A smart legal query method based on a multi-round pruning Skyline algorithm

Abstract

The application discloses a kind of wisdom legal inquiry method based on multi-wheel pruning Skyline algorithm, comprising: local equipment of legal department obtains local legal case according to the query command issued by central server;The local legal case obtained is uploaded to the central server in encrypted form, the central server carries out integration analysis, and the legal case with higher comprehensive document similarity is sent to local equipment.The scheduling strategy module of the central server obtains the physical node of local legal department or the corresponding legal department priority information according to the attributes such as name, gender, native place, address, case name, case-cracking event, capture event and report event in legal case.The application can make legal information retrieval personnel free from reading a large number of cases, save a lot of time and manpower and material resources, especially provides technical guarantee for finding similar cases, case-cracking clues and inducing crime trend.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent legal technology, and in particular to an intelligent legal query method based on a multi-round pruning Skyline algorithm. BACKGROUND

[0002] In recent years, with the joint development of software and information technology such as big data, computing power, operation logic, and core algorithms, the third wave of artificial intelligence is gradually coming, and the artificial intelligence technology in the legal field is also rapidly progressing. How to obtain the required information from the massive legal cases, judicial documents, and legal templates has become a problem to be solved in today's era. Learning-to-rank (LTR) has achieved great success in the field of information retrieval, and this success is particularly evident in commercial search engines such as Google and Bing, which have a large amount of user data resources. The effectiveness of the learning-to-rank algorithm cannot be achieved without a large amount of training data, and the data of search engine giants is often generated by the frequent interaction between the search engine and its tens of millions or even hundreds of millions of users every day. However, many small and medium-sized enterprises also have the need to build search engines or search systems, such as enterprise search or email search, but they do not have the strength to collect large-scale user data on their own. The data silo problem between small and medium-sized enterprises and enterprises, and even between departments of enterprises, leads to the scattered storage of a large amount of data, which is difficult to be connected due to reasons such as enterprise confidentiality protection and complex administrative procedures. At the same time, with the introduction of more and more data privacy protection regulations such as the European Union's General Data Protection Regulation (GDPR), even if the enterprise is willing to do so, it cannot freely exchange or share data containing user privacy as before. Various unfavorable factors make it difficult for small and medium-sized enterprises to build their own effective search systems with the support of powerful learning-to-rank algorithms and big data.

[0003] Legal case documents are criminal information recorded by legal staff during interrogation of criminal personnel, and the document information structure is usually poor, and the detective needs to spend a lot of time searching for similar cases. In the face of the millions of criminal cases growing every year, how to mine similar cases when breaking the case plays an important role in subsequent search for clues. SUMMARY

[0004] In view of the problems of slow legal case retrieval efficiency, low retrieval accuracy and legal security, the application provides a smart legal query method based on multi-round pruning Skyline algorithm in a federated learning environment, which can enable legal information retrieval personnel to avoid reading a large number of cases, saving a lot of time and manpower, and especially providing technical guarantee for finding similar cases, clues and crime trends.

[0005] To achieve the above object, the application provides a smart legal query method based on multi-round pruning Skyline algorithm, which comprises: a local device of a legal department acquires local legal cases according to a query command issued by a center server; the acquired local legal cases are uploaded to the center server in an encrypted form, and the center server performs integration analysis to obtain legal cases with high comprehensive document similarity and sends them to the local device.

[0006] Further, the scheduling strategy module of the center server obtains the physical node of the local legal department or the corresponding legal department priority information according to the attributes of the legal cases such as name, gender, place of origin, address, case name, case solving event, arrest event and report event.

[0007] Further, each local device in the task queue of the center server forms a node-task mapping table, and the scheduler scores the task mapping table on each node and selects the physical node with the highest score to allocate the query task.

[0008] Further, the local device performs inverted index pruning and then Skyline pruning on the local legal cases according to the query command issued by the center server to obtain legal cases meeting the query requirements.

[0009] Further, the center server aggregates the legal cases returned by the plurality of local devices after receiving them, and needs to consider the problems of query efficiency, data security, privacy and the like in the aggregation process.

[0010] Further, the search data function S(w) of the center server and the local device is:

[0011]

[0012] Wherein, m is the total number of local devices of legal departments, N is the total amount of data of all local devices, N i is the data amount of the i-th local device, S i (w) is the search function of the i-th local device.

[0013] Further, the index pruning is specifically scanning all legal cases, calculating the document importance of each legal case, and only taking the first k words when pruning each legal case; the Skyline pruning is mapping the results after pruning the legal cases to a self-defined grid, cutting off the dominated grid according to the grid domination relationship; the domination relationship between legal cases is compared to obtain a query set, and the query set is returned to the center server in an encrypted manner.

[0014] The index pruning is to determine the index length of each legal document according to the document importance; the present application measures the whole document by words, removes unimportant contents from the document index, so as to reduce the index length and greatly reduce the storage demand.

[0015] Further, the importance of a legal case can be measured by the link relationship and the document dispersion value.

[0016] The more links the legal case has, the higher the importance is; the LegalRank value reflects the number of links of the case.

[0017]

[0018] In (p i ) is the total sum of links to the legal document p i , Out (p j ) is the total sum of links out of the legal document p j , S is the total number of legal documents, lp j is the index length of the legal document p j , avg (p j ) is the average length of the document index of the whole document set, and k1, λ and σ are adjustment parameters.

[0019] For the content of the document, the similarity of a single document to the whole document set can be used to represent. Generally speaking, the more similar a document is to the whole document set, the more it represents the whole document, and the more important the document is in the document set. The document similarity can be calculated by the inverse document coefficient and the dispersion value, wherein the document dispersion value is an index for measuring the difference between documents, and the greater the value is, the greater the difference between documents is.

[0020] The document dispersion value is an index for measuring the difference between legal cases, and the greater the value is, the greater the difference between cases is.

[0021]

[0022] In the formula, T is the whole document term set, P is the language model of the legal document term d, and Q is the language model of the legal document set D.

[0023] The document similarity is as follows:

[0024]

[0025] In the formula, |D| is the total number of files owned by the legal department, j: P i ∈Q j represents the total number of files containing the word P i , n ij is the number of times the word appears in the document, and ∑ t∈T n kj represents the total sum of all words in the document.

[0026]

[0027] The present application uses document link relationships and document content to represent document importance:

[0028] P(d) = αP(LegalRank) + βP(LKL)

[0029] In the formula, P(LegalRank) is the converted value of the number of document links, P(LKL) is the converted value of the document dispersion, and α and β are proportional parameters, usually α = 0.5 and β = 0.5. The two methods of calculating the importance of legal documents are combined to obtain the comprehensive document importance. According to the comprehensive document importance, the documents with low comprehensive document importance are cut off.

[0030] When querying judicial documents, legal agency staff often need to spend a lot of time. If an index file is established for each legal case according to the importance, the retrieval efficiency can be greatly improved.

[0031] As a further improvement, the Skyline pruning is specific: there is a d-dimensional attribute space data S that measures the importance of documents, which includes the number of legal document link relationships LegalRank value, document dispersion LKL divergence, etc. P is a data set on the attribute space data S, and each document importance p ∈ P is represented as {x1, x2, …, x d}; assuming that the space grid width is λ, any document attribute in the data set is calculated by grid coordinates, that is Any grid is represented as Intkey j , and in the space grid, it is represented as (Intkey j x1, Untkey j x2, …, Intkey j x d ). Each attribute of the data set p measuring the importance of documents is mapped to the space grid.

[0032] As a further improvement, let Seed be the set of Skyline points, r be the candidate Skyline object in the grid, and SB is any point in the grid, U is a set of space points, t is any point in U; if r is in Seed, then r must be in U; if t dominates r, let T be a set of grid area, v is any point in T; W is a set of adjacent grids of T; there are two possibilities here, the first: W is located directly below or to the left of the grid set T, then The value of v point on some latitude is greater than the element in W, so v is not dominated by any object in W; the second: W is located in the lower left of the grid set T, then The value of v point on all latitudes is greater than the element in W, obviously any element in W is not dominated by v; in summary, if the value of v in at least one dimension is greater than the element in W, then v can dominate any element in the set W, and further prune the points in the W set.

[0033] Each grid can be calculated by coordinates, and the data in each grid is stored in the form of a key-value pair of <key, list[values]>, the key corresponds to the identifier of the grid, and the value in the list is the data point in the grid. The grid domination relationship is used to prune the points with low LegalRank values and high LKL divergence. The value of the grid width lambda depends on the sparsity of the spatial distribution of the document attribute values, if the document attribute value distribution is relatively dispersed, the lambda value can be appropriately large, and if the document attribute value distribution is relatively dense, the lambda value should be appropriately small.

[0034] The above technical scheme adopted by the present application has the advantages compared with the prior art:

[0035] (1) Similar legal cases can be mined when solving cases

[0036] In the face of millions of cases growing every year, querying similar cases plays an important role in subsequent clues for solving cases. The present application can use a multi-round pruning algorithm to prune the dominated cases in a large amount of data cases. This avoids spending a lot of time reading cases, and provides a technical guarantee for finding similar cases, solving clues, and summarizing crime trends.

[0037] (2) Ensure the security of legal data

[0038] Legal data as private data needs to be encrypted during data transmission. Federated learning can help different legal departments search and use data while meeting local data security requirements. Each legal department migrates data to the local user without uploading data to the center server, and interacts with the center server in an encrypted manner, effectively ensuring the security of local data. DETAILED DESCRIPTION ​

[0039] Figure 1 For federated learning query framework graph;

[0040] Figure 2 For multi-round pruning processing graph;

[0041] Figure 3 For legal document information graph;

[0042] Figure 4 For network pruning process graph. DETAILED DESCRIPTION

[0043] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application, that is, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments.

[0044] Therefore, the detailed description of the embodiments of the present application provided in the drawings below is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0045] Embodiment 1

[0046] As shown in Figures 1-4 The improved Skyline pruning algorithm in this embodiment runs in a federated learning environment, which can effectively solve the legal data security problem. At the same time, by taking advantage of the characteristics of federated learning in supporting non-independent and identically distributed data and efficient communication, the search efficiency is improved to some extent. Unlike traditional search methods, the structured legal documents are constructed, and the document similarity is measured from the aspects of link relationship and document content, and the documents with low comprehensive similarity are pruned. The multi-round Skyline pruning algorithm based on semantic requirements maps the document similarity to the grid, and further prunes the dominated documents, thereby improving the query efficiency of legal cases and positioning the relevant cases more scientifically and efficiently.

[0047] The present application provides a smart legal query method based on multi-round pruning Skyline algorithm, comprising: a local device of a legal department acquires local legal cases according to a query command issued by a center server; the acquired local legal cases are uploaded to the center server in an encrypted form, and the center server performs integration analysis to obtain legal cases with high comprehensive document similarity and sends them to the local device. The local device performs inverted index pruning and then Skyline pruning on the local legal cases according to the query command issued by the center server, and acquires legal cases meeting the query requirements.

[0048] The foregoing description of specific exemplary embodiments of the application has been presented for the purposes of illustration and description. It is not intended to be a limitation on the broad concepts of the application. Obviously, many modifications and variations of the specific exemplary embodiments described herein are possible in light of this disclosure, and it is intended that the scope of the application be limited only by the claims appended hereto and their equivalents. Examples of selected embodiments were chosen and described in order to explain the principles of the application and its practical application. Those skilled in the art will readily appreciate that various modifications and changes can be made to the specific exemplary embodiments without departing from the scope of the application. The scope of the application is defined only and exactly by the claims.

Claims

1. A smart legal query method based on a multi-round pruning Skyline algorithm, characterized in that, Comprise: The local device of the legal department obtains the local legal case according to the query command issued by the center server; the local legal case obtained is uploaded to the center server in an encrypted form, the center server performs integrated analysis, and the legal case with high comprehensive document importance is sent to the local device; The local device obtains the legal case meeting the query requirement by pruning the inverted index and then pruning the Skyline according to the query command issued by the center server; The index pruning is specifically scanning all legal cases, calculating the document importance of each legal case, and only taking the first k words when pruning each legal case; the Skyline pruning is mapping the result after pruning the legal case to a self-defined grid, pruning the dominated grid according to the grid domination relationship; comparing the domination relationship between legal cases to obtain a query set, and returning to the center server in an encrypted manner; The more the legal cases are linked, the higher the importance is; the LegalRank value reflects the number of links of the case: wherein, is the sum of the chain to the legal document , is the sum of the chain out of the legal document , is the sum of the legal document index length, is the average length of the document index over the entire document set, is a tuning parameter; The document divergence value is an index for measuring the difference between legal cases, and the larger the value is, the greater the difference between cases is: In the formula, T is the set of all document words; The document link relationship and the document content are used to represent the document importance: wherein is a converted value of the number of document links, is a converted value of the document dispersion, and , is a proportional parameter. 2.The intelligent legal query method based on the multi-round pruning Skyline algorithm of claim 1, wherein, The scheduling strategy module of the center server obtains the physical node of the local legal department or the corresponding legal department priority information according to the name, gender, place of origin, address, case name, case solving event, arrest event and report event attributes in the legal case. 3.The intelligent legal query method based on the multi-round pruning Skyline algorithm of claim 1, wherein, Each local device in the task queue of the center server forms a node-task mapping table, the scheduler scores the task mapping table on each node, and selects the physical node with the highest score to allocate the query task. 4.The intelligent legal query method based on the multi-round pruning Skyline algorithm of claim 1, wherein, The center server aggregates the legal cases returned by multiple local devices after receiving them. 5.The intelligent legal query method based on the multi-round pruning Skyline algorithm of claim 1, wherein, Let the query task on the local device of the legal department be a vector The center server and the local device search data function is: Where m is the total number of local devices in the legal department. It is the sum of data from all local devices. It is the first Data volume of a local device It is the first A local device search data function. 6.The intelligent legal query method based on the multi-round pruning Skyline algorithm of claim 1, wherein, The Skyline pruning specifically involves: having a d-dimensional attribute space data S that measures document importance, including the number of legal document link relationships. Values, document differences Divergence, where P is the dataset on the attribute space data S, represents the importance of each document. Represented as Assume the spatial grid width is Then, any document attribute in the dataset is calculated using grid coordinates, i.e. Any grid is represented as Then in the spatial grid it is represented as Each attribute of the document importance dataset p is mapped to the spatial grid.​ 7. The intelligent legal query method based on the multi-round pruning Skyline algorithm according to claim 6, characterized in that, Let Seed be the set of Skyline points, r is a candidate Skyline object in the grid, is an arbitrary point in the grid, U is the set of spatial points, t is an arbitrary point in U; Given r∈Seed, then r∈U must be true; If , then ∃t∈U, t dominates r, let T be a set of region grids, v is an arbitrary point in T; W is a set of adjacent grids of T; Here there are two possibilities, the first: W is located directly below or to the left of the grid set T, then ∃v∈T, v is greater than any object in W in some latitude, so v is not dominated by any object in W; The second: W is located in the lower left of the grid set T, then ∃v∈T, v is greater than any element in W in all latitudes, obviously any element in W is not dominated by v; In summary, v is at least greater than the elements in W in one dimension, then v can dominate any element in the set W, and further prune the points in the set W.

Images