A criminal case evidence pointing method based on NLP technology

Abstract

The application discloses a criminal case evidence guiding method based on NLP technology, which comprises case category judgment; specifically, on the basis of improving the accuracy of brief case information part-of-speech tagging, a text classification algorithm in NLP technology is used to predict the category of a case, so as to improve the accuracy of case category judgment; a recommendation engine collects case information; a recommendation result of case evidence collection content is generated; the recommendation result is loaded into a task flow to form a to-be-done task. The application provides a method of combining an artificial intelligence recommendation engine with a task flow, combines an automatic recommendation with a case evidence collection method and an evidence collection task flow, and is used for solving the mistakes in collecting criminal case evidence by experience and manual recording in the traditional mode, so as to avoid the situation that a case is returned to a procuratorate for supplementary investigation or the situation that evidence is insufficient due to mistakes, and even the situation that a case is exempted from prosecution, and the application can effectively improve the case handling efficiency and the final prosecution rate of a public security organ in criminal case investigation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of criminal case evidence collection, and particularly relates to a criminal case evidence guiding method based on NLP technology. BACKGROUND

[0002] In the criminal case investigation, firstly, the case category registration often appears inaccurate, and if only relying on hard coding to realize the evidence collection function of various cases, the evidence system will inevitably be inconsistent with the case category.

[0003] Secondly, due to different criminal constitutions of various crimes and different case characteristics of various cases, there are many differences in evidence forms, evidence systems, evidence thinking, evidence methods and evidence matters needing attention of various specific criminal cases. Therefore, in addition to following the general evidence standards, evidence thinking and evidence requirements of criminal cases, the evidence standards, evidence thinking and evidence requirements of specific criminal cases should also be combined with the constitution characteristics and case characteristics of various specific criminal cases, and personalized and differentiated evidence standards, evidence thinking and evidence requirements should be adhered to.

[0004] The personalized and differentiated evidence standards, evidence thinking and evidence requirements need the guidance of the experience and solid theoretical knowledge of the on-duty police. The criminal investigators should especially strengthen the evidence consciousness in the criminal case investigation, and should carefully learn and master the legal provisions of the concept of evidence, the types of evidence, the legal procedure of collecting evidence, the examination and verification of evidence, and should collect evidence in accordance with the law in a comprehensive, objective and practical manner, so that the criminal prosecution work can be carried out smoothly. Otherwise, in the actual work, the evidence collection is incomplete, the evidence thinking is unclear, and the criminal facts are not clear, the evidence is insufficient, and the inspection agencies do not adopt the situation, so that the case cannot be done, thereby affecting the efficiency and the case solving rate of the case investigation.

[0005] In the criminal case investigation, although the evidence forms are only divided into eight legal forms, (1) physical evidence; (2) documentary evidence; (3) witness testimony; (4) victim statement; (5) criminal suspect, defendant confession and defense; (6) expert opinion; (7) records of inspection, examination, identification, investigation experiment; (8) audio-visual materials, electronic data, but only in the classification of physical evidence, according to the different case categories, even the different specific conditions in the same case handling process, the content requirements of the physical evidence are very different. For example, rape cases need to extract fine hairs, bloodstains, traces, etc., intentional homicide cases need to extract crime tools, bloodstained clothes, bloodstains, etc., telecommunications fraud cases need to extract involved bank cards, network equipment for criminal activities, mobile phone cards, etc., which need to rely on the experience and theoretical knowledge of the criminal police.

[0006] In the process of handling such evidence extraction by public security organs, there are often insufficient evidence, and the procuratorate returns to supplement the investigation. Criminal police often do not investigate only one case at the same time in the process of investigating cases, and often multiple cases are intertwined in time, which inevitably leads to mistakes. The main reasons for returning to the public security organ for supplementary investigation are:

[0007] 1. The main criminal facts are unclear, and there are contradictions between the main evidence;

[0008] 2. The criminal constitutive elements are lacking;

[0009] 3. The main circumstances are not investigated;

[0010] 4. Important criminal facts and co-defendant suspects who should be held criminally responsible are omitted;

[0011] 5. Other important facts that have not been investigated. SUMMARY

[0012] The technical problem to be solved by the present application is to provide a criminal case evidence guiding method based on NLP technology, which can guide the evidence collection in criminal cases, so as to avoid the return of the case to the public security organ for investigation due to unclear main criminal facts, contradictions between main evidence and lack of criminal constitutive elements, and to improve the efficiency and accuracy of the public security organ in collecting evidence in criminal cases.

[0013] To solve the above technical problems, the technical solution adopted by the present application is as follows.

[0014] A criminal case evidence guiding method based on NLP technology combines artificial intelligence recommendation engine and task flow to realize the combination of the evidence collection process of automatic recommendation case and evidence collection task flow, specifically including:

[0015] Case category judgment; specifically, on the basis of improving the accuracy of brief case situation segmentation, using the text classification algorithm in NLP technology to predict the category of the case to improve the accuracy of case category judgment;

[0016] The recommendation engine collects case information;

[0017] Generate a recommended result of case evidence collection content;

[0018] The task flow loads the recommended result to form a to-do task.

[0019] Preferably, the method for improving the accuracy of brief case situation segmentation includes using conditional random field segmentation algorithm to segment the case, and annotating a public security domain corpus.

[0020] Preferably, the public security field corpus takes the State Language Commission modern balanced corpus as a training model, and selects the annotation specification "Modern Chinese Word Class Annotation Set Specification for Information Processing" (GB / T 20532-2006) consistent with the State Language Commission modern balanced corpus to expand the public security field corpus.

[0021] The raw corpus of the public security field corpus is derived from basic case data in the public security industry; firstly, the raw corpus is subjected to morphological analysis by the model generated by the State Language Commission modern balanced corpus, and the annotation is corrected in the output result to improve the annotation efficiency of the public security field corpus.

[0022] In the morphological analysis of the raw corpus, new words that are not registered need to be manually marked with their morphology to convert the raw corpus into cooked corpus; the conditional random field is continued to be used as a training algorithm to complete the training of the public security field corpus.

[0023] Preferably, the text classification algorithm is a linear support vector machine classifier, and the classification samples required by the linear support vector machine classifier and the raw corpus required in the process of improving the brief case situation word segmentation accuracy are managed by a sample training management module including sample preview and sample upload functions.

[0024] Preferably, the recommendation engine further includes determining a recommendation subject, a recommendation object, and determining a recommendation algorithm before collecting case information.

[0025] The recommendation subject is a case to be handled; after the case category judgment is completed, the recommendation subject is generated.

[0026] The recommendation object is various types of material evidence; eight types of evidence forms in the common evidence system of criminal cases are decomposed into the smallest granularity evidence elements as the recommendation object.

[0027] Preferably, the recommendation algorithm is to classify historical cases according to the recommendation category, and count the number of various types of evidence materials used in the case, which specifically includes the following steps:

[0028] Firstly, a case table is prepared, which contains case numbers and case categories;

[0029] Secondly, an evidence element table is prepared, which contains case numbers, evidence categories, and evidence elements.

[0030] Preferably, the recommendation engine collects case information by associating the case table with the evidence element table, and counting the evidence elements existing in the historical cases of the same category through a unary algorithm according to the judged case category.

[0031] Preferably, the one-dimensional algorithm of the evidence element is further improved to a two-dimensional or three-dimensional algorithm, that is, the correlation degree of the probability of occurrence of one evidence element in a case of one type and the probability of occurrence of another evidence element, and the evidence element with a high correlation degree is taken as a related evidence element for statistics.

[0032] Preferably, the recommendation result of the case evidence collection content is the proportion of all evidence elements in the case of the respective case type, and the proportion value is the recommended threshold value range, and the recommendation result is generated.

[0033] Preferably, the to-be-completed task is assigned to the corresponding police officer; after the police officer completes the task, the task flow automatically obtains the evidence collection audit leader according to the organizational structure of the unit where the police officer is located, the leader performs the audit, and after the leader completes the audit, the law department personnel are sent according to the relevant regulations for review.

[0034] Due to the adoption of the above technical solutions, the technical progress achieved by the present application is as follows.

[0035] The case type judgment method provided by the present application improves the accuracy of case type judgment by using the text classification algorithm in the NLP technology to predict the case type on the basis of improving the accuracy of brief case situation word segmentation, and avoids the problem that a part of the case type annotations of the cases in the process of manual processing of the cases are inaccurate or incorrect, thereby forming an obstacle for subsequent evidence guidance.

[0036] The automatic recommendation method provided by the present application can automatically recommend the evidence collection content of the case according to the case type and in combination with the historical cases of the same type, thereby helping the criminal police to make decisions in the evidence collection process, collect what kind of material evidence, perform what kind of investigation, identification, etc., and recommend how to perform the evidence collection mode, so as to avoid the important facts in the evidence collection process being ignored due to personal mistakes or limited experience.

[0037] The task flow method provided by the present application distributes the generated recommendation result to the corresponding police officer to form a to-be-completed task when the recommendation content of the case evidence collection content is generated, and the complete evidence chain and the main evidence information are all completed when the to-be-completed task is completed, thereby fundamentally solving the problem of unclear main criminal facts and the contradiction between the main evidence due to personal mistakes or limited experience.

[0038] The present application as a whole provides a method of combining an artificial intelligence recommendation engine with a task flow, which combines the evidence collection method and the evidence task flow of the case evidence collection process by using automatic recommendation, to solve the mistakes in collecting evidence of criminal cases in the traditional way relying on experience and manual recording, which leads to the case being returned by the procuratorate for supplementary investigation, or insufficient evidence due to mistakes, which leads to commutation or even exemption from prosecution, and effectively improves the case handling efficiency and the final prosecution rate of public security organs in criminal case investigation. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The flowchart of the present application. DETAILED DESCRIPTION

[0040] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0041] A criminal case evidence guidance method based on NLP technology, which adopts a method of combining an artificial intelligence recommendation engine with a task flow, combines the evidence collection method and the evidence task flow of the case evidence collection process by using automatic recommendation, to solve the mistakes in collecting evidence of criminal cases in the traditional way relying on experience and manual recording, which leads to the case being returned by the procuratorate for supplementary investigation, or insufficient evidence due to mistakes, which leads to commutation or even exemption from prosecution, and effectively improves the case handling efficiency and the final prosecution rate of public security organs in criminal case investigation.

[0042] Combination Figure 1 As shown, specifically comprising the following steps:

[0043] S1: case category judgment.

[0044] S2: initiate case category confirmation process.

[0045] S3: recommendation engine collects case information.

[0046] S4: generate recommendation results.

[0047] S5: task flow loads the recommendation results to form pending tasks.

[0048] Each step will be described in detail as follows:

[0049] S1: case category judgment.

[0050] The case category judgment is performed on the case to be filed, specifically, on the basis of improving the accuracy of brief case situation word segmentation, a text classification algorithm in NLP technology is used to predict the category of the case, to improve the accuracy of case category judgment, and the specific implementation method is:

[0051] S11: improve the accuracy of brief case situation word segmentation.

[0052] The word segmentation algorithm used in this invention is the conditional random field. The various indicators of the word segmentation accuracy rate of the conditional random field on the MSR (Microsoft Research Asia Named Entity Recognition Corpus) reach respectively: precision rate 96.86%, recall rate 96.64%, harmonic mean of precision rate and recall rate 96.75%, new word recall rate 71.54, registered word recall rate 97.33, which can fully meet the requirements of industrial applications.

[0053] In order to more accurately identify the proprietary terms in the cases, this invention selects to annotate a corpus in the field of public security to improve the word segmentation accuracy rate of the case details, and further improve the case classification accuracy rate. Considering the scale and quality of the corpus, the National Language Commission's Modern Balanced Corpus is used as the training model, and the annotation specification consistent with the National Language Commission's Modern Balanced Corpus - "Specification for Chinese Word Class Tagging Set for Information Processing" (GB / T 20532 - 2006) is selected to expand the corpus in the field of public security.

[0054] The raw corpus of this corpus in the field of public security comes from the basic case data in the public security industry. First, the morphological analysis of the raw corpus is carried out through the model generated by the National Language Commission's Modern Balanced Corpus, and the annotation is corrected in the output result to improve the annotation efficiency of the corpus in the field.

[0055] In the morphological analysis of the raw corpus, for the unregistered new words, their morphology needs to be manually marked. For example, in the output result, "packet loss fraud" is divided into "packet loss": gerund, "fraud": noun. At this time, it needs to be defined as a "packet loss fraud": noun in the output result of the raw corpus. Identify all such named entities in the field, and finally convert the raw corpus into a cooked corpus.

[0056] After the annotation is completed, the raw corpus becomes a cooked corpus, and the conditional random field is still used as the training algorithm to complete the training of the field model.

[0057] S12: Improve the case classification accuracy rate on the basis of improving the word segmentation accuracy rate.

[0058] After the word segmentation accuracy rate of the brief case details is improved, the text classification algorithm in NLP technology is used to predict the category of the case.

[0059] In order to maximize the accuracy rate of classification prediction, for the text classification algorithm in NLP technology, the "linear support vector machine classifier" is selected. This algorithm is a general algorithm, supported by corresponding papers, requires code implementation, and can also call the corresponding open-source software package.

[0060] In order to improve the operation efficiency of the whole NLP processing process, reduce the operation difficulty, the model management module needs to be implemented, the specific content is as follows: a sample training management module is established, the sample training management module includes "sample preview" and "sample upload" functions, which are used to manage the raw corpus required in the process of "improving the accuracy of brief case division", and the classification samples required by "linear support vector machine classifier".

[0061] S2: initiate case category confirmation process.

[0062] After the case category is judged, it will be assigned to the corresponding police officer to confirm the judged case category.

[0063] After the case category judgment and confirmation are completed, in order to realize the combination of artificial intelligence recommendation engine and task flow through steps S3, S4 and S5, realize the combination of evidence collection method and evidence collection task flow of automatic recommended case evidence collection process, the recommendation subject, the recommendation object and the recommendation algorithm need to be determined, as follows:

[0064] I. Determine the recommendation subject and the recommendation object:

[0065] The recommendation subject is the case to be handled; after the case category judgment is completed, the recommendation subject is generated.

[0066] The recommendation object is various types of material evidence; the eight types of evidence forms in the common evidence system of criminal cases are decomposed into the smallest granularity evidence elements, hereinafter referred to as evidence elements, as the recommendation object.

[0067] First, the material evidence

[0068] (1) The tool at the time of crime, such as the knife used for murder and the iron bar used for prying and pressing the safe during theft.

[0069] (2) The material object directly infringed by the criminal behavior, such as the stolen color TV, the robbed cash, the embezzled money, and the bribed small car.

[0070] (3) The article showing the social harm consequences of criminal behavior, such as damaged machine equipment, burned building, and stolen and cut telecommunication line.

[0071] (4) The illegal articles produced by criminal behavior, such as drugs, counterfeit currency, manufactured shotguns, and ammunition.

[0072] (5) The traces produced by criminal behavior, such as the traces left by prying the door.

[0073] (6) In the process of crime, the articles or traces left by the criminal in the crime scene, such as the clothes worn by the criminal, the hair torn off by the victim from the criminal suspect during rape, the fingerprints and footprints left by the criminal in the crime scene, etc.

[0074] (7) Various items and material traces forged by the criminal suspect to counter the investigation, such as leaving the clothes of someone else at the crime scene to divert attention.

[0075] (8) Items or traces that can prove the innocence of the criminal suspect or the defendant, such as a resident ID card used for hotel registration that is actually the criminal suspect's lost ID card.

[0076] (9) Other items or traces that can help determine the true circumstances of the case.

[0077] Second category: documentary evidence

[0078] (1) Contract

[0079] (2) Various public and private documents

[0080] (3) Lease agreement

[0081] (4) Marriage certificate

[0082] (5) Property deed

[0083] (6) Trademark

[0084] (7) Letter

[0085] (8) Telegram

[0086] (9) License plate

[0087] (10) Vehicle and ship tickets

[0088] (11) Various transportation documents

[0089] (12) Traffic accident responsibility identification

[0090] and so on, added according to the needs of the case.

[0091] Third category: witness testimony (consistent for all types of cases, no need to break down)

[0092] Fourth category: victim statement (consistent for all types of cases, no need to break down)

[0093] Fifth category: criminal suspect's confession and defense (consistent for all types of cases, no need to break down)

[0094] Sixth category: expert opinions

[0095] (1) Forensic identification: forensic pathological identification, forensic clinical identification, forensic psychiatric identification, forensic material evidence identification, forensic toxicological identification

[0096] (2) Material evidence identification: document identification, trace identification, and micro-identification

[0097] The seventh type: investigation, inspection, identification, and detection experiment records

[0098] (1) Investigation record

[0099] (2) Inspection record

[0100] (3) Identification record

[0101] (4) Detection experiment record

[0102] The eighth type: audio-visual materials and electronic data

[0103] (1) Email

[0104] (2) Electronic data exchange

[0105] (3) Online chat record

[0106] (4) Blog

[0107] (5) Micro-blog

[0108] (6) Mobile phone message

[0109] (7) Electronic signature

[0110] (8) Domain name

[0111] II. Determining the recommendation algorithm

[0112] After determining the subject and object of the recommendation, collect the content of the corresponding evidence materials of historical cases, and perform evidence material statistics and recommendation according to the predicted case category.

[0113] Recommendation priority calculation

[0114] Basic principle: The evidence materials of similar cases within a certain period of time are similar.

[0115] Classify historical cases according to the recommended category, and count the number of each type of evidence material used in the case.

[0116] First, prepare the case table tb_entity_criminalcase, which contains the case number (case_id) and case category (case_category_id).

[0117] Second, prepare the evidence element table tb_entity_evidence, which contains the case number (case_id), evidence category (evidence_category_id), and evidence element (evidence_element_id).

[0118] The case table is associated with the evidence element table, the evidence elements existing in the historical cases of the same category are counted according to the judged case category through a unary algorithm, and the proportion of all evidence elements in the cases of the respective case category is calculated, and the proportion value is the value range of the recommendation threshold.

[0119] In order to improve the accuracy of the recommended structure, the evidence element unary algorithm can be further improved to binary and ternary algorithms, that is, the correlation degree of the probability of the occurrence of one element in a case and the probability of the occurrence of another element, and the correlation degree high is recommended as the associated element.

[0120] The recommendation threshold is set to generate the recommendation information.

[0121] S3: The recommendation engine collects case information.

[0122] Step S3 associates the case table with the evidence element table according to the determination rules of the recommendation subject, the recommendation object and the recommendation algorithm, counts the evidence elements existing in the historical cases of the same category according to the judged case category, and completes the collection of case information.

[0123] S4: Generate a recommendation result.

[0124] Step S4 calculates the proportion of all evidence elements counted in step S3 in the cases of the respective case category according to the determination rules of the recommendation algorithm, the proportion value is the value range of the recommendation threshold, and generates a recommendation result, such as: for the case of Zhang San being robbed in Chaoyang district (case_id) type case, 90% contains the evidence of "material evidence" - "knife", please pay attention to collect.

[0125] S5: The task flow loads the recommendation result to form a to-do task.

[0126] After the recommendation result of the case evidence collection content is generated, the task flow loads the recommendation result, and according to the case information, the main and co-investigating officers of the case are obtained, and the generated recommendation result is distributed to the corresponding police officers, which can be realized by forming the to-do tasks of the main and co-investigating officers in the main information system used by the criminal police.

[0127] After the police complete the task, the task flow automatically obtains the evidence collection audit leader according to the organizational structure of the unit where the investigating officer is located, and the leader performs the audit, and after the leader completes the audit, the relevant regulations are sent to the personnel of the legal department for review.

[0128] When the present application is used, after all the to-do tasks are completed, the complete evidence chain and the main evidence information are all completed, which fundamentally solves the problems of unclear main criminal facts, contradictions between main evidence, and lack of criminal constitutive elements due to personal mistakes and limited experience, and improves the case investigation efficiency and the prosecution success rate.

[0129] The criminal case evidence guiding method based on the NLP technology described above is only one of the more general applications in the implementation of the present application, and in addition, the processing ideas and principles similar to the principles of the present application are also within the protection scope of the present application.

Claims

1. A criminal case evidence guiding method based on NLP technology, characterized in that: The method combines an artificial intelligence recommendation engine with a task flow, and realizes the combination of the evidence collection method of the automatically recommended case evidence collection process and the evidence task flow, and specifically comprises: Carrying out case category judgment; specifically, on the basis of improving the accuracy of brief case situation word segmentation, using a text classification algorithm in NLP technology to predict the category of the case, so as to improve the accuracy of case category judgment; The recommendation engine collects case information; Generating a recommendation result of case evidence collection content; The task flow loads the recommendation result to form a to-do task; The recommendation engine further comprises determining a recommendation subject, a recommendation object and determining a recommendation algorithm before collecting case information; The recommendation subject is a to-do case; after the case category judgment is completed, the recommendation subject is generated; The recommendation object is various types of material evidence; eight types of evidence forms in the common evidence system of criminal cases are decomposed into the smallest granularity evidence elements as the recommendation object; The recommendation algorithm is to classify historical cases according to the recommendation category, and count the number of various types of evidence materials used in the case, specifically comprising the following steps: First, prepare a case table, which contains case number and case category; Second, prepare an evidence element table, which contains case number, evidence category and evidence element; The recommendation engine collects case information by associating the case table with the evidence element table, and counting the evidence elements existing in the historical cases of the same category according to the judged case category through a unary algorithm; The unary algorithm for counting the evidence elements is further improved to a binary and ternary algorithm, that is, the correlation degree of the probability of the occurrence of one type of evidence element and the probability of the occurrence of another type of evidence element in one type of case, and the evidence elements with high correlation degree are counted as associated evidence elements; The recommendation result of the case evidence collection content is the proportion of all evidence elements in the case of the respective case category, and the proportion value is the recommendation threshold value range, and the recommendation result is generated.

2. The method of claim 1, wherein the method is based on NLP technology. The method for improving the accuracy of brief case situation word segmentation comprises using a conditional random field word segmentation algorithm to segment the case, and annotating a public security field corpus.

3. The method of claim 2, wherein the method is based on NLP technology. The public security field corpus uses the state language commission modern balanced corpus as a training model, and selects the annotation specification "Modern Chinese Word Class Labeling Set Specification for Information Processing" consistent with the state language commission modern balanced corpus to expand the public security field corpus; The raw corpus of the public security field corpus comes from the basic case data in the public security industry; first, the model generated by the state language commission modern balanced corpus is used for morphological analysis of the raw corpus, and the annotation is corrected in the output result to improve the annotation efficiency of the public security field corpus; In the morphological analysis of the raw corpus, new words that are not logged need to be manually marked with their morphology to convert the raw corpus into cooked corpus; Continue to use conditional random field as the training algorithm to complete the training of the public security field corpus.

4. The method of claim 3, wherein the method is characterized by: The text classification algorithm is a linear support vector machine classifier, and the classification samples required by the linear support vector machine classifier and the raw corpus required in the process of improving the accuracy of brief case situation word segmentation are managed through a sample training management module comprising a sample preview and sample upload function.

5. The method of claim 1, wherein the method is based on NLP technology. The to-do task is assigned to a corresponding police officer; after the police officer completes the task, the task flow automatically obtains an evidence collection auditing leader according to the organizational structure of the unit where the police officer is located, the leader performs auditing, and after the leader completes the auditing, the task flow sends the task to a legal department personnel for rechecking according to relevant regulations.

Images