Judicial expertise suggestion book generation method and device based on natural language model

Abstract

The invention provides a judicial appraisal suggestion book generation method and device based on a natural language model, and the method comprises the steps: obtaining an appraisal item request inputted by an appraisal person, determining an appraisal standard file according to the appraisal item request, obtaining an appraisal material inputted by the appraisal person, and determining the appraisal standard file according to the appraisal standard file, obtaining an identification abstract of the identification material through a text recognition algorithm; according to a first preset natural language model and the identification abstract, obtaining forensic clinical examination suggestions corresponding to the identification abstract; obtaining an examination result of forensic clinical examination, and obtaining an identification grade of the identified person according to a second preset natural language model, the examination result and the identification standard file; and finally, according to the identification standard file, the forensic clinical examination standard file, the examination result and the identification grade, generating an identification suggestion book. Based on the natural language model, the identification suggestion book is automatically generated, the time is shortened, the identification efficiency is improved, the error rate of the identification suggestion book is reduced, and judicial justice is guaranteed.

Description

Technical Field

[0001] The present application relates to the field of judicial identification technology, and in particular to a method and device for generating a judicial identification opinion based on a natural language model. Background Art

[0002] Forensic appraisal refers to the process by which a forensic expert applies scientific technology or specialized knowledge to identify and judge specialized issues involved in litigation and provide expert opinions. A client entrusts a qualified forensic expert institution with an appraisal and provides relevant test materials. The institution then conducts a series of forensic appraisal activities, including acceptance, implementation, review, and issuance, in accordance with the Ministry of Justice's "General Rules of Forensic Appraisal Procedure," and issues a forensic opinion. In the modern judicial system, forensic opinions are an important form of legal evidence, significantly influencing court decisions and litigation outcomes, and significantly impacting judicial credibility. Consequently, high standards are set for the authenticity, objectivity, accuracy, and professionalism of forensic opinions. These qualities, in turn, depend on the standardization, independence, objectivity, and impartiality of the forensic appraisal process.

[0003] Currently, most forensic appraisal institutions still employ a traditional appraisal model. The client provides the test materials, which the forensic appraisal institution receives and preserves. The forensic appraiser reviews the materials and conducts an on-site inspection of the individual being appraised. The appraiser then draws upon relevant appraisal standards and personal experience to form an appraisal opinion. After the appraisal is complete, the appraiser manually extracts key information from the extensive collection of test materials and related documentation, compiles relevant information from the appraisal process, and prepares a forensic appraisal opinion. The forensic appraisal opinion is then reviewed for accuracy, completeness, consistency, and compliance.

[0004] However, under the traditional appraisal model, there is a lot of manual participation in the process from the client providing the appraisal materials to obtaining the judicial appraisal opinion. For example, the appraisal opinion relies on relevant appraisal standards and the experience of the judicial appraiser, or the judicial appraisal opinion needs to be manually compiled, which makes the judicial appraisal time-consuming, inefficient, and prone to errors, affecting judicial fairness. Summary of the Invention

[0005] The present application provides a method and device for generating a judicial appraisal opinion based on a natural language model to solve the technical problems mentioned in the background technology.

[0006] In a first aspect, the present application provides a method for generating a judicial appraisal opinion based on a natural language model, characterized by comprising: Obtaining an appraisal request input by the person being appraised, and determining an appraisal standard document based on the appraisal request, wherein the appraisal request includes appraisal matters; Obtaining the identification materials input by the person being identified, and obtaining an identification summary of the identification materials through a text recognition algorithm; Obtaining a forensic clinical examination opinion corresponding to the identification summary based on a first preset natural language model and the identification summary, wherein the first preset natural language model is trained based on a pre-trained language BERT model, a glyph embedding model, and a global pointer model; Obtaining the results of a forensic clinical examination, and obtaining an identification grade of the person being identified based on a second preset natural language model, the examination results, and the identification standard document, wherein the second preset natural language model is trained based on a pre-trained language BERT model, a glyph embedding model, and a global pointer model; An appraisal opinion is generated based on the appraisal standard document, the forensic clinical examination standard document, the inspection results and the appraisal level.

[0007] Optionally, obtaining the identification summary of the identification material by a text recognition algorithm includes: Obtaining text content of the identification material according to optical character recognition; The identification summary is generated according to the text content of the identification material.

[0008] Optionally, before obtaining the appraisal level of the person being appraised based on the second preset natural language model, the inspection result, and the appraisal standard document, the method further includes: Obtain historical data related to forensic identification; Label and classify historical data; The second preset natural language model is obtained based on the annotated historical data and the natural language model to be trained corresponding to the second preset natural language model. The natural language model to be trained is a model that integrates the BERT model, the glyph embedding model and the global pointer model.

[0009] Optionally, obtaining the second preset natural language model according to the annotated historical data and the natural language model to be trained corresponding to the second preset natural language model includes: Inputting the annotated historical data into the natural language model to be trained; Encoding the annotated historical data according to the BERT model and the font embedding model to obtain encoding feature data corresponding to the annotated historical data; Decoding the processed encoded feature data according to the global pointer model to obtain an identification level output value of the natural language model to be trained; The natural language model to be trained is trained according to the identification level output value and the identification level annotation value of the annotated historical data to obtain the second preset natural language model.

[0010] Optionally, encoding the annotated historical data according to the BERT model and the font embedding model to obtain encoded feature data corresponding to the annotated historical data includes: Obtaining first encoding feature data according to the BERT model and the annotated historical data; Obtaining second encoding feature data according to the font embedding model and the annotated historical data; The encoding feature data is obtained according to the first encoding feature data and the second encoding feature data.

[0011] Optionally, after obtaining the identification materials input by the person being identified, the method further includes: The identification materials are stored on the blockchain.

[0012] Optionally, generating an appraisal opinion based on the appraisal standard document, the forensic clinical examination standard document, the inspection result, and the appraisal level includes: The identification standard document, the forensic clinical examination standard document, the inspection result and the identification grade are input into an identification opinion template to generate an identification opinion.

[0013] In a second aspect, the present application provides a device for generating a judicial appraisal opinion based on a natural language model, comprising: The identification item acquisition module is used to obtain the identification item request input by the identification person and determine the identification standard document according to the identification item request. The identification item request includes the identification items; The identification summary acquisition module is used to obtain the identification materials input by the person being identified and obtain the identification summary of the identification materials through the text recognition algorithm; An inspection opinion acquisition module is used to obtain a forensic clinical inspection opinion corresponding to the identification summary based on a first preset natural language model and the identification summary, where the first preset natural language model is trained based on a pre-trained language BERT model, a font embedding model, and a global pointer model; An identification grade acquisition module, configured to obtain the results of a forensic clinical examination and, based on a second preset natural language model, the examination results, and the identification standard document, obtain the identification grade of the person being identified. The second preset natural language model is trained based on a pre-trained language BERT model, a glyph embedding model, and a global pointer model. The generation module is used to generate an appraisal opinion based on the appraisal standard document, the forensic clinical examination standard document, the inspection results and the appraisal level.

[0014] Optionally, when the identification summary acquisition module obtains the identification summary of the identification material through a text recognition algorithm, it is specifically used to: Obtaining text content of the identification material according to optical character recognition; The identification summary is generated according to the text content of the identification material.

[0015] Optionally, it also includes: training module; Before the identification level acquisition module obtains the identification level of the identified person based on the second preset natural language model, the inspection result and the identification standard document, the training module is used to: Obtain historical data related to forensic identification; Label and classify historical data; The second preset natural language model is obtained based on the annotated historical data and the natural language model to be trained corresponding to the second preset natural language model. The natural language model to be trained is a model that integrates the BERT model, the glyph embedding model and the global pointer model.

[0016] Optionally, when the training module obtains the second preset natural language model based on the annotated historical data and the natural language model to be trained corresponding to the second preset natural language model, it is specifically configured to: Inputting the annotated historical data into the natural language model to be trained; Encoding the annotated historical data according to the BERT model and the font embedding model to obtain encoding feature data corresponding to the annotated historical data; Decoding the processed encoded feature data according to the global pointer model to obtain an identification level output value of the natural language model to be trained; The natural language model to be trained is trained according to the identification level output value and the identification level annotation value of the annotated historical data to obtain the second preset natural language model.

[0017] Optionally, the training module encodes the annotated historical data according to the BERT model and the font embedding model to obtain encoded feature data corresponding to the annotated historical data, specifically for: Obtaining first encoding feature data according to the BERT model and the annotated historical data; Obtaining second encoding feature data according to the font embedding model and the annotated historical data; The encoding feature data is obtained according to the first encoding feature data and the second encoding feature data.

[0018] Optionally, it also includes: a storage module; After the identification item acquisition module acquires the identification materials input by the person being identified, the storage module is used to: The identification materials are stored on the blockchain.

[0019] Optionally, when the generation module generates the appraisal opinion based on the appraisal standard document, the forensic clinical examination standard document, the inspection result, and the appraisal level, it is specifically used to: The identification standard document, the forensic clinical examination standard document, the inspection result and the identification grade are input into an identification opinion template to generate an identification opinion.

[0020] In a third aspect, the present application provides an electronic device including a processor and a memory; Memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory, so that the processor performs any method according to the first aspect.

[0021] In a fourth aspect, an embodiment of the present application provides a readable storage medium, including a program or instruction. When the program or instruction runs on a computer, any method of the above-mentioned first aspect is executed.

[0022] In a fifth aspect, an embodiment of the present application provides a computer program product, comprising a computer program, which implements the method described in any one of the first aspects when executed by a processor.

[0023] The method and device for generating a judicial appraisal opinion based on a natural language model provided by the present application obtains an appraisal request input by the person being appraised, and determines an appraisal standard document based on the appraisal request, wherein the appraisal request includes the appraisal matters, and obtains the appraisal materials input by the person being appraised, and obtains an appraisal summary of the appraisal materials through a text recognition algorithm; then, based on a first preset natural language model and the appraisal summary, obtains the forensic clinical examination opinion corresponding to the appraisal summary; then, obtains the inspection results of the forensic clinical examination, and obtains the appraisal level of the person being appraised based on a second preset natural language model, the inspection results and the appraisal standard document; finally, generates an appraisal opinion based on the appraisal standard document, the forensic clinical examination standard document, the inspection results and the appraisal level. The present application automatically generates an appraisal opinion based on a natural language model, shortens the appraisal time, improves the appraisal efficiency, and reduces the error rate of the appraisal opinion. Moreover, since the degree of manual participation in the appraisal is reduced, the probability of manual contact with the appraisal materials is reduced, thereby ensuring judicial fairness. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0025] Figure 1 A schematic diagram of the structure of a forensic expert opinion generation system based on a natural language model provided in one embodiment of the present application; Figure 2 A flowchart of a method for generating a judicial appraisal opinion based on a natural language model provided in one embodiment of the present application; Figure 3 A flowchart of a preset natural language model training method provided in one embodiment of the present application; Figure 4 A structural diagram of the BERT model provided in one embodiment of the present application; Figure 5 A structural diagram of a font embedding model provided in one embodiment of the present application; Figure 6 A structural diagram of a natural language model provided in one embodiment of the present application; Figure 7 A schematic diagram of the structure of a forensic expert opinion generation device based on a natural language model provided in one embodiment of the present application; Figure 8 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0026] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application are clearly and completely described below. Obviously, the described embodiments are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts also fall within the scope of protection of this application.

[0027] Currently, forensic appraisal institutions still employ a traditional appraisal model. The client provides the inspection materials, which the forensic appraisal institution receives and preserves. The forensic appraiser reviews the materials and conducts an on-site inspection of the individual being appraised, drawing upon relevant appraisal standards and personal experience to form an appraisal opinion. After the appraisal is complete, the appraiser manually extracts key information from the extensive inspection materials and related documentation, compiles relevant information from the appraisal process, and prepares a forensic appraisal opinion. The forensic appraisal opinion is then reviewed for accuracy, completeness, consistency, and compliance.

[0028] However, the traditional appraisal model involves a lot of manual participation from the client providing the appraisal materials to obtaining the judicial appraisal opinion. For example, the appraisal opinion relies on relevant appraisal standards and the experience of the judicial appraiser, or the judicial appraisal opinion needs to be manually compiled, which makes the judicial appraisal time-consuming, inefficient, and error-prone. In addition, there is a risk of tampering due to the high manual participation, which will affect judicial fairness.

[0029] Therefore, in order to solve the technical problems existing in the prior art, the present application proposes a method and device for generating a judicial appraisal opinion based on a natural language model. For the appraisal materials, an appraisal summary of the appraisal materials is obtained through a text recognition algorithm. The appraisal summary is analyzed according to the natural language model to obtain the forensic clinical examination opinion corresponding to the appraisal summary. Then, for the inspection results of the forensic clinical examination, the appraisal level of the person being appraised is obtained according to the natural language model and the said appraisal standard document. Finally, an appraisal opinion is generated based on the appraisal standard document, the forensic clinical examination standard document, the inspection results and the appraisal level. The appraisal time is shortened, the appraisal efficiency is improved, and the error rate of the appraisal opinion is reduced. In addition, since the degree of manual participation in the appraisal is reduced, the probability of manual contact with the appraisal materials is reduced, thereby ensuring judicial fairness.

[0030] Figure 1 This is a structural diagram of a forensic expert opinion generation system based on a natural language model provided in one embodiment of the present application. Figure 1 As shown, the system includes a server 100 and a terminal device 200. The terminal device 200 is provided with a website or application software. A user, such as the person being appraised in this application, operates the terminal device and obtains an appraisal opinion through the website or application software. The specific process of obtaining the appraisal opinion is described below.

[0031] Figure 2 This is a flow chart of a method for generating a judicial appraisal opinion based on a natural language model provided in one embodiment of the present application. Figure 2 As shown, the method includes: S201. Obtain the appraisal request input by the person to be appraised, and determine the appraisal standard document based on the appraisal request.

[0032] Among them, the request for appraisal matters includes appraisal matters.

[0033] In this step, the person to be identified may refer to the user who inputs the identification request. The user may be the real person to be identified or his / her client. For the sake of convenience, the user who inputs the identification request is referred to as the person to be identified.

[0034] This application embodiment uses work injury assessment as an example to illustrate: The person being assessed accesses the application software used to generate the assessment opinion on a terminal device. First, they select an assessment request on the terminal's interface. The interface displays options for all assessment items related to work-related injury assessments. The person being assessed selects one based on their assessment needs. For example, if they select "Assessment of the Degree of Disability Caused by Human Injury" on the interface, they have entered their assessment request.

[0035] After the server receives the "Human Body Injury Disability Degree Assessment" item selected by the person being assessed, it determines the assessment standard document as the "Human Body Injury Disability Degree Grading" based on the "Human Body Injury Disability Degree Assessment" item.

[0036] It should be noted that the server pre-sets the correspondence between each appraisal item and each appraisal standard file, so that after obtaining the appraisal item request input by the person being appraised, the server can determine the appraisal standard file according to the appraisal item.

[0037] S202: Obtain the identification materials input by the person being identified, and obtain an identification summary of the identification materials through a text recognition algorithm.

[0038] In this step, the person being assessed uploads assessment materials through the application software, where the assessment materials may include: the person being assessed's examination report, the person being assessed's hospitalization medical records, the person being assessed's discharge report, as well as a diagnosis certificate and a letter of authorization for work-related injury assessment.

[0039] Among them, since the uploaded appraisal materials are in picture format, the text recognition algorithm is used to perform text recognition on the content of the appraisal materials in picture format to obtain the appraisal summary corresponding to the appraisal materials. The appraisal summary includes the main content related to the work injury appraisal in the appraisal materials.

[0040] For example, based on the examination report of the person being appraised, the person's hospitalization medical record, the person's discharge report and the diagnosis certificate, the appraisal summary obtained includes: the basic information of the person being appraised, such as name, age, gender, as well as the process of the work injury, the admission diagnosis result, the operation content, the discharge diagnosis content, etc. Specifically, for example: Appraisal summary: The person being appraised was injured while working at a certain time, and went to the local hospital outpatient clinic for examination and treatment after the injury. Admission diagnosis: 1. Right femoral neck fracture; 2. Right hip contusion. After admission, the injured person underwent "right hip total hip replacement + hip synovectomy" on a certain day and was discharged on a certain day. The discharge diagnosis: 1. Femoral neck fracture; 2. Contusion.

[0041] Optionally, after obtaining the appraisal materials, since the appraisal materials are information uploaded by the person being appraised and are the fundamental data for judicial appraisal, blockchain technology is used for storage to prevent tampering, thereby improving confidentiality and storage security.

[0042] Optionally, a specific implementation of S202 is: S2021. Obtain text content of the identification material based on optical character recognition.

[0043] Specifically, the text content of the identification material is recognized according to optical character recognition (OCR) to obtain the text content.

[0044] S2022. Generate an appraisal summary based on the text content of the appraisal materials.

[0045] Specifically, the text content of the appraisal materials is typeset according to the preset format and content requirements of the appraisal summary to generate an appraisal summary.

[0046] S203. Obtain a forensic clinical examination opinion corresponding to the appraisal summary based on the first preset natural language model and the appraisal summary.

[0047] Among them, the first preset natural language model is obtained based on the pre-trained language BERT model, the font embedding model and the global pointer model training.

[0048] In this step, the first preset natural language model integrates the algorithms of the BERT model, the glyph embedding model, and the global pointer model. The BERT model and the glyph embedding model are used to encode the appraisal summary and extract its feature information. This feature information is then input into the corresponding decoding layer of the global pointer model. The global pointer model decodes the feature information of the appraisal summary to obtain the content of the appraisal summary, thereby obtaining keywords, and determining the forensic clinical examination opinion based on the keywords. The training process of the first preset natural language model is detailed below.

[0049] The first preset natural language model is used to analyze the appraisal summary, understand the work-related injury information of the person being appraised recorded in the appraisal summary, and obtain the keywords of the work-related injury information of the person being appraised, so as to determine the forensic clinical examination opinion corresponding to the person being appraised based on the keywords of the work-related injury information of the person being appraised and the forensic clinical examination opinions corresponding to the pre-set keywords of each work-related injury information.

[0050] For example, for the appraisal summary in the example of S202, the first preset natural language model analyzes the appraisal summary to obtain keywords of the work-related injury information of the person being appraised recorded in the appraisal summary, such as: femoral neck fracture, contusion. The forensic clinical examination opinion determined based on the keywords is: spinal injury, and a forensic examination must be performed according to the "Forensic Clinical Examination Standards" and the examination method for spinal injury.

[0051] S204: Obtain the examination results of the forensic clinical examination, and obtain the identification level of the person being identified based on the second preset natural language model, the examination results, and the identification standard document.

[0052] In this step, the second preset natural language model is obtained based on the pre-trained language BERT model, the font embedding model and the global pointer model.

[0053] After the forensic clinical examination opinion is determined in S203, the forensic doctor performs a forensic examination on the person to be examined according to the forensic clinical examination opinion and records the examination result on the application software so that the server can obtain the examination result.

[0054] The second preset natural language model analyzes the inspection results, obtains keywords in the inspection results, compares the keywords with the work-related injury information corresponding to the work-related injury level recorded in the appraisal standard file, and obtains the appraisal level of the person being appraised.

[0055] For example, the keywords recorded in the inspection results obtained by the second preset natural language model include "right femoral neck and femoral head base fracture undergoing "artificial total hip replacement"". Based on the comparison of this keyword with the "Classification of Disability of Human Injury", the appraisal level of the person being appraised is "According to the nine levels of disability, paragraph 5.9.6, and Article 5 of the "Classification of Disability of Human Injury", "after undergoing joint prosthesis replacement of any major joint of the limbs", the disability level of the person being appraised is assessed as level nine (nine)."

[0056] S205. Generate an appraisal opinion based on the appraisal standard document, the forensic clinical examination standard document, the inspection results, and the appraisal level.

[0057] In this step, according to the format and content requirements in the appraisal opinion, the appraisal standard documents, forensic clinical examination standard documents, inspection results and appraisal level contents are obtained, and an appraisal opinion letter is generated based on the appraisal standard documents, forensic clinical examination standard documents, inspection results and appraisal level contents.

[0058] Specifically, the identification standard document, the forensic clinical examination standard document, the inspection results and the identification level are doubled and input into the identification opinion template, wherein the identification opinion template stipulates the format and content of the identification opinion, thereby generating an identification opinion.

[0059] In this embodiment, the identification request input by the person to be identified is obtained, and the identification standard document is determined based on the identification request, the identification request includes the identification items, and the identification materials input by the person to be identified are obtained, and the identification summary of the identification materials is obtained through a text recognition algorithm; then, based on the first preset natural language model and the identification summary, the forensic clinical examination opinion corresponding to the identification summary is obtained; then, the inspection results of the forensic clinical examination are obtained, and the identification level of the person to be identified is obtained based on the second preset natural language model, the inspection results and the identification standard document; finally, an identification opinion is generated based on the identification standard document, the forensic clinical examination standard document, the inspection results and the identification level. This application automatically generates an identification opinion based on a natural language model, shortens the identification time, improves the identification efficiency, and reduces the error rate of the identification opinion. In addition, since the degree of manual participation in the identification is reduced, the probability of manual contact with the identification materials is reduced, thereby ensuring judicial fairness.

[0060] Figure 3 This is a flow chart of a method for training a preset natural language model according to an embodiment of the present application. Figure 3 As shown, the training methods include: S301. Obtain historical data related to forensic identification.

[0061] In this step, a large amount of historical data from the field of forensic identification, especially work-related injury identification data, such as legal documents, judgments, and identification reports, is collected. This historical data is then collated to obtain valid historical data. It should be noted that the historical data mentioned later is the valid historical data obtained after data collation.

[0062] S302: Label and classify historical data.

[0063] In this step, the historical data is labeled and classified, that is, the identification level label value corresponding to the historical data is marked according to the content of the historical data, and the historical data is classified according to the identification level, and historical data with the same identification level are divided into one category.

[0064] S303: Obtain a second preset natural language model based on the annotated historical data and the natural language model to be trained corresponding to the second preset natural language model.

[0065] Among them, the natural language model to be trained is a model that integrates the BERT model, the font embedding model and the global pointer model.

[0066] In this step, the annotated historical data is input into the natural language model to be trained, the natural language model to be trained analyzes the annotated historical data, obtains the keywords of the annotated historical data, determines the identification level output value based on the keywords, and obtains the second preset natural language model based on the identification level output value and the identification level annotation value of the annotated historical data.

[0067] Specifically, a possible implementation method of S302 is: S3021. Input the labeled historical data into the natural language model to be trained.

[0068] Specifically, the annotated historical data is input into the natural language model to be trained, and the natural language model to be trained extracts features of the annotated historical data.

[0069] S3022. Encode the annotated historical data according to the BERT model and the font embedding model to obtain encoded feature data corresponding to the annotated historical data.

[0070] Specifically, since the annotated historical data is in text format, computers cannot recognize historical data in text format. Therefore, it is necessary to encode the annotated historical data and encode the historical data into a language that can be recognized by computers.

[0071] Because historical data related to forensic identification, especially forensic identification of work-related injuries, involves the medical field, the corresponding historical data contains many entities, and the boundaries between entities overlap, making it difficult to identify and label the data. For example, "thymocyte carcinoma" has an entity nesting problem. The thymus is a lymphatic organ in the human body, thymocytes are medical cells, and thymocyte carcinoma represents a disease condition. Medical texts also have long sentences with large spans and abundant clauses, making analysis difficult. For example, "Spherocytes are common in hereditary spherocytosis and neonatal hemolytic disease with spherocytosis and hemolytic anemia caused by erythrocyte enzyme deficiency." The professional terms have certain similarities, but the meanings they express are different and refer to different individuals. Spherocytes are abnormal red blood cells, hereditary spherocytosis represents a disease symptom, and hemolytic disease and hemolytic anemia represent different anemia diseases.

[0072] Therefore, the quality of medical entity recognition has a significant impact on subsequent relationship extraction and graph construction. Therefore, it is necessary to improve the accuracy of medical entity recognition.

[0073] Therefore, the embodiments of this application adopt a BERT model and a glyph embedding model. By leveraging the bidirectionality of the BERT model, both the前文 and后文 of the sentence are considered simultaneously to learn long-distance medical texts. Moreover, due to the characteristics of medical texts, namely, the radical features of medical texts are relatively obvious, and there is a strong connection between the character structure and meaning. For example, body parts mostly have the radical "月", and most diseases have the radical "疒". Therefore, the glyph embedding model is used to learn the radical structure information of Chinese characters to compensate for the deficiency of the BERT model in processing morphological information. Therefore, when encoding the labeled historical data, the BERT model and the glyph embedding model are used for encoding, so that the encoded feature data fuses different levels of language features.

[0074] Optionally, a specific implementation manner of S3022 is as follows: S30221. Obtain the first encoded feature data according to the BERT model and the labeled historical data.

[0075] Specifically, as Figure 4 shown, the BERT model is mainly composed of deep and bidirectional Transformer units (Trm). Each part of the Trm contains various types of elements, such as multi-head attention, feed-forward neural network, residual connection, and normalization.

[0076] The multi-head attention mechanism divides the input sequence into multiple heads. Each head can calculate the attention weights in parallel, and then fuse the attention weights of multiple heads to obtain the final self-attention representation, which helps the model capture dependencies at different distances and improve the model's representation ability.

[0077] The feed-forward neural network layer performs non-linear transformation on the representation of the input sequence through a multi-layer perception mechanism to obtain higher-level abstract features, uses residual connection to handle the problem of small gradients, and at the same time uses normalization on the output of each layer to improve the stability and robustness of the model.

[0078] Therefore, the BERT model can learn general language features from a large amount of corpus and has stronger generalization ability in entity recognition.

[0079] Therefore, the BERT model extracts features from the labeled historical data to obtain the first encoded feature data. This first encoded feature data facilitates understanding the sentence context during entity recognition and improves the speed of entity recognition in the sentence.

[0080] S30222. Obtain the second encoded feature data according to the glyph embedding model and the labeled historical data.

[0081] Specifically, the glyph embedding model can be a convolutional neural network CNN, as Figure 5As shown in the figure, the character embedding model includes a radical embedding layer, a convolutional layer, and a maximum pooling layer. The radical embedding layer is the first layer, which is used to convert the radical information of Chinese characters into a vector representation. For example, a random initialization method is used to generate a radical embedding vector from a predefined range (such as a uniform distribution of [-1, 1]). The convolution layer is the core layer, used to extract features from the character features of Chinese characters. It uses convolution kernels of different sizes to capture local features of different sizes and performs convolution operations on the radical embedding vectors of Chinese characters to generate convolution features. The maximum pooling layer downsamples the convolutional features, selects the most significant eigenvalue in each convolutional feature, and uses it as the pooling result of the feature, reducing the feature dimension and the number of parameters while retaining important features and outputting the final vector representation of the word.

[0082] In other words, the font embedding model extracts the radicals of Chinese characters as font embeddings, combines them with the character embeddings themselves, and incorporates label information through iterative learning for named entity recognition, making full use of font features and text information.

[0083] Therefore, the Chinese characters in the annotated historical data are converted into radical encoding through the glyph embedding model. Each Chinese character is encoded into a vector segment through the embedding layer, convolution layer and maximum pooling layer, namely the second encoding feature data. The second encoding feature data contains the radicals, glyph structure morphology and meaning of the medical text.

[0084] S30223. Obtain coding feature data according to the first coding feature data and the second coding feature data.

[0085] Specifically, the first coding feature data and the second coding feature data encode the historical data from different perspectives. Therefore, the first coding feature data and the second coding feature data are combined. Specifically, Figure 6 As shown, the first coding data and the second coding feature data are both in vector form, and the first coding feature data and the second coding feature data are concatenated to obtain the coding feature data.

[0086] Based on radical embedding, the radical information of Chinese characters is extracted as glyph embedding, and combined with the vector obtained by BERT encoding to form a richer and more contextual input vector, which can capture the contextual information and morphological information of medical texts, thereby more accurately identifying entities in medical texts.

[0087] S3023. Decode the encoded feature data after processing according to the global pointer model to obtain an identification level output value of the natural language model to be trained.

[0088] Specifically, after encoding the historical data to obtain encoding feature data, the computer processes the encoding feature data and obtains "keywords" in the historical data through the encoding feature data. The "keywords" are computer-recognizable languages, and therefore need to be decoded. In this embodiment, decoding is performed using a global pointer model.

[0089] The global pointer model uses global pointers to capture contextual relationships between entities. By annotating each entity with a pointer vector containing the entity's starting and ending positions within a sentence, as well as its type, it can simultaneously identify entities within a sentence, identify nested entities within a sentence, and classify entity types. Furthermore, using a single pointer vector at the beginning and end of an entity enhances internal connections and provides a more global perspective.

[0090] Therefore, if Figure 6 As shown, the processed coded feature data is decoded through the global pointer model to obtain the entities in the sentence, thereby determining the keywords corresponding to the sentence based on the obtained entities, and determining and outputting the identification level output value based on the keywords.

[0091] S3024. Train the natural language model to be trained based on the identification level output value and the identification level annotation value of the annotated historical data to obtain a second preset natural language model.

[0092] Specifically, for the natural language model to be trained, due to the model error, especially the natural language model in the training stage, the identification level output value it outputs is not necessarily the identification level labeling value. There is a difference between the identification level output value and the identification level labeling value. The natural language model to be trained is trained based on the difference to obtain a second preset natural language model.

[0093] It should be noted that the training process of the first preset natural language model is the same as the training process of the second preset natural language model. According to the different model output values, when annotating and assigning historical data, classification is performed according to the model output values. The training process of the first preset natural language model will not be repeated here.

[0094] In this embodiment, when designing a natural language model, BERT is integrated at the encoding layer and combines the radical features of Chinese characters to enhance the domain expressiveness of the text. It can capture the global context information of the sentence and obtain the encoded feature data as the input of the natural language model. At the decoding layer, entity recognition is treated as a global labeling problem through global pointers, and the entity head and tail are recognized as a whole, which improves the accuracy of sentence recognition. This improves the efficiency and accuracy of obtaining appraisal opinions.

[0095] It should be noted that the training method of the first preset natural language model can refer to Figure 3 The embodiments shown are not described in detail here.

[0096] Figure 7 This is a schematic diagram of the structure of a forensic expert opinion generation device based on a natural language model provided in one embodiment of the present application. Figure 7 As shown, the device includes: an appraisal matter acquisition module 710, an appraisal summary acquisition module 720, an inspection opinion acquisition module 730, an appraisal level acquisition module 740 and a generation module 750.

[0097] Optionally, the device further includes: a training module 760.

[0098] Optionally, the device further includes: a storage module 770.

[0099] The appraisal item acquisition module 710 is used to obtain the appraisal item request input by the appraised person and determine the appraisal standard document according to the appraisal item request, wherein the appraisal item request includes the appraisal items; The appraisal summary acquisition module 720 is used to obtain the appraisal materials input by the appraised person and obtain the appraisal summary of the appraisal materials through a text recognition algorithm; The inspection opinion acquisition module 730 is used to obtain a forensic clinical inspection opinion corresponding to the identification summary based on a first preset natural language model and the identification summary, where the first preset natural language model is trained based on a pre-trained language BERT model, a font embedding model, and a global pointer model; Identification level acquisition module 740 is used to obtain the examination results of the forensic clinical examination and obtain the identification level of the person being identified based on a second preset natural language model, the examination results, and the identification standard document. The second preset natural language model is trained based on the pre-trained language BERT model, the font embedding model, and the global pointer model. The generation module 750 is used to generate an appraisal opinion based on the appraisal standard document, the forensic clinical examination standard document, the inspection results and the appraisal level.

[0100] Optionally, when the appraisal summary obtaining module 720 obtains the appraisal summary of the appraisal material through a text recognition algorithm, it is specifically used to: Obtaining text content of the identification material according to optical character recognition; The identification summary is generated according to the text content of the identification material.

[0101] Optionally, before the identification level acquisition module 740 obtains the identification level of the identified person based on the second preset natural language model, the inspection result, and the identification standard document, the training module 760 is configured to: Obtain historical data related to forensic identification; Label and classify historical data; The second preset natural language model is obtained based on the annotated historical data and the natural language model to be trained corresponding to the second preset natural language model. The natural language model to be trained is a model that integrates the BERT model, the glyph embedding model and the global pointer model.

[0102] Optionally, when the training module 760 obtains the second preset natural language model based on the annotated historical data and the natural language model to be trained corresponding to the second preset natural language model, it is specifically configured to: Inputting the annotated historical data into the natural language model to be trained; Encoding the annotated historical data according to the BERT model and the font embedding model to obtain encoding feature data corresponding to the annotated historical data; Decoding the processed encoded feature data according to the global pointer model to obtain an identification level output value of the natural language model to be trained; The natural language model to be trained is trained according to the identification level output value and the identification level annotation value of the annotated historical data to obtain the second preset natural language model.

[0103] Optionally, the training module 760 encodes the annotated historical data according to the BERT model and the font embedding model to obtain encoded feature data corresponding to the annotated historical data, specifically for: Obtaining first encoding feature data according to the BERT model and the annotated historical data; Obtaining second encoding feature data according to the font embedding model and the annotated historical data; The encoding feature data is obtained according to the first encoding feature data and the second encoding feature data.

[0104] Optionally, after the appraisal item acquisition module 710 acquires the appraisal materials input by the person being appraised, the storage module 770 is configured to: The identification materials are stored on the blockchain.

[0105] Optionally, when generating an appraisal opinion based on the appraisal standard document, the forensic clinical examination standard document, the inspection result, and the appraisal level, the generating module 750 is specifically configured to: The identification standard document, the forensic clinical examination standard document, the inspection result and the identification grade are input into an identification opinion template to generate an identification opinion.

[0106] The forensic appraisal opinion generation device based on the natural language model provided in the embodiment of the present application can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar and will not be repeated here.

[0107] Figure 8 Schematic diagram of the hardware structure of the electronic device provided in the embodiment of the present application. Figure 8 As shown, the electronic device includes: a processor 810 and a memory 820, The processor 810 and the memory 820 are connected via a bus 830 .

[0108] During the specific implementation process, the processor 810 executes the computer-executable instructions stored in the memory 820, so that the processor 810 performs the above method.

[0109] The specific implementation process of the processor 810 can be found in the above method embodiment. Its implementation principle and technical effects are similar and will not be repeated here in this embodiment.

[0110] In the above Figure 8 In the illustrated embodiment, it should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the application may be directly executed by a hardware processor or by a combination of hardware and software modules in the processor.

[0111] The memory may include high-speed RAM memory and may also include non-volatile storage NVM, such as disk storage.

[0112] A bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. Buses can be categorized as address buses, data buses, and control buses. For ease of illustration, the buses in the drawings of this application are not limited to just one bus or just one type of bus.

[0113] The present application also provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the above method is implemented.

[0114] The computer-readable storage medium mentioned above can be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The computer-readable storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0115] An exemplary readable storage medium is coupled to a processor so that the processor can read information from the readable storage medium and write information to the readable storage medium. Of course, the readable storage medium can also be an integral part of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the readable storage medium can also exist in the device as discrete components.

[0116] Those skilled in the art will appreciate that all or part of the steps in the above-described method embodiments can be implemented using hardware associated with program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for generating a judicial appraisal opinion based on a natural language model, characterized in that: include: Obtaining an appraisal request input by the person to be appraised, and determining an appraisal standard document according to the appraisal request, wherein the appraisal request includes appraisal matters; Obtaining the identification materials input by the person to be identified, and obtaining the identification summary of the identification materials through a text recognition algorithm; Obtaining a forensic clinical examination opinion corresponding to the identification summary according to a first preset natural language model and the identification summary, wherein the first preset natural language model is obtained by training based on a pre-trained language BERT model, a glyph embedding model, and a global pointer model; Obtaining the inspection results of the forensic clinical examination, and obtaining the identification level of the identified person according to a second preset natural language model, the inspection results and the identification standard document, wherein the second preset natural language model is obtained by training based on a pre-trained language BERT model, a font embedding model and a global pointer model; An appraisal opinion is generated based on the appraisal standard document, the forensic clinical examination standard document, the inspection results and the appraisal level.

2. The method according to claim 1, characterized in that: The step of obtaining the identification summary of the identification material by using a text recognition algorithm includes: Obtaining text content of the identification material according to optical character recognition; The identification summary is generated according to the text content of the identification material.

3. The method according to claim 1, characterized in that Before obtaining the appraisal level of the person to be appraised according to the second preset natural language model, the inspection result and the appraisal standard file, the method further includes: Obtain historical data related to forensic identification; Label and classify historical data; According to the annotated historical data and the natural language model to be trained corresponding to the second preset natural language model, the second preset natural language model is obtained, and the natural language model to be trained is a model that integrates the BERT model, the glyph embedding model and the global pointer model.

4. The method according to claim 3, characterized in that: The step of obtaining the second preset natural language model according to the annotated historical data and the natural language model to be trained corresponding to the second preset natural language model includes: Inputting the annotated historical data into the natural language model to be trained; Encode the annotated historical data according to the BERT model and the font embedding model to obtain encoding feature data corresponding to the annotated historical data; Decoding the processed coded feature data according to the global pointer model to obtain an identification level output value of the natural language model to be trained; The natural language model to be trained is trained according to the identification level output value and the identification level annotation value of the annotated historical data to obtain the second preset natural language model.

5. The method according to claim 4, characterized in that The encoding of the annotated historical data according to the BERT model and the font embedding model to obtain the encoded feature data corresponding to the annotated historical data includes: Obtaining first encoding feature data according to the BERT model and the annotated historical data; Obtaining second encoding feature data according to the font embedding model and the annotated historical data; The encoding feature data is obtained according to the first encoding feature data and the second encoding feature data.

6. The method according to any one of claims 1 to 5, characterized in that: After obtaining the identification materials input by the person to be identified, the method further includes: The identification materials are stored in blockchain.

7. The method according to any one of claims 1 to 5, characterized in that: The step of generating an appraisal opinion based on the appraisal standard file, the forensic clinical examination standard file, the inspection result and the appraisal level includes: The identification standard file, the forensic clinical examination standard file, the inspection result and the identification level are input into an identification opinion template to generate an identification opinion.

8. A device for generating a judicial appraisal opinion based on a natural language model, characterized in that: include: An appraisal item acquisition module is used to acquire an appraisal item request input by the person to be appraised, and determine an appraisal standard file according to the appraisal item request, wherein the appraisal item request includes the appraisal items; An identification summary acquisition module is used to acquire the identification materials input by the identified person and obtain the identification summary of the identification materials through a text recognition algorithm; An inspection opinion acquisition module, used to obtain a forensic clinical inspection opinion corresponding to the identification summary according to a first preset natural language model and the identification summary, wherein the first preset natural language model is obtained by training based on a pre-trained language BERT model, a font embedding model, and a global pointer model; An identification level acquisition module, used to obtain the inspection results of the forensic clinical examination, and obtain the identification level of the person to be identified according to a second preset natural language model, the inspection results and the identification standard file, wherein the second preset natural language model is obtained by training based on a pre-trained language BERT model, a font embedding model and a global pointer model; A generation module is used to generate an identification opinion based on the identification standard file, the forensic clinical examination standard file, the inspection result and the identification level.

9. An electronic device, characterized in that: include: Processor and memory; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory, so that the processor performs the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when a processor executes the computer-executable instructions, the method according to any one of claims 1 to 7 is implemented.

Images