protection of this application.

[0072] In a specific implementation provided by the present application, in the process of screening the second phrase set with preset attributes, the related phrases in the first phrase set can also be normalized. For example, the entity information such as license plate, ID card number, amount, organization name, etc. in the first phrase set is identified. That is, the phrases in the first phrase set are identified with corpus, and stop words are deleted, while other phrases remain unchanged.

[0073] S300: Establishing a mapping of the second phrase set in a legal text database, and obtaining a number of legal provisions having a mapping relationship with the second phrase set.

[0074] The legal text database here is a database obtained through preprocessing and recording relevant legal provisions. It can be understood that, compared with the first phrase set, the second phrase set has deleted irrelevant elements such as punctuation marks or stop words, reducing the number of elements in the set, thereby effectively increasing the matching efficiency between the second phrase set and the legal text database. The matching of the second phrase set with the legal text database is to establish a mapping in the legal database according to each element in the second phrase set, so as to obtain several legal provisions that have a mapping relationship with the case description. It can also be understood that candidate regulations with a certain degree of correlation with the case description are obtained.

[0075] It is worth noting that the regulatory text database here is also obtained by segmenting several regulatory clauses through the Ngram segmentation model. Here, the regulatory clauses before segmentation are regarded as standard regulatory texts. Standard regulatory texts are standard texts composed of coherent language formulated by relevant organizations; each standard text is composed of several word fragments. In this way, in the mapping process between the second phrase set and the regulatory text database, the mapping of the elements in the second phrase set and the elements in the regulatory text database in terms of expression form and semantics can be guaranteed. It can be understood that two word fragments may have the same semantics when their expression forms are different; or the expression forms are the same but the semantics are different. Therefore, segmenting the standard regulatory text can effectively avoid the phenomenon that the two elements have the same meaning but different expression forms, or the same expression forms but different meanings during the mapping process, thereby increasing the accuracy of the mapping between the second phrase set and the regulatory text database.

[0076] In practical applications, the legal text database may be in the form of an inverted index table, which records each word segment and the position of the word segment in the legal text database.

[0077] It is understandable that the specific value of N in the Ngram word segmentation model can be preset according to actual conditions. The larger the value of N, the longer the word fragment obtained; the smaller the value of N, the smaller the word fragment obtained. Moreover, the appearance of the Nth word fragment is related to the previous N-1 word fragments. Then the longer each word fragment is, the more accurate its contextual semantics is. Correspondingly, the shorter each word fragment is, the more possibilities its contextual semantics has. The present application uses Ngram word segmentation models with different preset N values ​​to segment standard regulatory texts respectively. In this way, word fragments with different granularities can be used, so that the second phrase set can be matched in more dimensions, thereby obtaining a standard text with a more accurate mapping.

[0078] Furthermore, in a preferred embodiment provided in the present application, the method for matching case facts with legal provisions also includes: segmenting the standard legal text through a Unigram segmentation model to generate a unigram legal text database consisting of a number of unigrams; segmenting the standard legal text through a Bigram segmentation model to generate a two-gram legal text database consisting of a number of bigrams; segmenting the standard legal text through a Trigram segmentation model to generate a three-gram legal text database consisting of a number of three-grams; merging the one-gram legal text database, the two-gram legal text database, and the three-gram legal text database to generate a legal text database.

[0079] Here, N=1 in the Unigram segmentation model. That is, each phrase fragment obtained by word segmentation consists of 1 word. Each phrase fragment can be regarded as a unigram. Here, N=2 in the Bigram segmentation model. That is, each phrase fragment obtained by word segmentation consists of 2 words. Each phrase fragment can be regarded as a bigram. Here, N=3 in the Trigram segmentation model. That is, each phrase fragment obtained by word segmentation consists of 3 words. Each phrase fragment can be regarded as a trigram.

[0080] Through experiments, N in the Ngram segmentation model is preset to 1, 2, and 3, respectively, which can take into account the mapping accuracy of the second phrase set and the regulatory text database while taking into account the mapping speed of the second phrase set and the regulatory text database, thereby increasing the mapping efficiency of the second phrase set and the regulatory text database. If N is small, the mapping accuracy of the second phrase set and the regulatory text database will be reduced; otherwise, the mapping speed of the second phrase set and the regulatory text database will be reduced, and it is impossible to take into account both the mapping accuracy and the mapping speed.

[0081] The mapping of the second phrase set with the legal text database, that is, the second phrase set is respectively mapped with a unigram legal text database consisting of a number of unigrams, a two-gram legal text database consisting of a number of bigrams, and a three-gram legal text database consisting of a number of three-grams, to obtain a standard text with a weight in each legal text database. In this way, the matching results of the second phrase set and the database with different granularity elements are effectively integrated.

[0082] In a specific implementation provided by the present application, after obtaining the word segmentation of the case description text input by the user, a word segmentation operation can be performed on it. After the word segmentation is completed, the obtained word fragments can be post-processed, that is, the first phrase set after the word segmentation is normalized according to the preset rules. A second phrase set can be generated based on the normalized word segmentation result. After that, the second phrase set can be matched with the legal text database (inverted index database), and a number of candidate regulations with a mapping relationship with the case description text can be obtained.

[0083] The preset rules here are preset attributes. The second phrase set with preset attributes is screened out, that is, irrelevant elements such as stop words and punctuation marks in the first phrase set are excluded. Therefore, the preset attributes can be understood as non-irrelevant elements such as non-stop words and non-punctuation marks. In addition, in practical applications, the attribute of entity information can also be added to the preset attributes to identify related entities in the first group set. For example, identify entity information such as license plates, ID numbers, amounts, and organization names. In this way, it is easy to determine the category of regulations corresponding to the case description text. That is, determine which field of regulations the case description text corresponds to.

[0084] S400: Determine the matching degree between the case fact description text and the plurality of legal provisions through a pre-trained Bert-wwm coding matching model.

[0085] It is understandable that the second phrase set is mapped with the legal text database, and candidate laws that have certain weights in the first, second, and third legal text databases in the legal text database can be obtained. However, since each element in the second phrase set is relatively short, its semantic uncertainty will affect the accuracy of the mapping result to a large extent during the database mapping process. Therefore, it is impossible to directly output the relevant candidate laws obtained by mapping as the final match, and it is necessary to fully combine the contextual semantics of each candidate legal text and the case description text to determine the matching degree in order to obtain a standard text with higher credibility.

[0086] The Bert-wwm coding matching model here can be understood as a coding model with matching degree calculation rules. The Bert-wwm coding model adds a whole word mask function based on the original Bert model. After training, the Bert-wwm coding model can recognize word fragments with longer length (i.e., large granularity) compared to the original Bert model. That is, when processing texts of the same length, the Bert-wwm coding model recognizes faster and with higher accuracy. Therefore, by calculating the matching degree between the case description text and the candidate regulations through the Bert-wwm coding matching model, the semantics of the case description text and the semantics of the candidate regulations can be fully combined, thereby facilitating the determination of target regulations with higher matching degrees.

[0087] Furthermore, in a preferred embodiment provided in the present application, the method of determining the degree of matching between the case fact description text and the several legal provisions through a pre-trained Bert-wwm coding matching model specifically includes: determining several first coding vectors of the several legal provisions based on the several legal provisions and through a pre-trained Bert-wwm coding matching model; determining the second coding vector of the case fact description text through a pre-trained Bert-wwm coding matching model based on the case fact description text; calculating the cosine value of the angle between the first coding vector and the second coding vector using a preset vector similarity calculation rule; and determining the degree of matching between the case fact description text and the several legal provisions based on the calculation result of the cosine value.

[0088] It can be understood that when a certain text is input into the Bert coding model, the corresponding vector of the text will be obtained. Similarly, when a case fact description text is input into the pre-trained Bert-wwm coding matching model, the vector corresponding to the case fact description text can be obtained. When a candidate regulation is input into the pre-trained Bert-wwm coding matching model, the vector corresponding to the candidate regulation can be obtained. Here, the vector corresponding to the candidate regulation is referred to as the first coding vector; the vector corresponding to the case fact description text is referred to as the second coding vector. Therefore, determining the matching degree between the case fact description text and several legal provisions (candidate regulations) is converted into calculating the similarity between the first coding vector and the second coding vector. It is worth noting that the vector similarity calculation rules here have been preset in the Bert-wwm coding matching model.

[0089] It is understandable that the calculation of vector similarity can be achieved by calculating the cosine value of the vector angle. Moreover, the angle interval between the two vectors is 0°-360°, and the corresponding cosine value interval is [-1,1]. The closer the cosine value of the angle between the two vectors is to 1 or -1, the more similar they are. Therefore, by calculating the cosine value of the angle between the first coded vector and the second coded vector, the matching degree between the case fact description text and the several legal provisions can be determined.

[0090] Furthermore, in another preferred embodiment provided in the present application, the determination of the degree of matching between the case fact description text and the several legal provisions through a pre-trained Bert-wwm coding matching model specifically includes: determining several first coding vectors of the several legal provisions according to the several legal provisions and through a pre-trained Bert-wwm coding matching model; determining the second coding vector of the case fact description text through a pre-trained Bert-wwm coding matching model according to the case fact description text; obtaining a reorganized third vector according to the first coding vector and the second coding vector and using a vector recombination rule; determining the degree of matching between the case fact description text and the several legal provisions according to the third vector and through a pre-trained neural network fully connected algorithm.

[0091] It can be understood that when a text is input into the Bert coding model, the vector corresponding to the text will be obtained. Similarly, when a case fact description text is input into the pre-trained Bert-wwm coding matching model, the vector corresponding to the case fact description text can be obtained. When a candidate regulation is input into the pre-trained Bert-wwm coding matching model, the vector corresponding to the candidate regulation can be obtained. Here, the vector corresponding to the candidate regulation is referred to as the first coding vector; the vector corresponding to the case fact description text is referred to as the second coding vector.

[0092] The process of obtaining the reorganized third vector according to the first coding vector and the second coding vector and using the vector reorganization rule can be understood as expanding the number of elements in the third vector. The vector reorganization rule here can be understood as a preset vector calculation method. In a specific embodiment provided in the present application, the vector reorganization rule is to subtract the absolute value of the first coding vector and the second coding vector bit by bit; or, to multiply the first coding vector and the second coding vector bit by bit. Bit by bit here can be understood as the elements located at the same position in the first coding vector and the second coding vector are calculated according to the preset calculation method. When the first coding vector and the second coding vector are calculated, the first coding vector, the second coding vector and the calculated vector obtained by calculation can be reorganized into a new vector. Therefore, in addition to the first coding vector and the second coding vector, the third vector also includes a calculated vector obtained by bitwise multiplication or bitwise subtraction. In this way, the semantics of each candidate regulation can be effectively combined with the semantics of the case description text, and the matching degree of the two can be calculated based on the separate semantics and associated semantics of the two.

[0093] Specifically, the matching degree between the case fact description text and several legal provisions is determined based on the third vector and through the pre-trained neural network fully connected algorithm. It is worth noting that the pre-trained neural network fully connected algorithm here has been preset in the Bert-wwm coding matching model. The neural network fully connected algorithm here mainly uses the fully connected layer of the neural network and the 2-way softmax classifier to obtain the similarity probability between the case fact description text and several legal provisions, thereby obtaining the matching degree between the two.

[0094] Furthermore, in a preferred embodiment provided in the present application, the training process of the pre-trained Bert-wwm coding matching model includes: using a number of standard regulatory texts and a number of training case fact description texts to perform negative feedback optimization of the Bert-wwm coding matching model.

[0095] The Bert-wwm coding matching model here can be understood as a coding model with matching degree calculation rules. The matching degree calculation rules here are mainly the preset neural network fully connected algorithm. Among them, the neural network fully connected algorithm mainly uses the fully connected layer of the neural network and the 2-way softmax classifier. The training of the Bert-wwm coding matching model is mainly to train the Bert-wwm coding ability and the preset neural network fully connected algorithm. The several standard regulatory texts here are standard regulations without word segmentation, which can be flexibly selected according to the actual application scenarios. The actual application scenarios here can be obtained from the case fact description text used for training. The negative feedback optimization here is to optimize the relevant parameters in the model according to the output value of the Bert-wwm coding matching model during the training process.

[0096] In a specific implementation provided in the present application, the Bert-wwm coding matching model is trained and a legal knowledge graph is also used to make the trained Bert-wwm coding matching model more applicable to the legal field, thereby increasing the accuracy of matching between the case description text and the regulatory text.

[0097] Please refer to Figure 2 , a device 100 for matching case facts with legal provisions provided in an embodiment of the present application is mainly used for matching case description text with legal provisions in the legal field. That is, according to the case description information, legal provisions with a certain degree of relevance to the current case are matched, so as to facilitate the subsequent work of relevant personnel. Specifically, a device 100 for matching case facts with legal provisions includes:

[0098] A word segmentation module 11 is used to segment the case fact description text to generate a first phrase set;

[0099] A screening module 12, configured to screen out a second set of phrases having a preset attribute from the first set of phrases;

[0100] A mapping module 13, used to establish a mapping of the second phrase set in a legal text database, and obtain a number of legal provisions having a mapping relationship with the second phrase set;

[0101] The matching degree calculation module 14 is used to determine the matching degree between the case fact description text and the plurality of legal provisions through a pre-trained Bert-wwm coding matching model.

[0102] The word segmentation module 11 is used to segment the case fact description text and generate a first phrase set. The case fact description text here mainly refers to the important fact description text used in the description of legal cases. In practical applications, it can be understood as a legal document, for example, it can be expressed as at least one legal document in the case acceptance approval form, conversation record, power of attorney, indictment, appeal, counterclaim, litigation preservation, evidence preservation, court ruling, judgment, mediation, and case closing report. The case fact description text is segmented, that is, the acquired case description text is segmented. After the case fact description text is segmented, a phrase set consisting of phrases in several case fact description texts is obtained. That is, a first phrase set is generated. The first phrase set here corresponds to a set consisting of several case fact description text phrases.

[0103] In a specific implementation provided in the present application, the text describing the facts of the case is segmented using an Ngram segmentation model. The Ngram segmentation model can perform a sliding window operation of size N on the content of the text according to the words, thereby forming a sequence of word fragments of length N. And during the segmentation process, the appearance of the Nth word is related to the previous N-1 words. That is, it is possible to fully combine the semantics of the above text to perform segmentation and obtain several phrase fragments of a preset length. In actual applications, the length of the word fragments obtained by segmentation depends on the preset value of N, which can be set according to actual conditions. For example, the case fact description text can be segmented by a Unigram segmentation model (i.e., N=1), and a first set of words consisting of several unigrams (i.e., a unigram set) can be obtained; or, the case fact description text can be segmented by a Bigram segmentation model (i.e., N=2), and a first set of words consisting of several bigrams (i.e., a bigram set) can be obtained; or, the case fact description text can be segmented by a Trigram segmentation model (i.e., N=3), and a first set of words consisting of several ternary words (i.e., a ternary set) can be obtained. The specific value of N here obviously does not constitute a limitation on the protection scope of this application.

[0104] Furthermore, in a preferred embodiment provided in the present application, the word segmentation module 11 is used to segment the case fact description text to generate a first phrase set, specifically used to: segment the case fact description text through a Unigram word segmentation model to generate a unigram phrase set consisting of a number of unigram phrases.

[0105] Here, the Unigram word segmentation model is preferred to perform word segmentation on the case fact description text, that is, N=1 in the Ngram word segmentation model. At this time, each phrase fragment obtained by word segmentation consists of one word. For example, the description of "people under the age of eighteen" in the case fact description text is segmented by the Unigram word segmentation model to obtain a unigram phrase set consisting of unigram phrases such as "under", "eighteen", "years old", "of", and "people". Compared with the original case fact description text, the unigram phrase set is more detailed and has a smaller granularity, which can effectively increase the accuracy of subsequent corresponding regulations matching.

[0106] The screening module 12 is used to screen out a second phrase set with preset attributes in the first phrase set. The preset attributes here can be understood as preset screening rules. It can be understood that the case fact description text can obtain a first phrase set composed of several word segmentation fragments after word segmentation processing. However, punctuation marks or stop words and other contents in the case fact description text will also exist in the first phrase set as relevant elements. Here, punctuation marks or stop words and other contents in the case fact description text are regarded as irrelevant elements in the first phrase set. These irrelevant elements do not increase the accuracy of the matching between the case description text and the regulatory terms, but will increase the amount of calculation in the matching process between the case description text and the regulatory terms. Therefore, the present application deletes these irrelevant elements so as to reduce the amount of calculation in the matching process without affecting the accuracy of the matching between the case description text and the regulatory terms, thereby improving the matching efficiency between the case description text and the regulatory terms. For example, when the case facts description text is segmented by a Unigram word segmentation model (i.e., N=1) to obtain a first phrase set (i.e., a unigram phrase set) consisting of several unigram phrases, the corresponding second phrase set is the screened unigram phrase set. Alternatively, when the case facts description text is segmented by a Bigram word segmentation model (i.e., N=2) to obtain a first phrase set (i.e., a bigram phrase set) consisting of several bigram phrases, the corresponding second phrase set is the screened bigram phrase set. Alternatively, when the case facts description text is segmented by a Trigram word segmentation model (i.e., N=3) to obtain a first phrase set (i.e., a trigram phrase set) consisting of several trigram phrases, the corresponding second phrase set is the screened trigram phrase set. The specific value of N here obviously does not constitute a limitation on the scope of protection of this application.

[0107] In a specific implementation provided by the present application, the screening module 12 can also perform normalization processing on the related phrases in the first phrase set during the screening of the second phrase set with preset attributes. For example, the entity information such as license plate, ID number, amount, organization name, etc. in the first phrase set is identified. That is, the phrases in the first phrase set are subjected to corpus recognition, and stop words are deleted, while other phrases remain unchanged.

[0108] The mapping module 13 is used to establish a mapping of the second phrase set in the legal text database, and obtain a number of legal provisions that have a mapping relationship with the second phrase set. The legal text database here is a database obtained through preprocessing and recording relevant legal provisions. It can be understood that, compared with the first phrase set, the second phrase set has deleted irrelevant elements such as punctuation marks or stop words, reducing the number of elements in the set, thereby effectively increasing the matching efficiency between the second phrase set and the legal text database. The matching of the second phrase set with the legal text database is to establish a mapping in the legal database according to each element in the second phrase set, so as to obtain a number of legal provisions that have a mapping relationship with the case description. It can also be understood that candidate regulations that have a certain degree of correlation with the case description are obtained.

[0109] It is worth noting that the regulatory text database here is also obtained by segmenting several regulatory clauses through the Ngram segmentation model. Here, the regulatory clauses before segmentation are regarded as standard regulatory texts. Standard regulatory texts are standard texts composed of coherent language formulated by relevant organizations; each standard text is composed of several word fragments. In this way, in the mapping process between the second phrase set and the regulatory text database, the mapping of the elements in the second phrase set and the elements in the regulatory text database in terms of expression form and semantics can be guaranteed. It can be understood that two word fragments may have the same semantics when their expression forms are different; or the expression forms are the same but the semantics are different. Therefore, segmenting the standard regulatory text can effectively avoid the phenomenon that the two elements have the same meaning but different expression forms, or the same expression forms but different meanings during the mapping process, thereby increasing the accuracy of the mapping between the second phrase set and the regulatory text database.

[0110] In practical applications, the legal text database may be in the form of an inverted index table, which records each word segment and the position of the word segment in the legal text database.

[0111] It is understandable that the specific value of N in the Ngram word segmentation model can be preset according to actual conditions. The larger the value of N, the longer the word fragment obtained; the smaller the value of N, the smaller the word fragment obtained. Moreover, the appearance of the Nth word fragment is related to the previous N-1 word fragments. Then the longer each word fragment is, the more accurate its contextual semantics is. Correspondingly, the shorter each word fragment is, the more possibilities its contextual semantics has. The present application uses Ngram word segmentation models with different preset N values ​​to segment standard regulatory texts respectively. In this way, word fragments with different granularities can be used, so that the second phrase set can be matched in more dimensions, thereby obtaining a standard text with a more accurate mapping.

[0112] Furthermore, in a preferred embodiment provided in the present application, the word segmentation module 11 is also used to: segment the standard regulatory text through the Unigram word segmentation model to generate a unigram regulatory text database consisting of several unigrams; segment the standard regulatory text through the Bigram word segmentation model to generate a bigram regulatory text database consisting of several bigrams; segment the standard regulatory text through the Trigram word segmentation model to generate a trigram regulatory text database consisting of several trigrams; merge the unigram regulatory text database, the bigram regulatory text database, and the trigram regulatory text database to generate a regulatory text database.

[0113] Here, N=1 in the Unigram segmentation model. That is, each phrase fragment obtained by word segmentation consists of 1 word. Each phrase fragment can be regarded as a unigram. Here, N=2 in the Bigram segmentation model. That is, each phrase fragment obtained by word segmentation consists of 2 words. Each phrase fragment can be regarded as a bigram. Here, N=3 in the Trigram segmentation model. That is, each phrase fragment obtained by word segmentation consists of 3 words. Each phrase fragment can be regarded as a trigram.

[0114] Through experiments, N in the Ngram segmentation model is preset to 1, 2, and 3, respectively, which can take into account the mapping accuracy of the second phrase set and the regulatory text database while taking into account the mapping speed of the second phrase set and the regulatory text database, thereby increasing the mapping efficiency of the second phrase set and the regulatory text database. If N is small, the mapping accuracy of the second phrase set and the regulatory text database will be reduced; otherwise, the mapping speed of the second phrase set and the regulatory text database will be reduced, and it is impossible to take into account both the mapping accuracy and the mapping speed.

[0115] The mapping of the second phrase set with the legal text database, that is, the second phrase set is respectively mapped with a unigram legal text database consisting of a number of unigrams, a two-gram legal text database consisting of a number of bigrams, and a three-gram legal text database consisting of a number of three-grams, to obtain a standard text with a weight in each legal text database. In this way, the matching results of the second phrase set and the database with different granularity elements are effectively integrated.

[0116] In a specific implementation provided by the present application, after obtaining the word segmentation of the case description text input by the user, a word segmentation operation can be performed on it. After the word segmentation is completed, the obtained word fragments can be post-processed, that is, the first phrase set after the word segmentation is normalized according to the preset rules. A second phrase set can be generated based on the normalized word segmentation result. After that, the second phrase set can be matched with the legal text database (inverted index database), and a number of candidate regulations with a mapping relationship with the case description text can be obtained.

[0117] The preset rules here are preset attributes. The second phrase set with preset attributes is screened out, that is, irrelevant elements such as stop words and punctuation marks in the first phrase set are excluded. Therefore, the preset attributes can be understood as non-irrelevant elements such as non-stop words and non-punctuation marks. In addition, in practical applications, the attribute of entity information can also be added to the preset attributes to identify related entities in the first group set. For example, identify entity information such as license plates, ID numbers, amounts, and organization names. In this way, it is easy to determine the category of regulations corresponding to the case description text. That is, determine which field of regulations the case description text corresponds to.

[0118] The matching degree calculation module 14 is used to determine the matching degree between the case fact description text and the several legal provisions through the pre-trained Bert-wwm coding matching model. It can be understood that the second phrase set is mapped with the legal text database, and the candidate regulations that have certain weights in the first, second, and third legal text databases in the legal text database can be obtained. However, since each element in the second phrase set is relatively short, its semantic uncertainty will affect the accuracy of the mapping result to a large extent during the database mapping process. Therefore, it is impossible to directly output the relevant candidate regulations obtained by mapping as the final match, and it is necessary to fully combine the contextual semantics of each candidate legal text and the case description text to determine the matching degree in order to obtain a standard text with higher credibility.

[0119] The Bert-wwm coding matching model here can be understood as a coding model with matching degree calculation rules. The Bert-wwm coding model adds a whole word mask function based on the original Bert model. After training, the Bert-wwm coding model can recognize word fragments with longer length (i.e., large granularity) compared to the original Bert model. That is, when processing texts of the same length, the Bert-wwm coding model recognizes faster and with higher accuracy. Therefore, by calculating the matching degree between the case description text and the candidate regulations through the Bert-wwm coding matching model, the semantics of the case description text and the semantics of the candidate regulations can be fully combined, thereby facilitating the determination of target regulations with higher matching degrees.

[0120] Furthermore, in a preferred embodiment provided in the present application, the matching degree calculation module 14 is used to determine the matching degree between the case fact description text and the several legal provisions through a pre-trained Bert-wwm coding matching model, and is specifically used to: determine the several first coding vectors of the several legal provisions based on the several legal provisions and through a pre-trained Bert-wwm coding matching model; determine the second coding vector of the case fact description text based on the case fact description text through a pre-trained Bert-wwm coding matching model; use a preset vector similarity calculation rule to calculate the cosine value of the angle between the first coding vector and the second coding vector; and determine the matching degree between the case fact description text and the several legal provisions based on the calculation result of the cosine value.

[0121] It can be understood that when a certain text is input into the Bert coding model, the corresponding vector of the text will be obtained. Similarly, when a case fact description text is input into the pre-trained Bert-wwm coding matching model, the vector corresponding to the case fact description text can be obtained. When a candidate regulation is input into the pre-trained Bert-wwm coding matching model, the vector corresponding to the candidate regulation can be obtained. Here, the vector corresponding to the candidate regulation is referred to as the first coding vector; the vector corresponding to the case fact description text is referred to as the second coding vector. Therefore, determining the matching degree between the case fact description text and several legal provisions (candidate regulations) is converted into calculating the similarity between the first coding vector and the second coding vector. It is worth noting that the vector similarity calculation rules here have been preset in the Bert-wwm coding matching model.

[0122] It is understandable that the calculation of vector similarity can be achieved by calculating the cosine value of the vector angle. Moreover, the angle interval between the two vectors is 0°-360°, and the corresponding cosine value interval is [-1,1]. The closer the cosine value of the angle between the two vectors is to 1 or -1, the more similar they are. Therefore, by calculating the cosine value of the angle between the first coded vector and the second coded vector, the matching degree between the case fact description text and the several legal provisions can be determined.

[0123] Furthermore, in another preferred embodiment provided in the present application, the matching degree calculation module 14 is used to determine the matching degree between the case fact description text and the several legal provisions through a pre-trained Bert-wwm coding matching model, and is specifically used to: determine the several first coding vectors of the several legal provisions according to the several legal provisions and through a pre-trained Bert-wwm coding matching model; determine the second coding vector of the case fact description text according to the case fact description text through a pre-trained Bert-wwm coding matching model; obtain a reorganized third vector according to the first coding vector and the second coding vector and using a vector recombination rule; determine the matching degree between the case fact description text and the several legal provisions according to the third vector and through a pre-trained neural network fully connected algorithm.

[0124] It can be understood that when a text is input into the Bert coding model, the vector corresponding to the text will be obtained. Similarly, when a case fact description text is input into the pre-trained Bert-wwm coding matching model, the vector corresponding to the case fact description text can be obtained. When a candidate regulation is input into the pre-trained Bert-wwm coding matching model, the vector corresponding to the candidate regulation can be obtained. Here, the vector corresponding to the candidate regulation is referred to as the first coding vector; the vector corresponding to the case fact description text is referred to as the second coding vector.

[0125] The process of obtaining the reorganized third vector according to the first coding vector and the second coding vector and using the vector reorganization rule can be understood as expanding the number of elements in the third vector. The vector reorganization rule here can be understood as a preset vector calculation method. In a specific embodiment provided in the present application, the vector reorganization rule is to subtract the absolute value of the first coding vector and the second coding vector bit by bit; or, to multiply the first coding vector and the second coding vector bit by bit. Bit by bit here can be understood as the elements located at the same position in the first coding vector and the second coding vector are calculated according to the preset calculation method. When the first coding vector and the second coding vector are calculated, the first coding vector, the second coding vector and the calculated vector obtained by calculation can be reorganized into a new vector. Therefore, in addition to the first coding vector and the second coding vector, the third vector also includes a calculated vector obtained by bitwise multiplication or bitwise subtraction. In this way, the semantics of each candidate regulation can be effectively combined with the semantics of the case description text, and the matching degree of the two can be calculated based on the separate semantics and associated semantics of the two.

[0126] Specifically, according to the third vector and through the pre-trained neural network fully connected algorithm, the matching degree between the case fact description text and several legal provisions is determined. It is worth noting that the pre-trained neural network fully connected algorithm here is preset in the Bert-wwm coding matching model. The neural network fully connected algorithm here mainly uses the fully connected layer of the neural network and the 2-way softmax classifier to obtain the similarity probability between the case fact description text and several legal provisions, thereby obtaining the matching degree between the two.

[0127] Furthermore, in a preferred embodiment provided in the present application, the training process of the pre-trained Bert-wwm coding matching model includes: using a number of standard regulatory texts and a number of training case fact description texts to perform negative feedback optimization of the Bert-wwm coding matching model.

[0128] The Bert-wwm coding matching model here can be understood as a coding model with matching degree calculation rules. The matching degree calculation rules here are mainly the preset neural network fully connected algorithm. Among them, the neural network fully connected algorithm mainly uses the fully connected layer of the neural network and the 2-way softmax classifier. The training of the Bert-wwm coding matching model is mainly to train the Bert-wwm coding ability and the preset neural network fully connected algorithm. The several standard regulatory texts here are standard regulations without word segmentation, which can be flexibly selected according to the actual application scenarios. The actual application scenarios here can be obtained from the case fact description text used for training. The negative feedback optimization here is to optimize the relevant parameters in the model according to the output value of the Bert-wwm coding matching model during the training process.

[0129] In a specific implementation provided in the present application, the Bert-wwm coding matching model is trained and a legal knowledge graph is also used to make the trained Bert-wwm coding matching model more applicable to the legal field, thereby increasing the accuracy of matching between the case description text and the regulatory text.

[0130] It should be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of further restrictions, an element defined by the phrase "comprises a ..." does not exclude the existence of other identical elements in the process, method, commodity or device including the element.

[0131] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for matching case facts with legal provisions, characterized in that: The following steps are involved: Segmenting the text describing the case facts to generate a first phrase set; specifically, segmenting the text describing the case facts by a Unigram word segmentation model to generate a unigram word set consisting of a plurality of unigram word sets; or segmenting the text describing the case facts by a Bigram word segmentation model to obtain a first phrase set consisting of a plurality of bigram word sets; or segmenting the text describing the case facts by a Trigram word segmentation model to obtain a first phrase set consisting of a plurality of trigram word sets; Filtering out a second set of phrases having a preset attribute from the first set of phrases; Segmenting the standard legal text by Unigram segmentation model to generate a unigram legal text database consisting of several unigrams; segmenting the standard legal text by Bigram segmentation model to generate a bigram legal text database consisting of several bigrams; segmenting the standard legal text by Trigram segmentation model to generate a trigram legal text database consisting of several trigrams; merging the unigram legal text database, the bigram legal text database and the trigram legal text database to generate a legal text database; the legal text database is embodied in the form of an inverted index table; Establishing a mapping of the second phrase set in the legal text database, specifically, mapping the second phrase set with a unigram legal text database consisting of a number of unigrams, a binary legal text database consisting of a number of binary phrases, and a ternary legal text database consisting of a number of ternary phrases, respectively, to obtain a number of legal provisions having a mapping relationship with the second phrase set; The matching degree between the case fact description text and the several legal provisions is determined through the pre-trained Bert-wwm coding matching model.

2. The method for matching case facts with legal provisions as claimed in claim 1, characterized in that: The pre-trained Bert-wwm coding matching model is used to determine the matching degree between the case fact description text and the several legal provisions, specifically including: Determine, according to the plurality of legal provisions and through a pre-trained Bert-wwm coding matching model, a plurality of first coding vectors of the plurality of legal provisions; Determine a second encoding vector of the case fact description text according to the case fact description text and through a pre-trained Bert-wwm encoding matching model; Using a preset vector similarity calculation rule, calculate the cosine value of the angle between the first encoding vector and the second encoding vector; According to the calculation result of the cosine value, the matching degree between the case fact description text and the several legal provisions is determined.

3. The method for matching case facts with legal provisions as claimed in claim 1, characterized in that: The pre-trained Bert-wwm coding matching model is used to determine the matching degree between the case fact description text and the several legal provisions, specifically including: Determine, according to the plurality of legal provisions and through a pre-trained Bert-wwm coding matching model, a plurality of first coding vectors of the plurality of legal provisions; Determine a second encoding vector of the case fact description text through a pre-trained Bert-wwm encoding matching model according to the case fact description text; Obtain a recombined third vector according to the first encoding vector and the second encoding vector and using a vector recombination rule; According to the third vector and through the pre-trained neural network fully connected algorithm, the matching degree between the case fact description text and the several legal provisions is determined.

4. The method for matching case facts with legal provisions as claimed in claim 3, characterized in that: The training process of the pre-trained Bert-wwm coding matching model includes: Several standard regulatory texts and several training case fact description texts are used to perform negative feedback optimization of the Bert-wwm coding matching model.

5. A device for matching case facts with legal provisions, characterized in that: include: A word segmentation module is used to segment the case fact description text to generate a first phrase set; specifically, it is used to: segment the case fact description text through a Unigram word segmentation model to generate a unigram phrase set consisting of a number of unigram phrases; or, segment the case fact description text through a Bigram word segmentation model to obtain a first phrase set consisting of a number of bigram phrases; or, segment the case fact description text through a Trigram word segmentation model to obtain a first phrase set consisting of a number of trigram phrases; A screening module, used to screen out a second set of phrases having a preset attribute from the first set of phrases; The word segmentation module is also used to: segment the standard legal text through the Unigram word segmentation model to generate a unigram legal text database consisting of several unigrams; segment the standard legal text through the Bigram word segmentation model to generate a bigram legal text database consisting of several bigrams; segment the standard legal text through the Trigram word segmentation model to generate a trigram legal text database consisting of several trigrams; merge the unigram legal text database, the bigram legal text database, and the trigram legal text database to generate a legal text database; the legal text database is embodied in the form of an inverted index table; A mapping module, used to establish a mapping of the second phrase set in the legal text database, specifically, the second phrase set is respectively mapped with a unigram legal text database consisting of a number of unigrams, a binary legal text database consisting of a number of binary phrases, and a ternary legal text database consisting of a number of ternary phrases, to obtain a number of legal provisions having a mapping relationship with the second phrase set; The matching degree calculation module is used to determine the matching degree between the case fact description text and the several legal provisions through a pre-trained Bert-wwm coding matching model.

6. The device for matching case facts with legal provisions as claimed in claim 5, characterized in that: The matching degree calculation module is used to determine the matching degree between the case fact description text and the several legal provisions through the pre-trained Bert-wwm coding matching model, specifically for: Determine, according to the plurality of legal provisions and through a pre-trained Bert-wwm coding matching model, a plurality of first coding vectors of the plurality of legal provisions; Determine a second encoding vector of the case fact description text according to the case fact description text and through a pre-trained Bert-wwm encoding matching model; Using a preset vector similarity calculation rule, calculate the cosine value of the angle between the first encoding vector and the second encoding vector; According to the calculation result of the cosine value, the matching degree between the case fact description text and the several legal provisions is determined.

7. The device for matching case facts with legal provisions as claimed in claim 5, characterized in that: The matching degree calculation module is used to determine the matching degree between the case fact description text and the several legal provisions through the pre-trained Bert-wwm coding matching model, specifically for: Determine, according to the plurality of legal provisions and through a pre-trained Bert-wwm coding matching model, a plurality of first coding vectors of the plurality of legal provisions; Determine a second encoding vector of the case fact description text through a pre-trained Bert-wwm encoding matching model according to the case fact description text; Obtain a recombined third vector according to the first encoding vector and the second encoding vector and using a vector recombination rule; According to the third vector and through the pre-trained neural network fully connected algorithm, the matching degree between the case fact description text and the several legal provisions is determined.

Images