A method for prison sentence prediction by incorporating auxiliary knowledge
By integrating auxiliary knowledge into the sentence prediction, processing legal document data and building an ERNIE_DPCNN hybrid network model, the problem of neglect of data processing and insufficient migration capabilities in the existing technology is solved, and more efficient and accurate sentence prediction results are achieved.
Patent Information
- Application Number
- CN202111491204.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-08
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2041-12-08
AI Technical Summary
The existing prison prediction technology spends a lot of effort on the model structure, but ignores the processing of training data, resulting in limited improvement in results, and traditional methods are difficult to interpret prediction results, and the migration ability in similar legal fields is weak, and the distribution of legal documents data Inhomogeneity affects the prediction effect.
Using a method that integrates auxiliary knowledge, by obtaining legal document sample data sets, designing extraction templates to extract fact descriptions and legal terms, performing data supplement and noise removal, building an ERNIE_DPCNN hybrid network model for training, and generating more accurate sentence prediction results .
It realizes a more convenient, accurate and efficient sentence prediction, reduces the influence of subjective factors, improves the uniformity of data distribution and word vector generation effect, reduces noise interference, and improves the accuracy of extraction.
Smart Images

Figure CN114239939B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of legal text data classification, and in particular to a method for sentence prediction by integrating auxiliary knowledge. Background Art
[0002] In the current process of sentencing, judges not only need to analyze complex and cumbersome legal documents, but also refer to previous similar cases to make judgments, which leads to low efficiency. This judgment method not only requires judges to have rich experience and knowledge, but also the judgment results are often affected by certain subjective factors.
[0003] Legal intelligence is an emerging industry in the legal field, which aims to give machines the ability to read, understand and analyze legal texts. Legal intelligence includes but is not limited to scenarios such as sentence prediction, crime prediction, similar case matching, legal intelligent question and answer, etc. It is expected to help judges, lawyers and other legal professionals to solve legal problems efficiently, and also help non-legal professionals to learn more about legal knowledge. In recent years, with the continuous development of deep learning technology, the legal intelligence industry has gradually emerged and received widespread attention from academia and industry.
[0004] Existing research on sentence prediction usually focuses on the model structure and ignores the processing of training data, so it is difficult to significantly improve the training results. In addition, the existing technology only processes the factual description part of the data, ignoring the importance of legal knowledge for sentence prediction. Secondly, the word vector generated by traditional Word2vec (a group of related models used to generate word vectors) software is affected by the training data, and the generated word vector effect is poor, which will affect the results of the sentence prediction task. Finally, sentence prediction is different from other text classification tasks. Due to the particularity of the legal field, it is very important to give a reasonable explanation for the sentence prediction results. The current research work lacks an explanation of the sentence prediction results. Secondly, the traditional machine learning algorithm has weak migration capabilities and is difficult to replicate in similar legal field tasks. Summary of the invention
[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method for sentence prediction by incorporating auxiliary knowledge.
[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for sentence prediction by integrating auxiliary knowledge, comprising the following steps:
[0007] Step 1: Get the legal document sample dataset, use the toolkit in Python language, call json.loads() method to extract the fact description, involved laws and sentence results from the legal document sample dataset and save them in a new file;
[0008] Step 2: To address the problem of uneven distribution of the legal document sample dataset, data supplement is performed on the types with fewer samples in the legal document sample dataset. The method is as follows:
[0009] Step 2.1: Download legal document data from the website and divide the legal document data into six parts: document header, accusation content, defense opinion, fact description, judgment content and document footer;
[0010] Step 2.2: Design an extraction template to extract the fact description part and the tail part of the legal document data respectively;
[0011] The extraction template is divided into a fact description part extraction template and a document tail part extraction template;
[0012] The extraction template of the description of the alleged facts is divided into two parts: the front template and the back template. The two templates are spliced by regular expressions in the middle, and the front template and the back template are matched at the same time during matching;
[0013] The document tail part extraction template is formed by simply splicing the preceding template with a regular expression, and is directly matched to the end of the legal document during matching.
[0014] Step 2.3: The extracted fact description part and the tail part of the document are merged to form training data, and the legal document sample data set is supplemented.
[0015] Step 3: The legal provisions involved in the case extracted in step 1 are integrated as auxiliary knowledge and fact description to achieve the integration of auxiliary knowledge of the supplemented data set. The process is as follows:
[0016] Step 3.1: Cut the length of the legal provisions involved in the case and make them shorter, and unify the length of the fact description part;
[0017] The method for performing the operation of truncating and supplementing the laws involved in the case is as follows: processing the laws into data with a length of 100, if the length of the laws is less than 100, use "[PAD]" to supplement it, if the length of the laws exceeds 100, directly truncate it; the role of the "[PAD]" is to fill in as supplementary data when the data length does not meet the required requirements.
[0018] Step 3.2: Connect the legal provisions and fact descriptions involved in the case to form a predicted text;
[0019] When splicing the legal provisions and factual description involved in the case, "[SEP]" is used to divide the two parts.
[0020] Step 3.3: Set the predicted label Label according to the sentence result and use it as the predicted value. The composed "prediction text prediction value" is used as the model input data.
[0021] The method for setting the predicted label Label according to the sentence result is as follows:
[0022] The sentence is divided into twelve sentence prediction results, including 1 year or less, 1-2 years, 2-3 years, 3-4 years, 4-5 years, 5-6 years, 6-7 years, 7-8 years, 8-9 years, 9-10 years, more than 10 years, and life imprisonment. These twelve sentence prediction results are represented by prediction labels Label, and the value of the Label is an integer 0-11, corresponding to the twelve sentence prediction results.
[0023] Step 4: Perform data enhancement processing on the fused data set to remove noise data; and randomly divide the data set into training set, test set and verification set;
[0024] The noise data includes punctuation marks, garbled characters, blank lines, English, stop words, time content, names of people and institutions, etc.
[0025] Step 5: Build the ERNIE_DPCNN hybrid network model, and train, verify and evaluate the built model. The process is as follows:
[0026] Step 5.1: Model parameter initialization, including ERNIE model parameter initialization and DPCNN model parameter initialization;
[0027] Step 5.2: Input the training set obtained in step 4 into the ERNIE model to vectorize the data set to achieve the trained language representation;
[0028] Step 5.3: Input the vectorized data set into the DPCNN model for calculation, and input the output of the ERNIE_DPCNN hybrid network model into the Softmax classifier for classification to obtain the probability value of the sentence prediction result of each category, thus completing the training of the model;
[0029] Step 5.4: Use the validation set as validation data to train the hyperparameters of the model;
[0030] Step 5.5: Use the test set to test the model and evaluate the test results.
[0031] The ERNIE_DPCNN hybrid network model is: ERNIE (Enhanced Representation through Knowledge Integration) and DPCNN (Deep Pyramid Convolutional Neural Networks) are integrated into a new ERNIE_DPCNN hybrid network model. Through a large number of experimental comparisons, including but not limited to TextCNN model, BERT model, BERT_CNN model, BERT_DPCNN model, ERNIE model, ERNIE_CNN model, ERNIE_DPCNN model, the ERNIE_DPCNN hybrid network model finally achieved the best sentence prediction results because the ERNIE_DPCNN hybrid network model exceeded other models in accuracy, recall rate, and Fn value, and then after model fine-tuning, the model achieved the best sentence prediction results.
[0032] Among them, the ERNIE (Enhanced Representation through Knowledge Integration) model is a pre-training model. Inputting the CAIL2018-Large dataset into the ERNIE model can train a pre-trained language model related to the legal field. Therefore, the language representation effect after training with the ERNIE network model far exceeds the language representation generated by the Word2Vec tool, and the training speed is faster and the cost is lower.
[0033] Among them, the DPCNN (Deep Pyramid Convolutional Neural Networks) model mainly performs downstream tasks - prison sentence prediction tasks. The advantage of DPCNN is that it can extract long text information by stacking CNN model layers, which is more suitable for prison sentence prediction tasks.
[0034] Step 6: Load the ERNIE_DPCNN hybrid network model that passed the evaluation in step 5, input the legal document sample to predict the sentence, and take the result with the highest probability as the final sentence prediction result.
[0035] The beneficial effects of adopting the above technical solution are:
[0036] 1. The method provided by the present invention ignores the influence of subjective factors and, based on deep learning technology, achieves more convenient, accurate and efficient sentence prediction by training a large number of criminal law documents in the legal field, which provides great convenience for judges in sentencing.
[0037] 2. The method provided by the present invention solves the problem that uneven distribution of data sets affects the sentence prediction results.
[0038] 3. The method provided by the present invention adopts knowledge integration to implement the enhanced representation model ERNIE to generate word vectors. Compared with the word vectors generated by the traditional Word2vec model software, the word vectors generated by the present invention are not affected by training data. The effect of generating word vectors is better, the training speed is faster, and the cost is lower.
[0039] 4. The method provided by the present invention first completes the extraction of the tail part of the legal document, and then extracts the two parts of the law involved and the sentence, which not only reduces the difficulty of extraction, but also reduces noise interference, greatly improving the extraction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 It is an overall flow chart of a method for sentence prediction by incorporating auxiliary knowledge in an embodiment of the present invention;
[0041] Figure 2 A flowchart of obtaining a data set and preprocessing the data set in an embodiment of the present invention;
[0042] Figure 3 This is a statistical diagram of the uneven distribution of the CAIL2018-Large dataset in an embodiment of the present invention;
[0043] Figure 4 is a flow chart of the auxiliary knowledge fusion process in an embodiment of the present invention;
[0044] Figure 5 This is a flowchart of building the ERNIE_DPCNN hybrid network model in an embodiment of the present invention. DETAILED DESCRIPTION
[0045] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.
[0046] like Figure 1 As shown, the method of incorporating auxiliary knowledge into prison sentence prediction in this embodiment is as follows, wherein the data preprocessing part includes steps 1 to 4, and the process is as follows: Figure 2 shown.
[0047] Step 1: Get the legal document sample dataset, use the toolkit in Python language, call json.loads() method to extract the fact description, involved laws and sentence results from the legal document sample dataset and save them in a new file;
[0048] In this embodiment, the CAIL2018-Large dataset released in the "China Law Research Cup---Judicial Artificial Intelligence Challenge" is used, which includes 1.5 million legal document sample data. The data is stored in Json format, mainly including fact description, defendant, fine (unit: yuan), crime, legal provisions involved in the case, and sentence results. In this embodiment, only the fact description, legal provisions involved in the case, and sentence results in the dataset are used. In the process of dataset processing, there will be uneven distribution of datasets, such as Figure 3 As shown, the data on life imprisonment are too small compared to the data on other prison terms, so the data on life imprisonment need to be supplemented.
[0049] Step 2: To address the problem of uneven distribution of the legal document sample dataset, data supplement is performed on the types with fewer samples in the legal document sample dataset. The method is as follows:
[0050] Step 2.1: Download legal document data from the website and divide the legal document data into six parts: document header, accusation content, defense opinion, fact description, judgment content and document footer;
[0051] Step 2.2: Design an extraction template to extract the fact description part and the tail part of the legal document data respectively;
[0052] The extraction template is divided into a fact description part extraction template and a document tail part extraction template;
[0053] The extraction template of the description of the alleged facts is divided into two parts: the front template and the back template. The two templates are spliced by regular expressions in the middle, and the front template and the back template are matched at the same time during matching;
[0054] The document tail part extraction template is formed by simply splicing the preceding template with a regular expression, and is directly matched to the end of the legal document during matching.
[0055] Step 2.3: The extracted fact description part and the tail part of the document are merged to form training data, and the legal document sample data set is supplemented.
[0056] In this example, we first download legal document data from the China Judgment Documents Network, and after reading a lot of legal documents and consulting with experts in the legal field, we divide a complete legal document into six parts: document header, accusation content, defense opinion, fact description, judgment content, and document footer. Then we design a complete set of extraction templates for matching and extract the fact description part in the legal document.
[0057] In this embodiment, the fact description part extraction template defines the preceding template existence set F n={f1, f2, ..., fn}, define the post-template set B m ={b1, b2, ..., b m}, define the middle regular expression as "(.*?)", and prepend the template set F n With post template set B m Perform pairwise splicing and splice regular expressions between the two sets. Template matching is shown in Table 1 below:
[0058] Table 1. Extraction template for fact description part
[0059]
[0060] Among them, the regular expression means matching the preceding template set F n With post template set B m The content of any number of characters between . This method is used to extract the fact description part of the legal document.
[0061] In this embodiment, the method for extracting the tail part of a document is the same as the method for extracting the fact description part. The template for extracting the tail part of a document is designed as shown in Table 2 below.
[0062] Table 2. Template for extracting the tail part of the document
[0063]
[0064] Among them, the post-template is no longer defined at the end of the legal document. It can directly match the pre-template concatenated regular expression and directly match it to the end of the legal document.
[0065] The purpose of extracting the tail part of the legal document is to assist in the subsequent identification and extraction of the two parts of the law and sentence involved in the case. The same template extraction method is used to extract the law and sentence involved in the case. The extraction characteristics of the law and sentence involved in the case are relatively obvious. The law involved in the case is often directly displayed at the end of the document in the form of an appendix to the relevant provisions of the Criminal Law of the People's Republic of China, and the sentence appears in the content of the judgment, such as a sentence of X years in prison.
[0066] Among them, if the full text of the legal document is used to extract the two parts of the law involved and the sentence, it is difficult and easy to be disturbed by other contents in the legal document. Therefore, the tail part of the legal document is extracted first, which greatly improves the extraction accuracy.
[0067] Step 3: The legal provisions involved in the case extracted in step 1 are integrated as auxiliary knowledge and fact description to achieve the integration of auxiliary knowledge of the supplemented data set. The process is as follows: Figure 4 As shown, the process is as follows:
[0068] Step 3.1: Cut the length of the legal provisions involved in the case and make them shorter, and unify the length of the fact description part;
[0069] The method for performing the operation of truncating and supplementing the laws involved in the case is as follows: processing the laws into data with a length of 100, if the length of the laws is less than 100, use "[PAD]" to supplement it, if the length of the laws exceeds 100, directly truncate it; the role of the "[PAD]" is to fill in as supplementary data when the data length does not meet the required requirements.
[0070] In this embodiment, the length of the fact description part is set to 300, and if it exceeds 300, it can be directly truncated.
[0071] Step 3.2: Connect the legal provisions and fact descriptions involved in the case to form a predicted text;
[0072] When splicing the legal provision and the fact description part involved in the case, "[SEP]" is used to separate the two parts. The function of "[SEP]" is to perform a separation mark function on the legal provision and the fact description part involved in the case.
[0073] Step 3.3: Set the predicted label Label according to the sentence result and use it as the predicted value. The composed "prediction text prediction value" is used as the model input data.
[0074] The method for setting the predicted label Label according to the sentence result is as follows:
[0075] The sentence is divided into twelve sentence prediction results, including 1 year or less, 1-2 years, 2-3 years, 3-4 years, 4-5 years, 5-6 years, 6-7 years, 7-8 years, 8-9 years, 9-10 years, more than 10 years, and life imprisonment. These twelve sentence prediction results are represented by prediction labels Label, and the value of the Label is an integer 0-11, corresponding to the twelve sentence prediction results.
[0076] Step 4: Perform data enhancement processing on the fused data set to remove noise data; and randomly divide the data set into training set, test set and verification set;
[0077] The noise data includes punctuation marks, garbled characters, blank lines, English, stop words, time content, names of people and institutions, etc.
[0078] There are many methods for processing text data, mainly based on experience to process the data into data that is more suitable for prison sentence prediction and to remove noise data as much as possible. The specific processing methods are as follows: remove meaningless symbols in the data set, such as punctuation, garbled characters, blank lines, etc.; remove interference factors such as English, stop words, time content, names of people, names of institutions, etc.
[0079] Among them, removing stop words requires building a stop word library, and removing stop words in the dataset based on the stop word library information.
[0080] Among them, there are two main methods for removing names of people and organizations. The first is to use the HanLP natural language processing package to efficiently recognize named entities of people and identify the names of people. The HanLP natural language processing package is an open source Chinese natural language processing software that can provide a named entity recognition API to facilitate the recognition of names of people and organizations.
[0081] HanLP natural language processing cannot recognize ambiguous names that often appear in legal documents, such as "Zhao, Zhao Moumou", and so on, and a reasonable extraction method needs to be designed. Designing a matching template "defendant (.*?)" can achieve name recognition and supplement the HanLP natural language processing package.
[0082] Furthermore, this embodiment also provides a method for fine-tuning relevant legal provisions, as described below:
[0083] Because neither the judgment documents nor the data in the "2018 China Law Research Cup Judicial" fully display the content of the law, and the basic knowledge displays the law number, it is necessary to match the law and the content of the law one by one, and update the corresponding data to the law data. The example is as follows: the data in the "2018 China Law Research Cup Judicial" is displayed in the form of "elevant_articles: 133". Among them, "133" corresponds to Article 133 of the Criminal Law, and the replacement content is roughly "[Traffic accident crime; dangerous driving crime] Violation of transportation management regulations, resulting in a major accident...".
[0084] In this embodiment, the training set, test set, and verification set are divided in a ratio of 5:1:1. Due to the uneven distribution of the data set, 42,000 evenly distributed data samples are extracted in this embodiment, of which the training set is 30,000 data, and the test set and verification set have 6,000 data respectively.
[0085] Step 5: Build the ERNIE_DPCNN hybrid network model, whose structure is as follows Figure 5 As shown, the built model is trained, verified and evaluated. The process is as follows:
[0086] Step 5.1: Model parameter initialization, including ERNIE model parameter initialization and DPCNN model parameter initialization;
[0087] In this embodiment, some important parameters of the ERNIE model are set as follows: Set the Epoch parameter to 3, that is, use all the data in the training set to train the model three times. Set Batch_Size to 2, that is, the number of samples in each batch of training is 2, and set Padding_Size to 400. Here, we have processed the data into the corresponding length in step 1. In order to make the program more robust, we also set it here. Learning_Rate is 5e-5, which sets the convergence speed. Set the ERNIE hidden layer Hidden_Size to 768.
[0088] In this embodiment, the initial domain of the convolution of the DPCNN model sets In_Channels (number of input channels) to 1, Out_Channels (number of output channels) to 250, Kernel_Size (core) to (3,768), and Stride (stride) convolution to 1 for each sliding step. Then the convolution sets In_Channels (number of input channels) to 250, Out_Channels (number of output channels) to 250, Kernel_Size (core) to (3,1), and Stride (stride) convolution to 1 for each sliding step. The pooling layer mainly performs feature dimensionality reduction. The parameters set Kernel_Size (core) to (3,1), and Stride (stride) convolution to 2 for each sliding step.
[0089] Step 5.2: Input the training set obtained in step 4 into the ERNIE model to vectorize the data set to achieve the trained language representation;
[0090] Step 5.3: Input the vectorized data set into the DPCNN model for calculation, and input the output of the ERNIE_DPCNN hybrid network model into the Softmax classifier for classification to obtain the probability value of the sentence prediction result of each category, thus completing the training of the model;
[0091] In this embodiment, the vectorized data set is input into the DPCNN model. First, Region embedding is performed, which uses multi-size convolution filters to avoid convolution of all adjacent positions. Then, two Conv operations are performed on the vector, and the size of the Feature Map is kept consistent with the original size according to the Padding_Size (input training data length). Figure 1The Pooling layer is then connected, which is mainly used to integrate the tiny feature points obtained after the convolution layer into new features, and then perform two Conv operations, such as Figure 5 As shown, repeat the operation to perform training stacking. Finally, perform the Pooling operation again to complete the ERNIE_DPCNN hybrid model splicing.
[0092] Step 5.4: Use the validation set as validation data to train the hyperparameters of the model;
[0093] Step 5.5: Use the test set to test the model and evaluate the test results.
[0094] In this embodiment, precision (P), recall (R), and F1-score (F1) are used to evaluate the performance of the ERNIE_DPCNN hybrid network model, where TP means that the correct label is positive, FP means that the wrong label is positive, FN means that the wrong label is negative, and TN means that the correct label is negative.
[0095]
[0096]
[0097]
[0098] Step 6: Load the ERNIE_DPCNN hybrid network model that passed the evaluation in step 5, input the legal document sample to predict the sentence, and take the result with the highest probability as the final sentence prediction result.
Claims
1. A method for predicting prison terms by incorporating auxiliary knowledge, characterized in that: The steps include: Step 1: Get the legal document sample dataset, use the toolkit in Python language, call json.loads() method to extract the fact description, involved laws and sentence results from the legal document sample dataset and save them in a new file; Step 2: To address the problem of uneven distribution of the legal document sample dataset, supplement the data for the types with fewer samples in the legal document sample dataset; Step 2.1: Download legal document data from the website and divide the legal document data into six parts: document header, accusation content, defense opinion, fact description, judgment content and document footer; Step 2.2: Design an extraction template to extract the fact description part and the tail part of the legal document data respectively; The extraction template is divided into a fact description part extraction template and a document tail part extraction template; The extraction template of the description of the alleged facts is divided into two parts: the front template and the back template. The two templates are spliced by regular expressions in the middle, and the front template and the back template are matched at the same time during matching; The document tail part extraction template is only composed of the preceding template and the regular expression, and is directly matched to the end of the legal document during matching; Step 2.3: The extracted fact description part and the tail part of the document are merged to form training data, and the legal document sample data set is supplemented; Step 3: The legal provisions involved in the case extracted in step 1 are integrated as auxiliary knowledge and fact description to achieve the integration of auxiliary knowledge in the supplemented data set; Step 4: Perform data enhancement processing on the fused data set to remove noise data; And randomly divide the data set into training set, test set and verification set; Step 5: Build the ERNIE_DPCNN hybrid network model, and train, verify and evaluate the built model; Step 5.1: Model parameter initialization, including ERNIE model parameter initialization and DPCNN model parameter initialization; Step 5.2: Input the training set obtained in step 4 into the ERNIE model to vectorize the data set to achieve the trained language representation; Step 5.3: Input the vectorized data set into the DPCNN model for calculation, and input the output of the ERNIE_DPCNN hybrid network model into the Softmax classifier for classification to obtain the probability value of the sentence prediction result of each category, thus completing the training of the model; Step 5.4: Use the validation set as validation data to train the hyperparameters of the model; Step 5.5: Use the test set to test the model and evaluate the test results; Step 6: Load the ERNIE_DPCNN hybrid network model that passed the evaluation in step 5, input the legal document sample to predict the sentence, and take the result with the highest probability as the final sentence prediction result.
2. The method for predicting prison terms by incorporating auxiliary knowledge according to claim 1, characterized in that: The process of step 3 is as follows: Step 3.1: Cut the length of the legal provisions involved in the case and make them shorter, and unify the length of the fact description part; Step 3.2: Connect the legal provisions and fact descriptions involved in the case to form a predicted text; Step 3.3: Set the predicted label Label according to the sentence result and use it as the predicted value. The composed "predicted text predicted value" is used as the model input data.
3. The method for predicting prison terms by incorporating auxiliary knowledge according to claim 2, characterized in that: The method for performing the operation of truncating and supplementing the laws involved in the case is: processing the laws into data with a length of 100, if the length of the laws is less than 100, use "[PAD]" to supplement it, if the length of the laws exceeds 100, directly truncate it; the role of the "[PAD]" is to fill it as supplementary data when the data length does not meet the required requirements.
4. The method for predicting prison terms by incorporating auxiliary knowledge according to claim 2, characterized in that: In step 3.2, "[SEP]" is used to separate the two parts when splicing the legal provisions and the fact description part involved in the case.
5. The method for predicting prison terms by incorporating auxiliary knowledge according to claim 2, characterized in that: The method for setting the predicted label Label according to the sentence result is as follows: The sentence is divided into twelve sentence prediction results, including 1 year or less, 1-2 years, 2-3 years, 3-4 years, 4-5 years, 5-6 years, 6-7 years, 7-8 years, 8-9 years, 9-10 years, more than 10 years, and life imprisonment. These twelve sentence prediction results are represented by prediction labels Label, and the value of the Label is an integer 0-11, corresponding to the twelve sentence prediction results.
6. The method for predicting prison terms by incorporating auxiliary knowledge according to claim 1, characterized in that: The noise data in step 4 includes punctuation marks, garbled characters, blank lines, English, stop words, time content, names of people and names of organizations.
Citation Information
Cited By
Data-driven criminal period prediction system
CN120931121A
