CN113553856B - Deep neural network-based dispute focus identification method - Google Patents
Deep neural network-based dispute focus identification method Download PDFInfo
- Publication number
- CN113553856B CN113553856B CN202110665262.XA CN202110665262A CN113553856B CN 113553856 B CN113553856 B CN 113553856B CN 202110665262 A CN202110665262 A CN 202110665262A CN 113553856 B CN113553856 B CN 113553856B
- Authority
- CN
- China
- Prior art keywords
- contradiction
- sentence
- layer
- dispute focus
- dispute
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 55
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 28
- 230000014509 gene expression Effects 0.000 claims abstract description 25
- 238000012549 training Methods 0.000 claims abstract description 21
- 230000008569 process Effects 0.000 claims abstract description 18
- 238000010276 construction Methods 0.000 claims abstract description 7
- 238000002372 labelling Methods 0.000 claims abstract description 5
- 238000007781 pre-processing Methods 0.000 claims abstract description 5
- 238000004140 cleaning Methods 0.000 claims abstract description 4
- 239000013598 vector Substances 0.000 claims description 93
- 238000001514 detection method Methods 0.000 claims description 35
- 238000004364 calculation method Methods 0.000 claims description 11
- 238000000605 extraction Methods 0.000 claims description 11
- 239000011159 matrix material Substances 0.000 claims description 8
- 238000011176 pooling Methods 0.000 claims description 8
- 238000012937 correction Methods 0.000 claims description 5
- 230000004927 fusion Effects 0.000 claims description 5
- 238000012935 Averaging Methods 0.000 claims description 3
- 230000007246 mechanism Effects 0.000 claims description 3
- 230000002068 genetic effect Effects 0.000 claims description 2
- 238000004458 analytical method Methods 0.000 abstract description 5
- 230000006870 function Effects 0.000 description 8
- 238000003058 natural language processing Methods 0.000 description 7
- 239000000463 material Substances 0.000 description 6
- 206010039203 Road traffic accident Diseases 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 230000004913 activation Effects 0.000 description 3
- 238000013145 classification model Methods 0.000 description 3
- 230000006378 damage Effects 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 239000013604 expression vector Substances 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 238000011835 investigation Methods 0.000 description 2
- 230000003340 mental effect Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 240000007643 Phytolacca americana Species 0.000 description 1
- 235000009074 Phytolacca americana Nutrition 0.000 description 1
- 229910052770 Uranium Inorganic materials 0.000 description 1
- 208000027418 Wounds and injury Diseases 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- PCHJSUWPFVWCPO-UHFFFAOYSA-N gold Chemical compound [Au] PCHJSUWPFVWCPO-UHFFFAOYSA-N 0.000 description 1
- 239000010931 gold Substances 0.000 description 1
- 229910052737 gold Inorganic materials 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 208000014674 injury Diseases 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 239000002184 metal Substances 0.000 description 1
- 229910052751 metal Inorganic materials 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 239000004575 stone Substances 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/18—Legal services
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Business, Economics & Management (AREA)
- Life Sciences & Earth Sciences (AREA)
- Tourism & Hospitality (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Technology Law (AREA)
- Financial Or Insurance-Related Operations Such As Payment And Settlement (AREA)
- Machine Translation (AREA)
Abstract
The invention relates to a dispute focus identification method based on a deep neural network, which comprises the following steps: acquiring a large number of referee documents, preliminarily cleaning the referee documents by using a regular expression method, extracting original appeal and referee dialect expression, carrying out dispute focus category labeling expressed in pairs by a legal expert, and completing construction of a dispute focus library; performing sentence-level and paragraph-level training on the model based on the deep neural network by using a dispute focus library to obtain a dispute focus recognizer; and (4) preprocessing the paired expressions of the two sides of the dispute to be identified to be used as input, transmitting the preprocessed paired expressions into the dispute focus recognizer obtained by training in the step two, and outputting the category to which the dispute focus belongs by the dispute focus recognizer. The process of the method can be automatically learned and finished through a machine, so that the labor cost is saved; the identification accuracy is improved; the method can better identify the dispute focus and provide support for rapid and accurate analysis of the key information of cases by judges, inspectors and other judicial personnel.
Description
Technical Field
The invention relates to the technical field of dispute focus identification, in particular to a dispute focus identification method based on a deep neural network.
Background
With the proposal of the intelligent judicial concept and the construction of the judicial public platform, the informatization of the judicial field is effectively promoted. Wisdom judicial is mainly based on Natural Language Processing (NLP) to process the data in the judicial field, because most resources in the judicial field are displayed in text form, such as judgment documents, legal opinions, etc., how to apply the NLP technology to the judicial field better and make the NLP technology benefit is gradually gaining wide attention in the chemical industry and the industrial industry.
There is a lot of high quality text data in the judicial domain, characterized by a longer space than other domain data, legal professionals often consider how to solve the task by rule-based and symbol-based methods, while NLP researchers are more concerned about data-driven and embedded methods. To facilitate the development of judicial artificial intelligence, a great deal of effort has been made by many researchers to early research the use of manually made rules or functions all the time due to the computational limitations at the time. In recent years, with the rapid development of deep learning, researchers have begun to apply deep learning techniques to the judicial field.
The dispute focus refers to the core dispute of both sides in evidence, facts and law application in litigation cases, and has high relevance to the property and case law of a case besides the fact, evidence and other factors of an individual case, so that the dispute focus of a specific case law case has certain characteristics, and can be classified into a limited number of types, thereby constructing a dispute focus library.
The dispute focus is identified as a basic task in the judicial field, the current research has not realized a great breakthrough, and how to apply the intelligent method to the problem and finally realize the improvement of the case trial efficiency becomes a research task with important significance. NLP has been applied in the field of law, but research on identification of the focus of dispute is still insufficient and remains to be studied in depth. In addition, the forensic material has unique structural characteristics, and can identify various dispute focuses, so that informatization process of judicial fields can be further developed.
Disclosure of Invention
The invention provides a dispute focus recognition method based on a deep neural network, aiming at solving the defects of dispute focus recognition in the prior art and aiming at the characteristics that a dissatisfied material has a longer text and a different length and is difficult to understand deep semantics by a model. The invention provides a dispute focus recognition method based on a deep neural network, which aims to automatically analyze dispute focuses in a case dispute process through an intelligent technology, provide support for rapid and accurate analysis of key information of cases by a judge, a scout and other judicial personnel, realize deep application of an NLP technology in the judicial field, and provide scientific and technical support for solving intelligent processing and analysis problems related to judicial services.
In order to realize the purpose, the invention adopts the following technical scheme:
a dispute focus identification method based on a deep neural network comprises the following steps:
the method comprises the following steps: acquiring a large number of referee documents, preliminarily cleaning the referee documents by using a regular expression method, extracting original appellations and compelling allegories from the referee documents, carrying out dispute focus category labeling expressed in pairs by a legal expert, and completing construction of a dispute focus library;
step two: performing sentence-level and paragraph-level training on the model based on the deep neural network by using the dispute focus library to obtain a dispute focus recognizer; the model based on the deep neural network comprises an input layer, a model layer, a correction layer and an output layer, and the second step comprises the following steps:
step two, firstly: the input layer takes the complaint party text, the dialect text and the complaint sentence obtained after the complaint text is preprocessed and matched as the input text, converts the input text into a tensor which is easy to understand by a model, and inputs the tensor into the embedding layer of the model layer for calculating the sentence-level feature representation and the paragraph-level feature representation by the model layer;
step two: on a model layer, receiving an appeal text, a dialect text and an appeal text which are obtained after matching, inputting a tensor of the appeal sentence and the dialect sentence, calculating sentence-level feature representation and paragraph-level feature representation, and further performing contradiction detection by adopting a contradiction detection model based on BERT-CBGA (belief-based genetic algorithm) for calculating contradiction detection feature representation, wherein the contradiction detection feature representation comprises sentence-level contradiction detection feature representation; carrying out contradiction classification by adopting a model with a convolution structure, and calculating contradiction classification characteristic representation which comprises sentence-level contradiction classification characteristic representation and paragraph-level contradiction classification characteristic representation; carrying out contradiction detection and classification at a sentence level, carrying out contradiction detection on the input tensor of the apperceive sentences obtained after matching of the apperceive texts to obtain vectors of whether the apperceive sentence pairs have contradictions, carrying out contradiction classification to obtain contradiction classification vectors of each sentence pair, and carrying out contradiction classification at a paragraph level to obtain contradiction classification vectors;
step two and step three: in a correction layer, correcting the contradiction classification vectors by using whether each sentence pair has a contradiction relation vector and the contradiction classification vectors in a cross-product mode to obtain contradiction classification vector representation of each sentence pair, extracting the corrected result according to the maximum value of each category of the sentence pair attributive complaint texts to obtain sentence-level whole complaint text contradiction classification vectors, and performing averaging calculation on the sentence-level complaint text contradiction classification results to obtain final dispute focus vector representation of the complaint texts;
step two, four: the output layer outputs the dispute focus vector of the final dispute text by using a dispute focus category label to obtain a dispute focus identifier;
step three: and preprocessing the paired expressions of the two sides of the dispute to be identified to be used as input, transmitting the preprocessed paired expressions of the two sides of the dispute to the dispute focus recognizer obtained by training in the step two, and outputting the category to which the dispute focus of the paired expressions of the two sides of the dispute to be identified belongs by the dispute focus recognizer.
Compared with the prior art, the invention has the following beneficial effects:
(1) the process of the dispute focus identification method based on the deep neural network can be completed through automatic learning of a machine, so that the labor cost is saved;
(2) the method can enable the model to better understand the deep semantics of the text, and improve the identification accuracy;
(3) the dispute focus recognition method based on the deep neural network can better perform dispute focus recognition and provides support for rapid and accurate analysis of key information of cases by judges, inspectors and other judicial personnel.
Drawings
FIG. 1 is a general flow chart of a deep neural network-based dispute focus identification method according to the present invention;
FIG. 2 is a diagram of a model structure of a deep neural network-based dispute focus identification method according to the present invention;
FIG. 3 is a diagram of a contradiction detection model based on BERT-CBGA;
FIG. 4 is an example of an appealing text;
fig. 5 is an example of a discourse sentence pair.
Detailed Description
The technical solution of the present invention will be described in detail with reference to the accompanying drawings and preferred embodiments.
In one embodiment, as shown in fig. 1, the present invention provides a deep neural network-based dispute focus identification method, which includes the following steps:
the method comprises the following steps: obtaining a large number of referee documents, preliminarily cleaning the referee documents by using a regular expression method, extracting original appeal and referee dialect expressions, carrying out dispute focus category labeling expressed in pairs by a legal expert, and completing construction of a dispute focus library.
In the first step, the complaint material is submitted by both complaints in the litigation process and mainly comprises complaints, answers, files and referee documents, and due to the reason that the complaints and the answers are not suitable for disclosure, a large number of referee documents are obtained through the Internet in the training stage, the referee documents are preliminarily cleaned by using a regular expression method, original complaints and the referee terms are extracted and expressed, and dispute focus categories expressed in pairs are labeled by a legal expert to complete construction of a dispute focus library.
Further, the category to which the dispute focus belongs is expressed into a binary vector, namely the vector value of the dispute focus category to which the dispute focus belongs is expressed in pairs by the dispute parties is 1, and the other vector values are 0, and the binary vector is used as a label vector of the dispute focus in pairs, so that the dispute focus automatic identification problem in the dispute material is converted into a text matching problem of multi-label classification.
Step two: and (4) carrying out sentence-level and paragraph-level training on the model based on the deep neural network by using a dispute focus library to obtain a dispute focus recognizer.
And training a model based on the deep neural network by using the dispute focus library obtained in the step one in combination with semantic information at sentence level and paragraph level to obtain a corresponding dispute focus recognizer and exporting the dispute focus recognizer.
Further, as shown in fig. 2, the deep neural network-based model performs sentence-level and paragraph-level training simultaneously, performs spear shield detection and classification at the sentence level, and performs contradiction classification at the paragraph level, and the deep neural network-based model mainly includes an input layer, a model layer, a modification layer, and an output layer, and the second step specifically includes the following steps:
step two, firstly: the input layer of the model receives the complaint text L (Xsc), the dialect text L (Xbc), and the complaint sentence Xsc obtained by matching the complaint text i (i 1.., m), dialect Xbc j (j 1., n) as input and converted into tensors easy to understand by the model, and sentence-level feature representations and paragraph-level feature representations are calculated by the model layer, wherein the sentence-level feature representations comprise an original tolling sentence partial feature representation, an original tolling sentence context feature representation, an original tolling sentence, paired embedding of a tolling sentence, a tolling sentence partial feature representation and a tolling sentence context feature representation, and the paragraph-level feature representations comprise an original tolling paragraph feature representation and a tolling sentence feature representation.
Step two: and in the model layer, a conflict detection model based on BERT-CBGA is adopted for conflict detection and is used for calculating conflict detection feature representation, and the conflict detection feature representation comprises sentence-level conflict detection feature representation. A model with a convolution structure is adopted to carry out contradiction classification and used for calculating contradiction classification feature representation, the contradiction classification feature representation comprises sentence-level contradiction classification feature representation and paragraph-level contradiction classification feature representation, the sentence-level contradiction classification feature representation and the sentence-level contradiction classification feature representation are received in the embedding stage of the model layer and are matched to obtain an input tensor of the sentence-level contradiction and the sentence-level contradiction classification feature representation, the contradiction classification feature representation is further calculated to obtain sentence-level contradiction classification feature representation or paragraph-level contradiction classification feature representation, specifically, various existing contradictions are identified, and the contradictions are output to be classified vectors through convolution pooling by a multi-layer perceptron (MLP). Performing spear shield detection and classification at a sentence level, performing contradiction detection on the apperceive sentence pair input tensor obtained after the apperceive text matching to obtain whether vectors of contradictions exist in the apperceive sentence pair, and performing contradiction classification to obtain a contradiction classification vector of each sentence pair to obtain final sentence-level representation; and carrying out contradiction classification at a paragraph level to obtain a contradiction classification vector.
Step two and step three: in the model training process, whether each sentence pair has a contradiction relation vector M or not is judged in a correction layer s Contradiction-related classification vector M ds The contradiction classification vectors are corrected in a cross multiplication mode to obtain the contradiction classification vector representation M of each sentence pair SO Extracting the maximum value of each category of the attributive apperceive text according to sentences by the corrected result to obtain a sentence-level whole apperceive text contradiction classification vector M SA Then, the classification result M is classified through contradiction with the discourse text at the paragraph level dd And carrying out average calculation to obtain a final dispute focus vector representation M of the appealing text. The calculation process for obtaining the sentence-level whole dispute text contradiction classification vector is described by taking a first-type dispute focus as an example, wherein n is the number of sentence pairs in one dispute text:
M SO ＝M S ⊙M ds
M＝mean(M SA ,M dd )
step two, four: and the output layer represents the final dispute focus vector of the dispute text to be output by the dispute focus category label, and finally the trained dispute focus recognizer is obtained.
Further, the contradiction detection model based on BERT-CBGA comprises an embedding layer, a feature extraction layer, a feature fusion layer and an output layer, and the process of carrying out the contradiction detection by adopting the contradiction detection model based on BERT-CBGA in the second step specifically comprises the following steps:
step two, step one: the embedding layer comprises word embedding and sentence pair embedding for calculating the original announcement appellation and the alleged announcement appellation, the embedding layer is a process for converting input sentences into dense feature vectors, and the embedding layer of the model adopts word embedding. Each word in the input sentence is modeled as a word vector by a word vector matrix. And initializing the word vector randomly, and selecting a word vector untrainable mode in the training process to obtain word embedding of the sentence of the both sides of the complaint. Sentence pair embedding is realized through BERT, a Chinese edition BERT-Base and 12-layer transform are selected, final vector calculation is carried out through the output of the second last layer and input _ mask input, the input _ mask is a mark during input and indicates whether the input at the position is effective text or not, sentences can be processed into the same length during input processing, the length which is not reached can be filled, the mark of a filling position is 0 at the moment, and the mark of a real input position is 1. Firstly, dimension expansion is carried out on the input _ mask to reach the dimension (including sentence length dimension and word vector dimension) which is the same as the dimension of the output of the second last layer, then the two dimensions are multiplied to enable the invalid position value to be 0, and finally the obtained value is added and divided by the sum of the input _ mask in the sentence length dimension to obtain the sentence pair vector.
Step two and step two: the feature extraction layer of the model is divided into two steps, the local feature representation of the original notice appellation and the defended appellation is calculated through convolution operation of independent sentences, pooling output feature vectors are obtained, features of the original notice appellation and the defended appellation with context information are calculated through the BiGRU, and the features are further extracted through an attention mechanism, and the attention output feature vectors are obtained. Two feature extraction methods used by the model will be described in detail below.
In text processing, the convolution kernel of the convolution layer only moves along the longitudinal direction, because the input vector after word embedding is received, the longitudinal direction is the sentence length, the transverse direction is the word vector dimension, in order to avoid changing the meaning of a word, so as to ensure that the whole word vector dimension of a word is operated when the convolution kernel moves each time, the pooling strategy selects the maximum pooling, receives the vector result passing through the convolution layer, extracts the maximum value of the column vector, and calculates as follows:
characteristic c i Is a filter
Is applied to one
Word window generation, where k is the sentence length, h is the window size, f is a non-linear function,
is a bias term, and the filter applied to such a window of words embedded in vector emb (x) produces a feature map c ═ c 1 ,c 2 ,...,c (k-h+1) ]Wherein
Using BiGRU to extract deep-level features of input text, and inputting at t-th time step
For example, gate z is updated in the GRU t Reset gate r t The current memory content
Final memory content h t The specific calculation method is as follows:
wherein sigma is sigmoid activation function, W (z) 、U (z) To update the gate weight matrix, W (r) 、U (r) To reset the gate weight matrix, W, U is the weight matrix of the candidate hidden states, b z 、b r Is an offset term, h t And is output for the current moment.
Input at the t-th time step in the BiGRU
For example, the resulting output h t The calculation method is as follows,
for the stitching operation:
the attention layer receives the k-dimensional output vector H ═ H of the BiGRU 1 ,h 2 ,...,h k ]And calculates the weight to be assigned by the following calculation methodThe probability weight is distributed to the corresponding vector, so that the text features are further extracted, the key information is more prominent, and the ith dimension vector H of the H is used i For example, the calculation process is as follows:
u i ＝tanh(W a h i +b)
wherein W a For the attention weight matrix, b is the bias term, u i For each word-to-sentence importance, u w For random initialization and training the resulting context vector, A is the output obtained through the attention layer.
Step two, step three: the fusion layer outputs the pooled output feature vector C obtained by the feature extraction layer sc 、C bc And attention output feature vector A sc 、A bc Pair of features Emb (SP) with BERT sentence X ) And splicing and fusing to obtain final characteristics, namely contradiction detection characteristics, which are used for inputting the multilayer perceptron, wherein the calculation formula is as follows:
step two, step four: and in order to improve the performance and generalization capability of the model, the output layer transmits the result to a full connection layer activated by Sigmoid for classification through a layer of Dropout and a layer of full connection layer (Dense) activated by ReLU. After Dropout, the ReLU function is used as a full connection layer of the nonlinear activation function, the variable smaller than 0 is set to be 0, the value of the variable larger than or equal to 0 is kept unchanged, and finally the variable is mapped to the interval of (0,1) by the Sigmoid function to obtain the final contradiction detection characteristic representation.
Step three: and (4) preprocessing any dispute party pair expression to be identified to serve as input, transmitting the preprocessed dispute party pair expression to the dispute focus recognizer obtained by training in the step two, identifying by the dispute focus recognizer, and outputting the category to which the dispute focus of the dispute party pair expression to be identified belongs to so as to realize identification of the dispute focus of the dispute material.
The method for identifying the dispute focus based on the deep neural network has the following beneficial effects:
(1) the process of the dispute focus identification method based on the deep neural network can be completed through automatic learning of a machine, so that the labor cost is saved;
(2) the method can enable the model to better understand the deep semantics of the text, and improve the recognition accuracy;
(3) the dispute focus recognition method based on the deep neural network can better perform dispute focus recognition and provides support for rapid and accurate analysis of key information of cases by judges, inspectors and other judicial personnel.
The technical solution of the present invention is further described in detail with reference to the specific examples of the apperceive text.
1. The forensic material is submitted by both sides in the litigation process and mainly comprises an appeal form, a response form, a file and referee documents, wherein due to the reason that the appeal form and the response form are not suitable for disclosure, a large number of referee documents are obtained through the Internet in the training stage, the referee documents are preliminarily cleaned by using a regular expression method, original advices and quilt debate expressions are extracted, dispute focus categories expressed in pairs are labeled by a legal expert, and the construction of a dispute focus library is completed;
a. the Chinese referee document network is selected as a data source, because the Chinese referee document network is used as a first established judicial open platform, the document coverage is comprehensive, the quality is high, and in the open referee documents, the number of civil cases is more than 60 percent through investigation statistics, so that the Chinese referee document network is an important part for forming the referee cases. Because the times of retrieval and browsing of users are more due to livelihood, the document selects a civil first-pass referee document as a data base for dispute focus identification.
b. Writing rules, and extracting the original appeal and the defended dialect expression in the document. Specifically, the paragraph of the referee document takes the original report as the initial character, and contains keywords such as the title and the litigation request, and the character is the expression of the original report title; the paragraph of the referee document takes the "defendant" as the initial character and contains the keyword of "dialect", and the character of the paragraph is the dialect expression.
As shown in FIG. 4, the extracted original declaration such as "original declaration XX, XXX" makes a litigation request to the Hospital 1, the declaration of the body damage of the original declaration traffic accident with reimbursement totals 267557.998 yuan, and the defended insurer undertakes the payment responsibility in the insurance scope; 2. the litigation cost is borne by the notice. The fact and reason are that when a motorcycle is driven by XXX on a certain day in a certain month in a certain year, the motorcycle is driven from the place A to the place B, and when the motorcycle is driven for 30 minutes to the place C, the motorcycle collides with a small special passenger car which is informed of XXX driving of a certain type, so that XXX injury is killed after the motorcycle is sent to a hospital for rescue, and traffic accidents that the motorcycle and the passenger car are damaged are caused. On a certain day of the same month, the police officer makes a road traffic accident subscription book, and identifies XXX as the main responsibility of the accident and XXX as the secondary responsibility of the accident. According to the investigation, the reported XXX is a driver of a small special bus with a certain model, the reported XXX company is a legal owner of the bus, the bus is guaranteed to be protected by YYY company, the accident happens in the insurance period, the forced insurance and the liability insurance risk of the third party of 50 ten thousand special commercial buses are invested, and the exemption is not counted. 1, death compensation 28335 Yuan/year x20 years which is 566700 Yuan; 2. the life cost of the supported people XXX20660 yuan/year multiplied by 5 years divided by 6 people which are 17216.66 yuan; 3. handling accident error work cost 2000 Yuan; 4. the traffic fee is 1000 yuan; 5. 40000 yuan of mental soothing gold; 6. the vehicle loses 3610 yuan, and the cost is 630526.66 yuan. After the responsibility proportion is calculated, the quilt is also informed that 267557.998 yuan should be paid. In court trial, trawlers charge is increased by 250 yuan for original reports XX and XXX, and the total sum of litigation targets is changed to 267808 yuan. "
The extracted notice dialogs such as the notice XXX dialogs are 1, and no objection exists for the passing and responsibility division of the accident occurrence; 2. the owner of a small special passenger car with a certain model is XXX company, and is a driver when an accident occurs; the vehicle is informed that YYY company guarantees strong traffic insurance and 50 thousands of special commercial vehicles third party responsibility insurance (including exemption), and accidents occur in an insurance period; 3. the indemnity of the original claim agrees with the insurance company opinions. The company of the reported XXX is dialectical of 1, and has no objection to the passing of the accident and the division of the accident responsibility; 2. the company pays funeral 24787.5 yuan, funeral 17000 yuan and mortuary 2070 yuan for the deceased XX in the case, and totals 43857.5 yuan, and the funeral, the mortuary and the mortuary are required to be processed together in the case and directly paid to the company by an insurance company; 3. the vehicle insurance application condition is consistent with that of an insurance company, and XXX is a driver employed by the company; 4. the indemnity of the original claim agrees with the insurance company opinions. The reported YYY company has the dialectical that 1 is not objected to the process of accident occurrence and the division of accident responsibility; 2. a small-sized special passenger car of a certain model guarantees that the traffic intensity insurance and the responsibility insurance of the third party of a special commercial car are 50 ten thousand yuan (including no free claims) in our company, and accidents occur in the insurance period; 3. the company pays 1787.7 Yuan medical fee and 27212.5 Yuan funeral fee for the dead in the case, and the requirements are processed together in the case; 4. the death compensation is calculated for 20 years according to the rural standard for each compensation claim of the original report; the living cost of the supported people is calculated according to the rural standard for 5 years, and the supported people number is 6; the working cost of the traffic accident personnel is 900 yuan according to 100 yuan multiplied by 3 people multiplied by 3 times; the traffic fee is approved to be 500 yuan; the metal is pacified by the mental damage, and the dead has serious mistakes and is approved to be 20000 yuan; the vehicle repair fee is up to 3000 dollars. "
c. Labeling dispute focus labels by a legal expert, and recording the labeled original data set by the following steps of 8: 1: the proportion of 1 is divided into a training set, a verification set and a test set, wherein the training set is used for training the model based on the deep neural network.
2. Firstly, directly segmenting words of an original notice appeal and a notice appeal, sending the words into a contradiction classification model, then carrying out sentence level segmentation by taking a period number as a mark, segmenting the segmented sentences, and sending the segmented sentences into a contradiction detection model and a contradiction classification model.
For the contradiction detection model, word embedding and sentence pair embedding are firstly carried out through an embedding layer, each word in an input sentence is modeled into a word vector through a word vector matrix, the dimensionality is 300 dimensions, the sentence pair vector is obtained through BERT, and the dimensionality is 768 dimensions. And secondly, the word vector representation is sent into a feature extraction layer, local features are extracted through convolution operation on independent sentences, the features with context information are obtained through a BiGRU, and then the features are further extracted through an attention mechanism. And when local features are extracted, the operation on the whole word vector dimension of one word is ensured when the convolution kernel moves each time, the pooling strategy selects the maximum pooling, receives the vector result passing through the convolution layer, extracts the maximum value of the column vector, and outputs the vector dimension of 128 dimensions. During the context feature extraction, the BiGRU is used for extracting deep features of the input text, and the output vector of the BiGRU is sent to an attention layer to calculate the weight to be distributed, so that the key information is more prominent, and the dimension of the output vector is 256 dimensions. And then, splicing the features of 2 128-dimensional pooling output feature vectors, 2 256-dimensional attention output feature vectors and 768-dimensional BERT sentences of the feature extraction layer at the fusion layer to obtain a final feature with dimensions of 1536, and transmitting the final feature into the multilayer sensor. And finally, in order to improve the performance and generalization capability of the model, the multi-layer perceptron classifies the results through a layer of Dropout and a layer of Dense activated by ReLU and then transmits the results into a layer of Dense activated by Sigmoid (Dense is full connection). After Dropout, the ReLU function is used as a full connection layer of the nonlinear activation function, the variable smaller than 0 is set to be 0, the value of the variable larger than or equal to 0 is kept unchanged, and finally the variable is mapped to the interval of (0,1) by the Sigmoid function to obtain a final result. An example of a litigant sentence pair is shown in fig. 5, for the sentence (Xsc) "in the prosecution text with a litigant, who eats overnight somewhere with one, and then one, to go together to a garden to find her girls, when they reach a garden where one person comes to call them away, and when they are ready to go away, for less than a minute, the notifier either takes a helper to take a spring knife to go without asking the cause to stab them, and the notifier pokes both arms of the first. "and sentence (Xbc) in the dialect text" think that the mislabor cost of the original reporter A accompanied by civil litigation is too high. "there is no contradiction, but rather than sentence (Xbc) in the dialectic text" some of the defenders argue that they do not take a knife or a stone to get a person, and do not have any opinion on other facts dictated by the public prosecution department. There are contradictions.
For the contradiction classification model, each word in an input sentence is firstly modeled into a word vector through a word vector matrix through word embedding by an embedding layer, and the dimension is 300 dimensions. And secondly, adopting a convolution structure to accept the word vector as input to carry out feature extraction, and obtaining 768-dimensional feature expression vectors. And finally, transmitting the feature expression vector into a multilayer perceptron to obtain a final result, namely a contradiction classification vector.
In the model training process, the contradiction classification vectors are corrected by sentence-level contradiction relation vectors and contradiction classification vectors in a cross-product mode to obtain contradiction classification vector representation of sentence pairs, the corrected result is extracted according to the maximum value of each category of the sentence-level attributive complaint texts to obtain the sentence-level whole complaint text contradiction classification vectors, and then the sentence-level complaint text contradiction classification results and the paragraph-level complaint text contradiction classification results are subjected to averaging calculation to obtain final dispute focus vector representation of the complaint texts.
3. And finally expressing the vectors as labels for the two sections of input texts to output, namely all dispute focuses existing in the solved appealing texts, and saving the model which is trained to obtain the best model as a dispute focus recognizer.
4. The method comprises the steps of preprocessing any dispute party paired expression as an input, inputting the preprocessed dispute focus identifier obtained through training, and outputting the identified category of the dispute focus of the dispute party paired expression by the dispute focus identifier, such as dispute focus type labels L6_ CCPCCFW, L8_ SWPCFW _ y, L9_ JFWJ, L11_ RHPC, L13_ GZZR and L13_ GZZR.
The technical features of the embodiments described above may be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the embodiments described above are not described, but should be considered as being within the scope of the present specification as long as there is no contradiction between the combinations of the technical features.
The above-mentioned embodiments only express several embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention. Therefore, the protection scope of the present patent should be subject to the appended claims.
Claims (3)
1. A dispute focus identification method based on a deep neural network is characterized by comprising the following steps:
the method comprises the following steps: acquiring a large number of referee documents, preliminarily cleaning the referee documents by using a regular expression method, extracting original appellations and compelling allegories from the referee documents, carrying out dispute focus category labeling expressed in pairs by a legal expert, and completing construction of a dispute focus library;
step two: performing sentence-level and paragraph-level training on the model based on the deep neural network by using the dispute focus library to obtain a dispute focus recognizer; the model based on the deep neural network comprises an input layer, a model layer, a correction layer and an output layer, and the second step comprises the following steps:
step two, firstly: the input layer takes the complaint party text, the dialect text and the complaint sentence obtained after the complaint text is preprocessed and matched as the input text, converts the input text into a tensor which is easy to understand by a model, and inputs the tensor into the embedding layer of the model layer for calculating the sentence-level feature representation and the paragraph-level feature representation by the model layer;
step two: in the model layer, receiving an appeasing sentence and an appeasing sentence input tensor obtained after matching an appeasing text, a dialectic text and an appeasing text, calculating sentence-level feature representation and paragraph-level feature representation, and further performing contradiction detection by adopting a contradiction detection model based on BERT-CBGA (belief-based genetic algorithm) for calculating contradiction detection feature representation, wherein the contradiction detection feature representation comprises sentence-level contradiction detection feature representation; carrying out contradiction classification by adopting a model with a convolution structure, and calculating contradiction classification characteristic representation which comprises sentence-level contradiction classification characteristic representation and paragraph-level contradiction classification characteristic representation; carrying out contradiction detection and classification at a sentence level, carrying out contradiction detection on the input tensor of the apperceive sentences obtained after matching of the apperceive texts to obtain vectors of whether the apperceive sentence pairs have contradictions, carrying out contradiction classification to obtain contradiction classification vectors of each sentence pair, and carrying out contradiction classification at a paragraph level to obtain contradiction classification vectors;
step two and step three: in a correction layer, correcting the contradiction classification vectors by using whether each sentence pair has a contradiction relation vector and the contradiction classification vectors in a cross-product mode to obtain contradiction classification vector representation of each sentence pair, extracting the corrected result according to the maximum value of each category of the sentence pair attributive complaint texts to obtain sentence-level whole complaint text contradiction classification vectors, and performing averaging calculation on the sentence-level complaint text contradiction classification results to obtain final dispute focus vector representation of the complaint texts;
step two, four: the output layer outputs the dispute focus vector of the final dispute text by using a dispute focus category label to obtain a dispute focus identifier;
step three: and preprocessing the paired expressions of the appealing parties and the dispute parties to be identified to serve as input, transmitting the preprocessed paired expressions of the appealing parties and the dispute parties to the dispute focus recognizer obtained through training in the step two, and outputting the category of the dispute focus of the paired expressions of the appealing and the dispute parties to be identified by the dispute focus recognizer.
2. The method of claim 1, wherein the deep neural network-based dispute focus identification method,
and representing the dispute focus category by using a binary vector, wherein the binary vector is used as a label vector expressed by the dispute parties in pairs.
3. The method for identifying the dispute focus based on the deep neural network as claimed in claim 1, wherein the contradiction detection model based on the BERT-CBGA comprises an embedding layer, a feature extraction layer, a feature fusion layer and an output layer, and the process of carrying out the contradiction detection by adopting the contradiction detection model based on the BERT-CBGA comprises the following steps:
step two, step one: the embedding layer comprises word embedding and sentence pair embedding for calculating original notice appellations and reported allegories, each word in the appellation sentences and the dialect sentences is modeled into a word vector through a word vector matrix, random initialization is adopted for initializing the word vector, a word vector untrainable mode is selected in the training process, word embedding of the sentences of both sides of the appellations is obtained, and sentence pair embedding is realized through BERT;
step two and step two: the feature extraction layer is divided into two steps, the local feature representation of the original notice appellation and the noticed appellation is calculated through convolution operation on independent sentences, pooling output feature vectors are obtained, the features of the original notice appellation and the noticed appellation with context information are calculated through the BiGRU, and the features are further extracted through an attention mechanism, so that attention output feature vectors are obtained;
step two, step three: the fusion layer splices and fuses the pooled output feature vector and the attention output feature vector obtained by the feature extraction layer with the BERT sentence to obtain final features, namely contradiction detection features;
step two, step four: in order to improve the performance and generalization capability of the model, the output layer transmits the result to a full connection layer activated by Sigmoid for classification through a Dropout layer and a full connection layer activated by ReLU, so as to obtain the final contradiction detection characteristic representation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110665262.XA CN113553856B (en) | 2021-06-16 | 2021-06-16 | Deep neural network-based dispute focus identification method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110665262.XA CN113553856B (en) | 2021-06-16 | 2021-06-16 | Deep neural network-based dispute focus identification method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113553856A CN113553856A (en) | 2021-10-26 |
| CN113553856B true CN113553856B (en) | 2022-08-26 |
Family
ID=78102180
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110665262.XA Active CN113553856B (en) | 2021-06-16 | 2021-06-16 | Deep neural network-based dispute focus identification method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113553856B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115017917B (en) * | 2022-08-09 | 2022-10-28 | 北京肇祺信息科技有限公司 | Judgment document dispute focus identification method based on multi-head attention mechanism |
| CN116186574B (en) * | 2022-09-09 | 2023-12-12 | 武汉中数医疗科技有限公司 | Thyroid sampling data identification method based on artificial intelligence |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112651853A (en) * | 2020-11-17 | 2021-04-13 | 四川大学 | Judgment and opinion mining method and system based on referee document |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109992664B (en) * | 2019-03-12 | 2023-04-18 | 平安科技（深圳）有限公司 | Dispute focus label classification method and device, computer equipment and storage medium |
| CN111143550B (en) * | 2019-11-27 | 2022-05-03 | 浙江大学 | Method for automatically identifying dispute focus based on hierarchical attention neural network model |
| CN112347269A (en) * | 2020-11-11 | 2021-02-09 | 重庆邮电大学 | Method for recognizing argument pairs based on BERT and Att-BilSTM |
| CN112613582B (en) * | 2021-01-05 | 2022-06-24 | 重庆邮电大学 | Deep learning hybrid model-based dispute focus detection method and device |
| CN113468323B (en) * | 2021-06-01 | 2023-07-18 | 成都数之联科技股份有限公司 | Dispute focus category and similarity judging method, system and device and recommending method |
-
2021
- 2021-06-16 CN CN202110665262.XA patent/CN113553856B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112651853A (en) * | 2020-11-17 | 2021-04-13 | 四川大学 | Judgment and opinion mining method and system based on referee document |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113553856A (en) | 2021-10-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110287320B (en) | Deep learning multi-classification emotion analysis model combining attention mechanism | |
| CN110427623B (en) | Semi-structured document knowledge extraction method and device, electronic equipment and storage medium | |
| CN107632968B (en) | Method for constructing evidence chain relation model for referee document | |
| CN108492118B (en) | Two-stage extraction method for return visit text data of automobile after-sale service quality evaluation | |
| CN113553856B (en) | Deep neural network-based dispute focus identification method | |
| CN110990525A (en) | Natural language processing-based public opinion information extraction and knowledge base generation method | |
| CN107315738A (en) | A kind of innovation degree appraisal procedure of text message | |
| US20200142856A1 (en) | Systems and methods for extracting specific data from documents using machine learning | |
| CN111177383B (en) | Text entity relation automatic classification method integrating text grammar structure and semantic information | |
| CN111079985A (en) | Criminal case criminal period prediction method based on BERT and fused with distinguishable attribute features | |
| CN111177382B (en) | Intelligent legal system recommendation auxiliary system based on FastText algorithm | |
| CN112508600A (en) | Vehicle value evaluation method based on Internet public data | |
| CN113168499A (en) | Method for searching patent document | |
| CN111159336A (en) | Semi-supervised judicial entity and event combined extraction method | |
| CN111339284A (en) | Product intelligent matching method, device, equipment and readable storage medium | |
| CN109492230A (en) | A method of insurance contract key message is extracted based on textview field convolutional neural networks interested | |
| CN113196277A (en) | System for retrieving natural language documents | |
| CN111858940A (en) | Multi-head attention-based legal case similarity calculation method and system | |
| CN113919366A (en) | Semantic matching method and device for power transformer knowledge question answering | |
| CN108829671A (en) | Method, apparatus, storage medium and the terminal device of decision based on survey data | |
| CN115952292B (en) | Multi-label classification method, apparatus and computer readable medium | |
| CN113282711A (en) | Internet of vehicles text matching method and device, electronic equipment and storage medium | |
| CN110334192B (en) | Text abstract generation method and system, electronic equipment and storage medium | |
| CN106202299B (en) | Disabled person authoritative user recommendation method based on disabled person characteristics | |
| CN115017879A (en) | Text comparison method, computer device and computer storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |