CN111309900B - A method for judging and pushing the similarity of similar legal cases - Google Patents

Abstract

The invention discloses a legal class similarity judging and pushing method, wherein the related judging method comprises the following steps: classifying target legal cases, and extracting historical cases of the same category from a historical case database to form a candidate set according to the obtained case category; performing event sequence representation on the target legal case and each similar historical case in the candidate set; calculating the distance between the event sequence corresponding to the target legal case and the event sequence corresponding to each historical case in the candidate set according to the event sequence measurement model; and calculating the similarity between the target legal case and the candidate centralized historical case based on the distance of the event sequence and the scoring function. The method can realize more comprehensive and accurate class case identification; meanwhile, the legal documents are expressed as a time sequence event sequence, similarity calculation is carried out based on an unsupervised mode, and historical cases with higher scores are selected for pushing, so that labor input is greatly reduced, and pushing intellectualization can be better realized.

Description

A method for judging and pushing the similarity of similar legal cases

technical field

The invention relates to the field of legal intelligence, in particular to a method for judging and pushing the similarity of legal similar cases.

Background technique

At present, the theory and technology of artificial intelligence are becoming more and more mature, and the scope of application is constantly expanding. In 2017, the National Artificial Intelligence Strategy "New Generation Artificial Intelligence Development Plan" proposed to build a smart court, promote the application of artificial intelligence in evidence collection, case analysis, legal document reading and analysis, and realize the intelligentization of the court's trial system and trial capacity. Among them, the realization of similar cases and similar judgments through artificial intelligence technology has become an important research content that is close to the needs of judges.

As an auxiliary tool, the purpose of similar judgments is to find similar or even identical cases for the cases the judges are dealing with, so as to inspire and expand the judges' thinking in judging cases, help the judges make correct judgments, and make the judgment results of the same or similar cases more consistent. Minor deviation. However, the existing similar case retrieval system is inaccurate in pushing cases and cannot effectively solve the problem that judges need. For example, the pushed cases are not "same case" or even "similar"; the number of pushed cases is too high, which does not really save the judge's time, and still requires a lot of manual screening.

Since most legal case records are electronic documents, their form is text expressed in natural language. Therefore, class case recognition can be regarded as an application scenario of text similarity measurement. The application of existing natural language processing methods can realize the identification of similar cases to a certain extent, but it is still difficult to accurately identify the core distinguishing points of the elements of a case. The main issues are as follows:

1) The method based on keyword matching is not accurate enough. Keyword search is actually "sampling verification", and the conclusions drawn with the help of a small number of samples are not complete. At the same time, the number of similar cases obtained by this method is too large, making it difficult for judges to identify cases with important reference value.

2) The method of constructing a neural network by expressing words as vectors based on word2vec requires a large amount of labeled and structured training corpus, while the current legal field lacks massive and detailed labeled legal data, and lacks both legal and technical talents .

3) The main reference value of similar cases is to push forward the judgment ideas and practices of judges in similar historical cases based on some legal details or difficulties in the cases. However, there is no legal document similarity measurement model designed for the characteristics of the legal industry.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for judging and pushing the similarity of legal similar cases, which solves the shortcomings of the existing method, such as the need for a large number of manual annotations, inaccurate push of similar cases, redundant information, and lack of legal issues pertinence.

The purpose of this invention is to realize through the following technical solutions:

A method for judging similarity of similar legal cases, including:

Classify the target legal case, according to the obtained case category, extract the historical cases of the same category from the historical case database to form a candidate set;

Represent the event sequence for the target legal case and each similar historical case in the candidate set;

According to the event sequence measurement model, calculate the distance between the event sequence corresponding to the target legal case and the event sequence corresponding to each historical case in the candidate set;

Based on the distance of the event sequence and combined with the scoring function, the similarity between the target legal case and the historical cases in the candidate set is calculated.

A method for pushing similar legal cases, comprising: using the aforementioned method to calculate the similarity between a target legal case and a historical case in a candidate set, and then sorting the historical cases in the candidate set in descending order of similarity scores, and extracting The top M historical cases are pushed.

It can be seen from the technical solution provided by the present invention that, by classifying legal documents and analyzing the distribution of case subjects, the historical cases that are most similar to the semantic information described in the target case can be selected while ensuring the same case category. Realize more comprehensive and accurate identification of similar cases; at the same time, by representing legal documents as time-series event sequences and performing similarity calculation based on an unsupervised method, historical cases with higher scores are selected for push, which greatly reduces manpower investment and can Better realization of push intelligence.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the following briefly introduces the accompanying drawings used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without any creative effort.

1 is a flowchart of a method for judging similarity of legal similar cases provided by an embodiment of the present invention;

2 is a schematic diagram of an event time chain represented by extracting events corresponding to each time period provided by an embodiment of the present invention;

3 is a schematic diagram of an event time chain extracted from a specific case provided by an embodiment of the present invention;

FIG. 4 is a schematic diagram of aligning event sequence time points according to an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present invention.

An embodiment of the present invention provides a method for judging the similarity of legal similar cases, as shown in FIG. 1 , including:

Step 1. Classify the target legal case, and according to the obtained case category, extract historical cases of the same category from the historical case database to form a candidate set.

Similar case identification aims to find similar or even identical cases to the target case in similar historical cases, so case classification should be carried out first.

Each case in the existing historical case database has a corresponding category label. According to the case category distribution, a certain number of historical cases can be selected for each case category to construct a data set. By using a relatively mature deep learning text classification algorithm in natural language processing as a classifier, the case category corresponding to the target case is obtained. According to the obtained case category, similar historical cases in the historical case database are extracted to form a candidate set.

Step 2. Perform event sequence representation on the target legal case and each similar historical case in the candidate set.

In this step, a text representation model is designed according to the characteristics of legal cases, which can effectively represent the connection between events and present the evolution process of the case over time, which has higher practical application value. The preferred embodiment of this step is as follows:

1) For any legal case, use information extraction technology to extract the case elements in the legal case documents. The case elements at least include: the defendant's position, the defendant's attitude after the case, time, core words, and the amount involved.

Among them, the core words are extracted by the method of dependency syntax analysis. In short, the sentences in the legal case document are divided into sentences according to periods, exclamation marks and question marks. Get n core words.

Dependency syntactic analysis is a relatively mature technology at present. Its principle is: the entire sentence structure is expressed through the dependencies between words. These dependencies express the semantic dependencies between the components of the sentence and the dependencies between all words. A syntax tree is formed, and the root node of the tree is the core predicate of the sentence, which is used to express the core content of the entire sentence. The core predicate is the core word corresponding to the sentence.

2) Independently retain the defendant's position and the defendant's attitude after the incident.

3) The remaining case elements are qualitatively represented by the time sequence relationship between the events according to different time nodes. For the events with no overlap and no overlap in occurrence time, they are regarded as independent events, and the remaining cases are combined to form a case. Plot event chain.

Those skilled in the art can understand that legal case documents are usually described in chronological order. For example, two cases given below, non-overlapping time periods correspond to different events. The event is first located by time, and then the corresponding event description sentence within the time is found. Then, the unstructured event description sentence is structured into a combination of various case elements through event extraction technology (extracting each case element).

4) Numerically represent the case elements with events occurring at time i to obtain event _i =(e _i1 , e _i2 ,..., e _in ), where n is the number of case elements with events occurring at time i.

将权重向量与event_i中每个元素对应相乘，得到时间i事件的最终表示vector_i＝(w₁e_i1，w₂e_i2，…，w_ne_in)；。5) Initialize the weight vector weight=(w ₁ , w ₂ , . . . , _wn ), where,

Multiply the weight vector with each element in event _i correspondingly to obtain the final representation vector _i of the event i at time i = (w ₁ e _i1 , w ₂ e _i2 , . . . , w _n e _in );.

In the embodiment of the present invention, the weight vector is a priori knowledge and is pre-initialized through experience. In the follow-up learning process, the results can be updated, and the specific update method can be selected by the user according to the situation or experience.

6) Connect the event representations of all times to obtain a vectorized sequential event chain EventSequence=(vector ₁ , vector ₂ , ..., vector _m ), where vector _i represents the vector representation corresponding to the event that occurs at time i, m is the total number of independent events and combined events extracted from legal case documents.

Step 3: Calculate the distance between the event sequence corresponding to the target legal case and the event sequence corresponding to each historical case in the candidate set according to the event sequence measurement model.

Sequence distance measurement is performed according to the event sequence obtained in step 2. When measuring the sequence similarity of events, there are cases where the time lengths of the sequences to be compared are inconsistent, so the sequences need to be aligned. Using the dynamic time warping DTW method, the event sequence corresponding to the target legal case is matched with each event sequence corresponding to the historical case, starting from the starting point, and each time a point is reached, the distance between the corresponding two points is calculated, and all previously passed through are accumulated. Point distance, and finally select the smallest accumulated distance as the event sequence distance Distance _{EventSequence} . This method can minimize the point-to-point matching of the distance between the two sequences by finding the correspondence between the points.

Step 4: Calculate the similarity between the target legal case and the historical case in the candidate set based on the distance of the event sequence and combining with the scoring function.

In the embodiment of the present invention, considering the similarity of the subject, the distance of the event sequence obtained in step 3, and the similarity of the position of the defendant obtained in step 2 and the attitude of the defendant after the incident, the main factors are as follows:

1) Through the topic analysis model, perform topic analysis on the target legal case and the historical cases in the candidate set, and obtain the corresponding topic probability distribution. According to the topic probability distribution, calculate the semantic distance between the target legal case and the historical cases in the candidate set as the topic similarity Distance _topic .

2) In the event sequence representation, the two elements of the defendant's position and the defendant's attitude after the incident were extracted from the target legal case and the historical cases in the candidate set, respectively, and the cosine distance was used to calculate the target legal case and the historical cases in the candidate set. The similarity of the defendant's position and the defendant's attitude after the incident was recorded as Distance _position and Distance _attitude respectively.

3) The event sequence distance between the target legal case and the historical cases in the candidate set is recorded as Distance _{EventSequence} .

4) Calculate the similarity between the target legal case and the historical cases in the candidate set using the following formula:

score=α ₁ Distance _topic +α ₂ Distance _{EventSequence} +α ₃ Distance _position +α ₄ Distance _attitude

Among them, α ₁ , α ₂ , α ₃ and α ₄ are all weights, and α ₁ +α ₂ +α ₃ +α ₄ =1.

For ease of understanding, the following description is given in conjunction with specific examples, and the case types and case information involved in the following examples are all examples.

Case 1: During the period when the defendant, Mr. Sun, was a salesperson of Company A, he repeatedly misappropriated the project receivables of Company A for personal use. The specific facts are as follows: 1. During the period from April to July 2017, the defendant, Mr. Sun, received 96,000 yuan for the B1 community project paid by Wu, and then handed over 65,000 yuan to Company A, and returned the remaining 31,000 yuan to the company. personal use. 2. During October 2017, the defendant, Mr. Sun, received 58,000 yuan for the B2 community project paid by Mr. Wang, but did not hand it over to Company A for personal use. 3. During December 2017, the defendant Sun received 113,000 yuan for the B3 community project paid by Zhou, but did not hand it over to Company A for personal use. After the incident, the defendant Sun Mou truthfully confessed the facts of the crime, returned RMB 100,000 to the victim unit (namely Company A) and obtained an understanding.

Case 2: During the period when the defendant Ye was a sales consultant of an automobile sales service company (hereinafter referred to as Company C), he took advantage of his position to sell the company's vehicles to customers without permission for many times, and used part of the car money for himself without handing it over. to Company C. The specific criminal facts are as follows: 1. On November 15, 2017, the defendant Ye received a car payment of RMB 130,500 from the customer Li for his own use. 2. On July 31, 2018, the defendant Ye received the car payment of RMB 52,678 from the customer Zhou for his own use. 3. On August 4, 2018, the defendant Ye received a deposit of RMB 20,000 from the customer to purchase a car on his behalf and used it for himself. After the incident, the defendant, Ye Mou, has refunded all the economic losses of Company C and obtained a written understanding from the company.

See the method provided above:

Step 1. First, classify the cases. Select the text classification algorithm in deep learning, such as BERT, FastText, DPCNN, etc., as the classifier, classify the events of Case 1, and get the category as "crime of misappropriation of funds". Then, select all "crime of misappropriation of funds" cases from the historical case database to form a candidate set, and sequentially select the cases in the candidate set to compare the similarity with case 1, taking case 2 in the candidate set as an example.

Step 2. Represent the event sequence by the following methods:

1) Use the information extraction method to extract the case elements in a structured manner. It was obtained that the defendant in case 1 was "salesman", and his attitude was "refunded" after the incident; the defendant's position in case 2 was "sales consultant", and his attitude was "refunded" after the incident. Corresponding time, key words, amount involved and purpose of the case: Case 1: April to July 2017: collection, project payment, 31,000 yuan, personal use, October 2017: collection, project payment, 58,000 RMB, personal use, December 2017: collection, project payment, 113,000 yuan, personal use; case 2: November 15, 2017: collection, car purchase, 130,500 yuan, own use, July 31, 2018: Collection, car purchase, 52,678 yuan, own use, August 4, 2018: collection, car purchase deposit, 20,000 yuan, own use.

2) According to the extraction results, the position of the defendant in Case 1 was independently reserved as "salesman", and the attitude of the defendant was "refunded" after the incident; the position of the defendant in Case 2 was "sales consultant", and the attitude of "refund" after the incident occurred.

3) As shown in Figure 2, the time series case elements are organized according to the temporal qualitative relationship. Events with no overlap or overlap in occurrence time are regarded as independent events and represented separately, and events are combined for other cases to obtain the event chain shown in Figure 3, including case elements such as time, trigger words, motives involved, and amount of money.

4) Use the word2vec training word vector method in gensim to quantify the case elements in the event chain.

5) Perform vector calculation according to the weight. For example, the event event _{in October 2017 in} case 1 = (collection, project payment, 58000, self-use), assuming that the initialization weight is weight=(0.3, 0.2, 0.2, 0.3), then the event event is calculated as _{10 in 2017} The vectorization of _month is expressed as vector _{October 2017} = (0.3e _{for collection} , 0.2e for _{project payment} , 0.2e for ₅₈₀₀₀ , 0.3e for _{own use} ), e _{element 1} represents the vector value of case element 1.

6) Connect the event representations at all times, that is, to obtain a vectorized sequential event chain.

The event sequence obtained in step 3 and step 2 is subjected to sequence distance measurement. Using the dynamic time warping method shown in Figure 4, for the event sequence EventSequence corresponding to Case 1, _{Case 1} = (vector _{April to July 2017} , vector _{October 2017} , vector _{December 2017} ) and events corresponding to Case 2 Sequence EventSequence _{case 2} = (vector _{on November 15, 2017} , vector _{on July 31, 2018, vector on August 4, 2018} ) for time point alignment and distance calculation, then in case one, the vector is _{from April to July 2017} Corresponds to the vector in case 2 _{on November 15, 2017} , and so on. Get the distance calculation result of the event sequence Distance _{EventSequence} .

Step 4. Score the overall similarity of the sequences by the following methods:

1) Calculate the distance _topic similarity between the two cases. Using the LDA topic analysis model in gensim, input the legal documents of Case 1 and Case 2 to obtain the corresponding topic distribution. According to the topic probability distribution, the semantic distance of the two legal documents is calculated by the distance formula KL distance, that is, the topic similarity Distance _topic ;

2) Use the cosine distance to calculate the similarity Distance _position of the positions "salesperson" and "sales consultant" in the two cases, and the similarities Distance _attitude of "refund" and "refund" after the case;

3) Scoring based on the subject similarity, event sequence distance, position and attitude similarity after the incident of the case

score=α ₁ Distance _topic +α ₂ Distance _{EventSequence} +α ₃ Distance _position +α ₄ Distance _attitude .

Another embodiment of the present invention also provides a method for pushing legal similar cases, which calculates the similarity between the target legal case and the historical cases in the candidate set by using the similarity discrimination method provided in the foregoing embodiment, and then calculates the similarity according to the similarity score from high to high. Sort the historical cases in the candidate set in the lowest order, extract the top M historical cases and push them. The value of M can be set according to the actual situation, for example, M=10.

From the description of the above embodiments, those skilled in the art can clearly understand that the above embodiments can be implemented by software or by means of software plus a necessary general hardware platform. Based on this understanding, the technical solutions of the above embodiments may be embodied in the form of software products, and the software products may be stored in a non-volatile storage medium (which may be CD-ROM, U disk, mobile hard disk, etc.), including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present invention.

The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited to this. Substitutions should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (5)

1. A legal class similarity discrimination method is characterized by comprising the following steps:

classifying target legal cases, and extracting historical cases of the same category from a historical case database to form a candidate set according to the obtained case category;

performing event sequence representation on the target legal case and each similar historical case in the candidate set;

calculating the distance between the event sequence corresponding to the target legal case and the event sequence corresponding to each historical case in the candidate set according to the event sequence measurement model;

calculating the similarity between the target legal case and the candidate centralized historical case based on the distance of the event sequence and the scoring function;

wherein the event sequence representation of the target legal case and each similar historical case in the candidate set comprises:

for any legal case, extracting case elements in the legal case document by using an information extraction technology, wherein the case elements at least comprise: the positions of the defenders, the postscript attitudes, the time, the core words and the involved amount of the defenders; the core words are extracted by a dependency syntactic analysis method, namely, sentences in the legal case document are divided according to periods, exclamation marks and question marks, dependency analysis is carried out on the sentence division results, and corresponding core words are obtained;

two case elements of the position of the announced person and the after-case attitude of the announced person are independently reserved;

representing the time sequence relation among the events of the rest case elements according to different time nodes, regarding the events without crossing and overlapping in occurrence time as independent events, respectively representing, and combining the events of the rest cases to form a case scenario event chain;

numerically expressing case elements of the events occurring at the time i to obtain the events _i ＝(e _i1 ,e _i2 ,…,e _in ) Wherein n is the number of case elements of the event occurring at the time i;

initialization weight vector weight ═ w ₁ ,w ₂ ,…,w _n ) Wherein, in the step (A),

associating the weight vector with the event _i Multiplying each element correspondingly to obtain a final representation vector of the time i event _i ＝(w ₁ e _i1 ,w ₂ e _i2 ,…,w _n e _in )；

Connecting the event representations of all the time to obtain a sequence event chain (vector) of vectorized representation ₁ ,vector ₂ ,…,vector _m ) Vector therein _i And m is the total number of independent events and combined events extracted from the legal case document.

2. The legal class similarity determination method of claim 1, wherein the classifying the target legal case comprises: and classifying the target legal case by using a deep learning text classification algorithm in natural language processing as a classifier to obtain a case category corresponding to the target legal case.

3. The method as claimed in claim 1, wherein the calculating the distance between the event sequence corresponding to the target legal case and the event sequence corresponding to each historical case in the candidate set according to the event sequence metric model comprises:

matching an event sequence corresponding to a target legal case with each event sequence corresponding to a historical case by adopting a Dynamic Time Warping (DTW) method from a starting point, calculating the Distance between the corresponding two points when each event sequence reaches one point, accumulating the distances of all points which pass before, and finally selecting the minimum accumulated Distance as the Distance of the event sequence _{EventSequence} 。

4. The legal class similarity discrimination method of claim 1, wherein calculating the similarity between the target legal case and the candidate centralized history case based on the distance of the event sequence and the scoring function comprises:

performing topic analysis on the target legal case and the candidate centralized history case through a topic analysis model to obtain corresponding topic probability distribution, and calculating the semantic Distance between the target legal case and the candidate centralized history case as topic similarity Distance according to the topic probability distribution _topic ；

When the event sequence is expressed, two case elements of the position of the advertiser and the post-event attitude of the advertiser are respectively extracted from the target legal case and the candidate centralized history case, the similarity of the position of the advertiser and the post-event attitude of the advertiser in the target legal case and the candidate centralized history case is calculated by utilizing the cosine Distance, and the similarity is respectively marked as Distance _position 、Distance _attitude ；

Event sequence distance between target legal case and candidate centralized historical caseIs recorded as Distance _{EventSequence} ；

Calculating the similarity of the target legal case and the candidate centralized historical case by using the following formula:

score＝α ₁ Distance _topic +α ₂ Distance _{EventSequence} +α ₃ Distance _position +α ₄ Disyance _attitude

wherein alpha is ₁ 、α ₂ 、α ₃ And alpha ₄ Are all weighted, α ₁ +α ₂ +α ₃ +α ₄ ＝1。

5. A legal class pushing method is characterized by comprising the following steps: the method of any one of claims 1 to 4 is utilized to calculate the similarity between the target legal case and the historical cases in the candidate set, then the historical cases in the candidate set are ranked according to the similarity score from high to low, and M historical cases with the top rank are extracted for pushing.

Images