A machine learning-based judicial expertise information extraction method

Abstract

The application discloses a kind of based on machine learning's judicial expertise information extraction method, comprising the following steps: reading identification material, extracting fragment, number and original text index, form node table;According to position, number consistent and process support relationship build text evidence graph;Build occupation conservation account book, record role, occupation chain, shared mark and state;Evidence graph and account book are input into improved DiffPool, and role cluster is pooled according to occupation constraint;With test material consistency mask, exclude number conflict cluster;With the opinion role cluster as anchoring column, generate evidence chain cluster and update state;Back mapping original text, delete chain check, output structured judicial expertise information.The application relates to the technical field of judicial expertise information processing, improves the accuracy of judicial expertise information extraction, evidence chain integrity and review efficiency.

Description

Technical Field

[0001] This invention relates to the field of forensic identification information processing technology, and in particular to a method for extracting forensic identification information based on machine learning. Background Technology

[0002] Forensic identification data includes commission information, sample information, testing methods, identification process, expert opinions, and signed / stamped content. These documents come from various sources, have significantly different formats, and exhibit overlap, repetition, and omissions in their expression. Current methods for extracting forensic identification information primarily rely on manual input, template matching, keyword retrieval, or ordinary sequence labeling models, which perform field recognition on text fragments to output structured results.

[0003] The aforementioned methods focus on single-field extraction and lack an evidence chain organization mechanism that considers the identification matter, sample, testing method, identification process, and identification opinion. When the same text fragment is reused in multiple identification opinions, existing methods struggle to record exclusive, shared, and unused states, easily leading to the same evidence being repeatedly used by conflicting evidence chains. Ordinary graph pooling models aggregate text fragments based on node features but do not introduce evidence occupancy constraints in the role allocation matrix or a sample consistency pooling mask in the evidence chain attribution stage; therefore, conflicting role clusters can still enter the same identification opinion evidence chain. Existing output results lack original text index back-mapping and chain deletion verification, resulting in insufficient capabilities for source location, evidence closure judgment, and review processing of structured results. Therefore, providing a machine learning-based forensic information extraction method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a machine learning-based method for extracting forensic identification information. This invention utilizes an improved DiffPool structure, an evidence occupancy conservation ledger, and a sample consistency pooling mask to extract, assign, and verify text fragments, numbered information, and evidence chains of expert opinions in forensic identification materials. It has the advantages of clear evidence chain organization, accurate control of duplicate occupancy, traceability of original text, and high review efficiency.

[0005] A method for extracting forensic identification information based on machine learning according to an embodiment of the present invention includes the following steps: Read forensic identification data, extract text fragments, numbering information, and original text index, and generate a forensic identification text fragment node table; Using text fragments as graph nodes, graph edges are established according to text positional relationships, numbering consistency relationships, and identification process support relationships to generate a judicial identification text evidence graph. Establish an evidence occupancy conservation ledger using text fragments, and register the fields of role, occupancy chain number, sharing marker, and occupancy status; The judicial appraisal text evidence diagram and evidence occupancy conservation ledger are input into the improved DiffPool structure. An occupancy constraint term is introduced in the generation of the role allocation matrix. The text fragments are constrained into prohibited selection, shareable selection and optional selection according to the state of exclusive, shared and unoccupied. The constrained role allocation matrix is ​​pooled into appraisal role clusters. In the evidence chain attribution stage of the improved DiffPool structure, a sample consistency pooling mask is introduced to retain the identification role clusters with consistent numbers as optional objects of the evidence chain attribution matrix, and to exclude the identification role clusters with conflicting numbers from the same identification opinion evidence chain cluster. Using the expert opinion role cluster as the anchor column of the evidence chain attribution matrix, the attribution position of the expert opinion role cluster is determined within the limit of the sample consistency pooling mask. The pooling generates the expert opinion evidence chain cluster, and the occupancy status is updated according to the evidence occupancy conservation ledger. The evidence chain clusters of expert opinions are back-mapped to the original text index. The chain deletion verification is performed on the evidence chain clusters of expert opinions. Based on the original text mapping results, chain deletion verification results, occupancy status and number consistency results, structured judicial appraisal information is output.

[0006] Optionally, the operation of generating the forensic identification text fragment node table includes: The forensic identification data was segmented according to page order to obtain text fragments; Extract the numbering information from the text fragments, and configure the text fragment numbering and original text index; Write the text fragment number, text fragment, number information, and original text index into the same row of records to generate a forensic identification text fragment node table.

[0007] Optionally, the operation of generating the forensic text evidence diagram includes: Read the text fragment number, number information, and original text index from the forensic text fragment node table; Create text position edges based on adjacent text segments in the original index; Create edges with the same numbering information for text segments; Establish a supporting framework for the identification process in the order of identification matters, sample materials, testing methods, identification process, and identification opinions; Write the text location edge, the numbering consistency edge, and the identification process support edge into the graph edge record to generate a judicial identification text evidence graph.

[0008] Optionally, the operation of establishing an evidence-based conservation ledger from text fragments includes: Read the text fragment number, text fragment, and number information from the forensic identification text fragment node table; Configure field roles based on the identification matter, sample, testing method, identification process, and identification opinion to which the text fragment belongs; Configure the occupancy chain number for the text fragment according to the numbering information; Text fragments that can be shared by several expert opinions will be registered as shared tags, and text fragments that correspond to only a single expert opinion will be registered as exclusive tags; Text fragments not included in the evidence chain cluster of the expert opinion are registered as unoccupied, and an evidence occupancy conservation ledger containing text fragment number, field role, occupancy chain number, sharing tag and occupancy status is generated.

[0009] Optionally, the operation of introducing an occupancy constraint term in the role allocation matrix generation includes: Read the occupancy status and sharing marker in the evidence occupancy conservation ledger according to the text fragment number; Mark the corresponding matrix position of the text fragment whose occupancy status is exclusive in the role allocation matrix as unselectable; Mark the corresponding matrix position of the text fragment with the shared tag in the role assignment matrix as shareable; Mark the corresponding matrix position of the text fragment with an unoccupied status in the role allocation matrix as optional; Based on the marked results of prohibited, shareable, and optional roles, occupancy constraints are generated and written into the corresponding matrix positions of the role allocation matrix to obtain the constrained role allocation matrix.

[0010] Optionally, the operation of pooling the constrained role allocation matrix into an identification role cluster includes: Read the role assignment matrix after constraints and mark the matrix positions that are shareable and optional. Merge the text fragment numbers according to the field role to form the role clusters of identification matters, evidence samples, testing methods, identification process, and identification opinions. Exclude text fragments marked as forbidden from the identification role cluster; Create corresponding records within the same identification role cluster by assigning text fragment numbers, field roles, number information, and original text indexes to the same identification role cluster. Write the identification role cluster number and the corresponding record within the cluster into the evidence occupancy conservation ledger.

[0011] Optionally, the operation of introducing a sample consistency pooling mask in the evidence chain attribution stage of the improved DiffPool structure includes: Read the text fragment number and number information from the corresponding record within the identified role cluster; Extract the case number, evidence number, and sample number from the numbering information; Using the expert opinion role cluster as the target chain index, the case number, sample number and specimen number of the expert role cluster to be assigned are compared with the expert opinion role cluster item by item; When all three numbers are consistent, the cluster of roles to be identified is marked as admitted in the sample consistency pooling mask. If any number is inconsistent, the role cluster to be identified will be marked as prohibited in the sample consistency pooling mask; Clusters of expert witness roles marked as admitted are retained as optional objects in the evidence chain attribution matrix, while clusters of expert witness roles marked as prohibited are excluded from the same evidence chain cluster of expert opinions.

[0012] Optionally, the operation of using the expert opinion role cluster as the anchor column of the evidence chain attribution matrix includes: Read the identification role clusters marked as admission from the evidence consistency pooling mask, configure the admission identification role cluster numbers as row indices of the evidence chain attribution matrix, and configure the identification opinion role cluster numbers as column indices of the evidence chain attribution matrix. The identification process support edge and occupied chain number between the identification role cluster and the identification opinion role cluster are compared line by line. When there is a supporting edge for the identification process and the occupied chain number is consistent, the attribution value of the corresponding row and column position in the evidence chain attribution matrix is ​​set to 1; If there is no supporting edge for the identification process or the occupied chain number is inconsistent, the attribution value of the corresponding row and column position in the evidence chain attribution matrix is ​​set to 0; The expert opinion role cluster number corresponding to the column index with an attribution value of 1 is used as the main index of the evidence chain cluster. The expert opinion evidence chain cluster is generated by summarizing the expert opinion role cluster numbers corresponding to the row index with an attribution value of 1.

[0013] Optionally, the operation of updating the occupancy status according to the evidence-based occupancy conservation ledger includes: Read the expert opinion evidence chain cluster number and the expert role cluster number from the expert opinion evidence chain cluster; In the evidence occupancy conservation ledger, search for the corresponding text fragment number, sharing tag, occupancy chain number, and occupancy status by the identification role cluster number; When the shared token is set to the exclusive token and the occupancy status is unoccupied, the occupancy status is changed to the exclusive status. When the shared flag is set to shared and the occupied status is unoccupied, the occupied status is changed to shared occupied status. When the shared marker is a shared marker and the occupancy status is a shared occupancy status, the evidence chain cluster number of the expert opinion will be appended to the occupancy chain number; When the shared token is an exclusive token and the occupancy status is exclusive, the occupancy status remains unchanged.

[0014] Optionally, the operation of back-mapping the evidence chain cluster of the expert opinion to the original text index and performing chain deletion verification includes: Read the evidence chain cluster number, the expert role cluster number, and the corresponding text fragment number of the expert opinion; Search the original text index by text fragment number in the judicial appraisal text fragment node table to generate the original text mapping result; The role cluster is deleted from the evidence chain cluster of the expert opinion, resulting in the evidence chain to be verified. Check the integrity of the supporting edges of the identification process between the identification matter role cluster and the identification opinion role cluster in the evidence chain to be verified. If the edges are complete, the evidence chain is marked as closed; if they are missing, the evidence chain is marked as unclosed. Generate deletion verification results based on the closed state of the evidence chain, and output structured forensic identification information based on the original text mapping results, deletion verification results, occupancy status and number consistency results.

[0015] The beneficial effects of this invention are: (extending from the above) (1) This invention divides judicial appraisal data into text fragments, and generates a judicial appraisal text fragment node table by combining numbering information and original text index, and then establishes a judicial appraisal text evidence diagram, so that the appraisal matters, evidence samples, testing methods, appraisal process and appraisal opinions form a calculable evidence chain structure, thereby improving the organizational integrity of judicial appraisal information extraction results.

[0016] (2) The present invention introduces an evidence occupancy conservation ledger and occupancy constraint terms into the improved DiffPool structure, and performs differentiated pooling constraints on text fragments in the states of exclusive, shared and unoccupied status, thereby reducing the problem of the same text fragment being repeatedly occupied by multiple conflict identification opinions and improving the accuracy of identification role cluster generation.

[0017] (3) In the evidence chain attribution stage, the present invention introduces a sample consistency pooling mask and uses the expert opinion role cluster as the anchor column of the evidence chain attribution matrix to filter out expert role clusters with conflicting numbers and generate expert opinion evidence chain clusters; then, combined with the original text back mapping and chain deletion verification, the structured judicial appraisal information is output to improve the traceability of the results and the efficiency of review. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a machine learning-based forensic information extraction method proposed in this invention; Figure 2 This invention presents a schematic diagram of evidence diagrams and ledger generation for a machine learning-based forensic identification information extraction method. Figure 3 This is a schematic diagram of the improved DiffPool processing method for extracting forensic identification information based on machine learning proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A machine learning-based method for extracting forensic identification information includes the following steps: Read forensic identification data, extract text fragments, numbering information, and original text index, and generate a forensic identification text fragment node table; Forensic identification materials include the identification entrustment form, acceptance record, sample transfer record, testing record, the main text of the identification report, and the signature page. When reading forensic identification materials, the page content is read page by page in the document's page order. The title area, main text area, table area, and signature area are identified. Continuous text in the title and main text areas is segmented by paragraph boundaries, text in the table area is segmented by cell boundaries, and the institution name, personnel name, and date fields in the signature area are segmented separately to obtain text fragments. Numbering information is extracted from the text fragments, including case number, entrustment number, sample number, identification matter number, and page sequence number. The original text index consists of the document page number, paragraph sequence number, line sequence number, character start and end positions, table row and column positions, and signature area positions. The forensic identification text fragment node table establishes row records according to the text fragment number. Each row record records at least the text fragment number, text fragment, numbering information, original text index, and fragment source area.

[0021] The Forensic Expertise Text Fragment Node Table serves as the foundation for text fragments to enter subsequent graph processing, ledger processing, and back-mapping processing. The table includes fields for text fragment number, document page number, fragment source area, text fragment, case number, commission number, evidence number, sample number, expert opinion number, page number, and original text index. The text fragment number field is the primary key. The document page number field records the page where the text fragment is located. The fragment source area field records the title area, body area, table area, or signature area. The text fragment field records the original text content. The case number, commission number, evidence number, sample number, and expert opinion number fields record the corresponding number values ​​extracted from the original text content. The original text index field records the position of the text fragment within the forensic expertise document.

[0022] Using text fragments as graph nodes, graph edges are established according to text positional relationships, numbering consistency relationships, and identification process support relationships to generate a judicial identification text evidence graph. When creating a forensic evidence graph, each text fragment record in the forensic text fragment node table is converted into a graph node, and the text fragment number is used as the graph node number. Text position relationships are determined by the original text index; adjacent text fragments within the same page have text position edges, as do text fragments corresponding to consecutive titles or tables across pages. Number consistency relationships are determined by numbering information; text fragments with the same case number, evidence number, sample number, or forensic item number have number consistency edges. Forensic process support relationships, edges are created in the following order: forensic item points to evidence sample, evidence sample points to testing method, testing method points to forensic process, and forensic process points to forensic opinion. Graph edge records register the starting text fragment number, ending text fragment number, edge type, and edge source fields. Edge types include text position edges, number consistency edges, and forensic process support edges.

[0023] The forensic evidence diagram consists of node records and edge records. Node records are derived from the forensic text fragment node table, while edge records include fields for edge number, starting text fragment number, ending text fragment number, edge type, number consistency type, and process support type. Text position edges preserve the original writing order of the forensic materials; number consistency edges preserve the attribution relationship between the same case, the same evidence, the same sample, and the same forensic matter; and forensic process support edges represent the evidence support path between forensic matters, evidence samples, testing methods, forensic processes, and forensic opinions.

[0024] Establish an evidence occupancy conservation ledger using text fragments, and register the fields of role, occupancy chain number, sharing marker, and occupancy status; The evidence occupancy conservation ledger uses text fragment numbers as ledger row indexes. Field roles are determined by the text fragment content and its position on the graph edge. Field roles include identification matter role, evidence sample role, testing method role, identification process role, expert opinion role, and signature role. The occupancy chain number is a temporary number generated by combining the case number, evidence number, and sample number. After the expert opinion evidence chain cluster is generated, the expert opinion evidence chain cluster number is written into the occupancy chain number to identify the candidate expert opinion evidence chain to which the text fragment belongs. The sharing marker is used to distinguish whether a text fragment can be used by multiple expert opinion evidence chain clusters. Testing method descriptions, general testing standards, and general instrument descriptions are registered as sharing markers, while evidence sample descriptions, expert conclusion content, and signature content are registered as exclusive markers. Occupancy status includes unoccupied status, exclusive status, and shared occupancy status. When generating the evidence occupancy conservation ledger, text fragments that have not yet entered the expert opinion evidence chain cluster are uniformly registered as unoccupied status.

[0025] The evidence occupancy conservation ledger includes fields for text fragment number, field role, occupancy chain number, shared marker, occupancy status, expert opinion role cluster number, and expert opinion evidence chain cluster number. The text fragment number field connects the forensic text fragment node table and the forensic text evidence graph; the field role field records the role of the text fragment in the forensic process; the occupancy chain number field records which expert opinion evidence chain cluster occupies the text fragment; the shared marker field records whether it is a shared or exclusive marker; the occupancy status field records whether it is unoccupied, exclusively occupied, or shared; the expert opinion role cluster number field records the expert opinion role cluster to which the text fragment belongs after pooling; and the expert opinion evidence chain cluster number field records the expert opinion evidence chain cluster into which the text fragment ultimately enters.

[0026] The judicial appraisal text evidence diagram and evidence occupancy conservation ledger are input into the improved DiffPool structure. An occupancy constraint term is introduced in the generation of the role allocation matrix. The text fragments are constrained into prohibited selection, shareable selection and optional selection according to the state of exclusive, shared and unoccupied. The constrained role allocation matrix is ​​pooled into appraisal role clusters. The improved DiffPool structure receives graph node features, graph edge records, and ledger line records from the forensic evidence text graph and the evidence occupancy conservation ledger. Graph node features are encoded from text fragment content, field roles, numbering information, and original text index. A role allocation matrix records the candidate relationships of text fragments assigned to different forensic role clusters. Before generating the role allocation matrix, the occupancy status and sharing flag are queried in the evidence occupancy conservation ledger according to the text fragment number, generating occupancy constraint items. For text fragments in the "exclusively occupied" state, all allocable positions in the corresponding role allocation matrix are constrained to be prohibited; for text fragments with a shared flag, positions in the corresponding role allocation matrix associated with the detection method role cluster and the forensic process role cluster are constrained to be shareable; for text fragments in the "unoccupied" state, positions in the corresponding role allocation matrix matching the field role are constrained to be optional. After the role allocation matrix completes the constraints, prohibited positions do not participate in pooling, shareable positions are allowed to enter multiple forensic role clusters, and optional positions enter one forensic role cluster according to the field role. After pooling, the system forms role clusters for identification matters, samples, testing methods, identification processes, and identification opinions. Text fragments related to signatures and seals are retained in the original text index and structured output records.

[0027] The standard DiffPool structure generates node-to-cluster allocation relationships based on graph node features and edge structure, and performs pooling aggregation on the graph nodes. The improved DiffPool structure used in this invention adds an evidence occupancy conservation ledger input to the role allocation matrix generation stage of the standard DiffPool structure. This input converts occupancy status and sharing tags into occupancy constraints, which are then written into the corresponding positions in the role allocation matrix. The improved DiffPool structure also adds a sample consistency pooling mask to the evidence chain attribution stage, and uses the expert opinion role cluster as the anchor column of the evidence chain attribution matrix. This ensures that the text fragment pooling results are jointly constrained by the evidence occupancy rule, the sample number consistency rule, and the expert opinion attribution rule.

[0028] In the evidence chain attribution stage of the improved DiffPool structure, a sample consistency pooling mask is introduced to retain the identification role clusters with consistent numbers as optional objects of the evidence chain attribution matrix, and to exclude the identification role clusters with conflicting numbers from the same identification opinion evidence chain cluster. The evidence chain attribution phase is executed after the expert role clusters are generated. The evidence consistency pooling mask uses the cluster number to be assigned as the row and the expert opinion cluster number as the column. The intersection of the row and column is used to mark whether the cluster to be assigned can enter the corresponding expert opinion evidence chain cluster. When generating the evidence consistency pooling mask, the case number, evidence number, and sample number are extracted from the corresponding record within the expert role cluster, and the corresponding case number, evidence number, and sample number are extracted from the expert opinion cluster. When all three numbers of the cluster to be assigned match those of the expert opinion cluster, the corresponding position in the evidence consistency pooling mask is marked as "admitted"; if any number is inconsistent, the corresponding position in the evidence consistency pooling mask is marked as "prohibited". The evidence chain attribution matrix only accepts expert role clusters corresponding to admitted positions; expert role clusters corresponding to prohibited positions do not enter the same expert opinion evidence chain cluster.

[0029] The evidence consistency pooling mask is a two-dimensional marker table. Row fields register the cluster number of the role to be assigned to, column fields register the cluster number of the expert opinion, and the intersection positions register either an admission marker or a prohibition marker. An admission marker indicates that the case number, evidence number, and sample number of the role cluster to be assigned to are all identical to those of the expert opinion. A prohibition marker indicates that there is an inconsistency among these three numbering elements. When reading the evidence consistency pooling mask in the evidence chain attribution matrix, only the row and column positions corresponding to the admission markers are read, and the row and column positions corresponding to the prohibition markers are removed from the candidate attribution range.

[0030] Using the expert opinion role cluster as the anchor column of the evidence chain attribution matrix, the attribution position of the expert opinion role cluster is determined within the limit of the sample consistency pooling mask. The pooling generates the expert opinion evidence chain cluster, and the occupancy status is updated according to the evidence occupancy conservation ledger. The evidence chain attribution matrix uses the admission identification role cluster number as the row index and the identification opinion role cluster number as the column index. Each column corresponds to an identification opinion role cluster, and each row corresponds to an identification role cluster to be assigned. When determining the attribution position, the identification role cluster number to be assigned is read row by row, the identification process support edge between the identification role cluster to be assigned and the identification opinion role cluster is found, and the occupied chain number of the identification role cluster to be assigned and the identification opinion role cluster are compared. When an identification process support edge exists and the occupied chain number is consistent, the attribution value of the corresponding row and column position in the evidence chain attribution matrix is ​​set to 1; when no identification process support edge exists or the occupied chain number is inconsistent, the attribution value of the corresponding row and column position in the evidence chain attribution matrix is ​​set to 0. The identification opinion role cluster number corresponding to the column index with an attribution value of 1 is used as the main index of the evidence chain cluster. The identification matter role cluster number, sample role cluster number, testing method role cluster number, and identification process role cluster number corresponding to the row index with an attribution value of 1 are summarized to generate the identification opinion evidence chain cluster. After generating the evidence chain cluster of expert opinions, the evidence occupancy conservation ledger is checked back according to the cluster number of the expert role, and the text fragments that have entered the evidence chain cluster of expert opinions are updated to be in an exclusive or shared occupancy state.

[0031] The anchor columns of the evidence chain attribution matrix consist of the expert opinion role cluster numbers. Each expert opinion role cluster number corresponds to a column, and each column is used to form an expert opinion evidence chain cluster. When the attribution value in the evidence chain attribution matrix is ​​1, the expert role cluster of the corresponding row index is assigned to the expert opinion evidence chain cluster of the corresponding column index; when the attribution value is 0, the expert role cluster of the corresponding row index is not included in the expert opinion evidence chain cluster of the corresponding column index. The expert opinion evidence chain cluster record table sets up fields for expert opinion evidence chain cluster number, expert opinion matter role cluster number, evidence sample role cluster number, testing method role cluster number, expert opinion role cluster number, evidence chain cluster main index, and attribution value source.

[0032] The evidence chain clusters of expert opinions are back-mapped to the original text index. The chain deletion verification is performed on the evidence chain clusters of expert opinions. Based on the original text mapping results, chain deletion verification results, occupancy status and number consistency results, structured judicial appraisal information is output.

[0033] During backmapping, the expert opinion evidence chain cluster number and corresponding text fragment number are read. The original text index is then searched in the forensic text fragment node table by text fragment number. A correspondence is established between the original text index and the expert opinion evidence chain cluster number to obtain the original text mapping result. During chain deletion verification, the expert opinion evidence chain cluster is used as the verification object. The sample role cluster, testing method role cluster, and expert process role cluster are deleted respectively, forming evidence chains to be verified. For each evidence chain to be verified, the continuity of the expert process support edge between the expert matter role cluster and the expert opinion role cluster is checked. If continuous, the evidence chain is marked as closed; if the expert process support edge is missing, the evidence chain is marked as unclosed. Structured forensic information includes the expert opinion evidence chain cluster number, expert matter field, sample field, testing method field, expert process field, expert opinion field, original text mapping result, chain deletion verification result, occupancy status, and number consistency result.

[0034] Structured forensic information is used to create output records based on the evidence chain cluster number of the forensic opinion. The output records include fields for: Forensic Opinion Evidence Chain Cluster Number, Case Number, Evidence Number, Sample Number, Forensic Matter, Evidence Sample, Testing Method, Forensic Process, Forensic Opinion, Original Text Mapping Result, Chain Deletion Verification Result, Occupancy Status, Number Consistency Result, and Review Prompt. The Original Text Mapping Result field records the page number, region type, paragraph number, and character start and end positions corresponding to each output field. The Chain Deletion Verification Result field records the closure status of the evidence chain. The Occupancy Status field records the final status of the text fragment in the evidence occupancy conservation ledger. The Number Consistency Result field records the admission and prohibition markers in the evidence consistency pooling mask. The Review Prompt field registers unclosed evidence chains, conflicting number evidence chains, and exclusively / repeatedly occupied text fragments.

[0035] In this embodiment, the operation of generating the forensic identification text fragment node table includes: The forensic identification data was segmented according to page order to obtain text fragments; When segmenting forensic identification data by page order, the data is converted into readable page text and processed in ascending order of page number. Each page is first divided into title area, body text area, table area, and signature / stamp area, then segmented according to area type. The title area is segmented by title line, the body text area by paragraph break and punctuation ending position, the table area by row and column units, and the signature / stamp area by institution name, expert name, issuance date, and stamp position. After segmentation, each continuous section is treated as a single text segment.

[0036] After the text is segmented, it is assigned a segment number within the page according to the page number and the order within the page. In the main text area, consecutive text segments spanning multiple lines are merged into a single text segment before the paragraph ends; in the table area, text spanning multiple lines within the same cell is merged into a single text segment; in the signature area, the organization name, the appraiser's name, the date of issuance, and the location of the seal are each formed into an independent text segment.

[0037] Extract the numbering information from the text fragments, and configure the text fragment numbering and original text index; Numbering information is obtained through field name matching, fixed format matching, and context position matching. Field name matching is used to identify case numbers, commission numbers, evidence numbers, sample numbers, and identification matter numbers. Fixed format matching is used to identify number strings containing numbers, letters, hyphens, and parentheses. Context position matching is used to bind the number string to the identification matter, evidence sample, or expert opinion in adjacent text fragments. Text fragment numbers are generated in the order of document number, page number, and fragment sequence number within the page. The original text index records the page number, area type, paragraph sequence number, table row and column position, character start position, and character end position of the text fragment.

[0038] Write the text fragment number, text fragment, number information, and original text index into the same row of records to generate a forensic identification text fragment node table.

[0039] When writing to the same row, the text fragment number is used as the row primary key, the original text content of the text fragment is registered in the text fragment field, the case number, commission number, evidence number, sample number, and appraisal matter number are registered in the number information field, and the page number, area type, paragraph number, table row and column position, character start position, and character end position are registered in the original text index field. After the forensic identification text fragment node table is completed, each text fragment corresponds to a text fragment number, and each text fragment number can be used to retrieve the original text position in the forensic identification materials.

[0040] The row records in the forensic identification text fragment node table can also register the candidate role field. The candidate role field is written with the corresponding candidate role when the text fragment contains identification matters, specimen name, testing method, identification process, identification opinion or signature information, and is used for subsequent field role configuration.

[0041] In this embodiment, the operation of generating a forensic identification text evidence diagram includes: Read the text fragment number, number information, and original text index from the forensic text fragment node table; During reading, the text fragment number is used as the graph node number. The corresponding numbering information and original text index are retrieved, and the page number, paragraph position, table position, and signature position of the text fragment are used as reference data for edge construction. Each graph node in the forensic evidence text diagram corresponds to a row record in the forensic text fragment node table.

[0042] Create text position edges based on adjacent text segments in the original index; When creating text position edges, within the same page, text segments are sorted by paragraph number, line number, and character start and end positions, and adjacent segments are connected as text position edges. For consecutive paragraphs spanning multiple pages, text position edges are created based on the page sequence continuity between the text segment at the end of the previous page and the text segment at the beginning of the next page. In table areas, text position edges are created based on adjacent cells in the same row and adjacent cells in the same column.

[0043] Create edges with the same numbering information for text segments; When establishing a matching edge, text segments with the same case number are connected to form a matching edge; text segments with the same evidence number are connected to form a matching edge; text segments with the same sample number are connected to form a matching edge; and text segments with the same identification item number are connected to form a matching edge. The start and end points of the matching edge are determined by the text segment numbers.

[0044] Establish a supporting framework for the identification process in the order of identification matters, sample materials, testing methods, identification process, and identification opinions; When establishing support edges for the identification process, first identify the field roles, and then establish directed edges according to the directions: identification matter role points to the sample role, sample role points to the testing method role, testing method role points to the identification process role, and identification process role points to the identification opinion role. Each support edge for the identification process registers the starting field role, the ending field role, the starting text fragment number, and the ending text fragment number.

[0045] Write the text location edge, the numbering consistency edge, and the identification process support edge into the graph edge record to generate a judicial identification text evidence graph.

[0046] The graph edge record includes fields for starting text fragment number, ending text fragment number, graph edge type, and graph edge source. Text position edges, numbered matching edges, and edges supporting the identification process are registered in the graph edge type field. The forensic identification text evidence graph consists of a set of text fragment nodes and a set of graph edge records, serving as the input graph for the improved DiffPool structure.

[0047] After the forensic text evidence diagram is completed, duplicate edges between text fragments with the same starting point number and the same ending point number are merged. During the merging process, all edge types in the edge type field are retained, and multiple edge source fields are merged into a single edge source record. The merged forensic text evidence diagram retains the positional order, numbering consistency, and supporting relationships of the forensic process among the text fragments.

[0048] In this embodiment, the operation of establishing an evidence occupancy conservation ledger using text fragments includes: Read the text fragment number, text fragment, and number information from the forensic identification text fragment node table; During reading, the text fragment number is used as the ledger row index of the evidence occupancy conservation ledger, and the text fragment, number information, and original text index are used as the basic fields of the ledger row. The text fragment number is used to maintain a correspondence with the graph node number in the forensic text evidence diagram.

[0049] Configure field roles based on the identification matter, sample, testing method, identification process, and identification opinion to which the text fragment belongs; When configuring field roles, the system identifies the field names, field values, and adjacent edges in the forensic evidence text diagram within the text fragment. When a text fragment contains an identification request, identification item, or identification scope, it is configured as an identification matter role; when it contains the name of the evidence, evidence number, sample name, or sample number, it is configured as a evidence sample role; when it contains testing standards, testing instruments, testing method names, or experimental conditions, it is configured as a testing method role; when it contains testing steps, observation results, comparison processes, or analysis processes, it is configured as an identification process role; and when it contains opinions, conclusions, identification opinions, or judgment results, it is configured as an identification opinion role.

[0050] Configure the occupancy chain number for the text fragment according to the numbering information; The occupancy chain number is generated by combining the case number, evidence number, sample number, and candidate number for the expert opinion role cluster. If a candidate number for the expert opinion role cluster is missing, a temporary occupancy chain number is first generated based on the case number, evidence number, and sample number. After the expert opinion evidence chain cluster is generated, the temporary occupancy chain number is updated to the expert opinion evidence chain cluster number. The occupancy chain number is used to record which expert opinion evidence chain cluster the text fragment enters.

[0051] Text fragments that can be shared by several expert opinions will be registered as shared tags, and text fragments that correspond to only a single expert opinion will be registered as exclusive tags; When a shared marker is registered, the text fragments corresponding to the detection method role and the general identification process role are marked as shared markers. When an exclusive marker is registered, the text fragments corresponding to the sample role, the expert opinion role, and the signature role are marked as exclusive markers. After a text fragment is registered with a shared marker, its corresponding position in the subsequent role allocation matrix can be constrained to be shareable; after a text fragment is registered with an exclusive marker, it will not enter other expert opinion evidence chain clusters with conflicting numbers after entering one expert opinion evidence chain cluster.

[0052] Text fragments not included in the evidence chain cluster of the expert opinion are registered as unoccupied, and an evidence occupancy conservation ledger containing text fragment number, field role, occupancy chain number, sharing tag and occupancy status is generated.

[0053] When generating the ledger, a ledger line record is created for each text fragment number. The ledger line record includes the text fragment number, field role, occupancy chain number, sharing flag, occupancy status, identification role cluster number, and identification opinion evidence chain cluster number. Initially, the identification role cluster number and identification opinion evidence chain cluster number are null values, and the occupancy status is recorded as unoccupied.

[0054] The occupancy status update rules are written to the ledger rule record of the evidence occupancy conservation ledger. The ledger rule record includes a shared marker field, an original occupancy status field, a target occupancy status field, and an occupancy chain number processing field. When the shared marker is an exclusive marker and the original occupancy status is unoccupied, the target occupancy status is exclusive, and the occupancy chain number processing field records the current expert opinion evidence chain cluster number. When the shared marker is a shared marker and the original occupancy status is unoccupied, the target occupancy status is shared, and the occupancy chain number processing field records the current expert opinion evidence chain cluster number. When the shared marker is a shared marker and the original occupancy status is shared, the target occupancy status remains shared, and the occupancy chain number processing field adds the current expert opinion evidence chain cluster number. When the shared marker is an exclusive marker and the original occupancy status is exclusive, the target occupancy status remains exclusive, and the occupancy chain number processing field does not add anything.

[0055] In this embodiment, the operation of introducing an occupancy constraint term in the generation of the role allocation matrix includes: Read the occupancy status and sharing marker in the evidence occupancy conservation ledger according to the text fragment number; Before generating the role assignment matrix using the improved DiffPool structure, the text fragment numbers are read one by one according to the graph node numbers in the forensic evidence graph. The occupancy status and sharing flag corresponding to the same text fragment number are then searched in the evidence occupancy conservation ledger. The reading results form the input record for the occupancy constraint item, which includes the text fragment number, occupancy status, sharing flag, and field role.

[0056] Mark the corresponding matrix position of the text fragment whose occupancy status is exclusive in the role allocation matrix as unselectable; In the role allocation matrix, rows correspond to text fragment numbers, and columns correspond to candidate numbers for role clusters. When a text fragment is in an exclusive state, the position in the row containing the text fragment number that corresponds to all candidate numbers for role clusters is marked as forbidden. Forbidden positions do not participate in role pooling and do not generate new role cluster affiliations.

[0057] Mark the corresponding matrix position of the text fragment with the shared tag in the role assignment matrix as shareable; When a text fragment is registered with a sharing tag, the position in the row containing the text fragment number that corresponds to the detection method role cluster and the identification process role cluster is marked as shareable. Shareable positions allow the same text fragment number to enter multiple evidence chain clusters of identification opinions, but must be limited by the evidence consistency pooling mask.

[0058] Mark the corresponding matrix position of the text fragment with an unoccupied status in the role allocation matrix as optional; When a text fragment is in an unoccupied state, the position corresponding to the candidate number of the identification role cluster that matches the field role in the row containing the text fragment number is marked as optional. The optional position enters an identification role cluster after role pooling and waits for the occupancy status to be updated in the evidence occupancy conservation ledger.

[0059] Based on the marked results of prohibited, shareable, and optional roles, occupancy constraints are generated and written into the corresponding matrix positions of the role allocation matrix to obtain the constrained role allocation matrix.

[0060] Occupancy constraints are represented by three constraint states: prohibited, shareable, and optional. After generating occupancy constraints, these constraints are overwritten in the role allocation matrix to cover the positions corresponding to the same text fragment number and the same identification role cluster candidate number. After overwriting, each matrix position in the role allocation matrix carries one of the constraint states: prohibited, shareable, or optional, resulting in a constrained role allocation matrix.

[0061] After the occupancy constraint is written into the role allocation matrix, the matrix position is read according to the constraint state. The matrix position corresponding to the forbidden state is skipped during pooling read; the matrix position corresponding to the shareable state is retained during pooling read and is filtered by the evidence consistency pooling mask during the evidence chain attribution stage; the matrix position corresponding to the optional state is retained during pooling read and merged into an identification role cluster within the same field role range.

[0062] In this embodiment, the operation of pooling the constrained role allocation matrix into an identification role cluster includes: Read the role assignment matrix after constraints and mark the matrix positions that are shareable and optional. Merge the text fragment numbers according to the field role to form the role clusters of identification matters, evidence samples, testing methods, identification process, and identification opinions. During pooling, the constrained role allocation matrix is ​​read row by row. When the matrix position marker is optional, the corresponding text fragment number is added to the identification role cluster that matches the field role. When the matrix position marker is shareable, the corresponding text fragment number is added to the detection method role cluster or the identification process role cluster. After merging according to field roles, the text fragment numbers corresponding to the identification matter role form the identification matter role cluster, the text fragment numbers corresponding to the sample role form the sample role cluster, the text fragment numbers corresponding to the detection method role form the detection method role cluster, the text fragment numbers corresponding to the identification process role form the identification process role cluster, and the text fragment numbers corresponding to the identification opinion role form the identification opinion role cluster.

[0063] Exclude text fragments marked as forbidden from the identification role cluster; When a forbidden selection marker is encountered, the original occupancy status of the text fragment number in the evidence occupancy conservation ledger is preserved, and the text fragment number is not added to any newly generated identification role cluster. The original text index, numbering information, and occupancy chain number of the forbidden text fragment remain unchanged.

[0064] Create corresponding records within the same identification role cluster by assigning text fragment numbers, field roles, number information, and original text indexes to the same identification role cluster. The corresponding record within each cluster is indexed by the identification role cluster number, recording the field role, number information, and original text index corresponding to each text fragment number within the cluster. The identification role cluster number is configured according to the field role type and generation order. The corresponding record within the cluster is used for subsequent evidence consistency pooling mask reading of the number information.

[0065] Write the identification role cluster number and the corresponding record within the cluster into the evidence occupancy conservation ledger.

[0066] When writing to the evidence occupancy conservation ledger, the ledger line is located according to the text fragment number. The identification role cluster number is registered in the same ledger line, and the corresponding record within the cluster is associated with the text fragment number. After registration, the evidence occupancy conservation ledger simultaneously records the text fragment's field role, occupancy chain number, sharing flag, occupancy status, and identification role cluster number.

[0067] The identification role cluster record includes fields for Identification Role Cluster Number, Field Role, Text Fragment Number within Cluster, Case Number, Specimen Number, Sample Number, and Original Text Index Set. The Identification Item Role Cluster records the text fragment numbers related to the identification item; the Specimen / Sample Role Cluster records the text fragment numbers related to the specimen and sample; the Testing Method Role Cluster records the text fragment numbers related to the testing standards, instruments, and methods; the Identification Process Role Cluster records the text fragment numbers related to the testing steps, observation results, and analysis process; and the Identification Opinion Role Cluster records the text fragment numbers related to the opinion conclusion and judgment result.

[0068] In this embodiment, the operation of introducing a sample consistency pooling mask in the evidence chain attribution stage of the improved DiffPool structure includes: Read the text fragment number and number information from the corresponding record within the identified role cluster; At the start of the evidence chain attribution phase, the corresponding records within all expert witness clusters are read. For each expert witness cluster, the numbering information corresponding to the text fragment numbers within the cluster is summarized to form an expert witness cluster numbering information record. The numbering information record must include at least the case number, the evidence number, and the sample number.

[0069] Extract the case number, evidence number, and sample number from the numbering information; Case number is used to limit the scope of the same forensic identification case, evidence number is used to limit the same evidence object, and sample number is used to limit the specific sample under the same evidence object. During extraction, the number field name and number field value are matched. If a field name is missing, the number field name is filled in according to the field role in the adjacent text segment.

[0070] Using the expert opinion role cluster as the target chain index, the case number, sample number and specimen number of the expert role cluster to be assigned are compared with the expert opinion role cluster item by item; The target chain index is configured by the expert opinion role cluster number. Each expert opinion role cluster corresponds to a candidate expert opinion evidence chain cluster. During item-by-item comparison, the case number of the expert opinion role cluster to be assigned is compared with the case number of the expert opinion role cluster, the specimen number of the expert opinion role cluster to be assigned is compared with the specimen number of the expert opinion role cluster, and the sample number of the expert opinion role cluster to be assigned is compared with the sample number of the expert opinion role cluster.

[0071] When all three numbers are consistent, the cluster of roles to be identified is marked as admitted in the sample consistency pooling mask. When all three numbers are identical, it indicates that the cluster of roles to be assigned and the cluster of roles in the expert opinion belong to the same case, the same evidence, and the same sample. The evidence consistency pooling mask registers an admission marker at the corresponding positions of the cluster numbers to be assigned and the cluster numbers in the expert opinion. The cluster of roles corresponding to the admission marker enters the selectable objects of the evidence chain attribution matrix.

[0072] If any number is inconsistent, the role cluster to be identified will be marked as prohibited in the sample consistency pooling mask; When there are inconsistencies in case numbers, evidence numbers, or sample numbers, the evidence consistency pooling mask registers a prohibition mark at the corresponding position of the expert role cluster number to be assigned and the expert opinion role cluster number. The expert role cluster corresponding to the prohibition mark will not enter the same expert opinion evidence chain cluster.

[0073] Clusters of expert witness roles marked as admitted are retained as optional objects in the evidence chain attribution matrix, while clusters of expert witness roles marked as prohibited are excluded from the same evidence chain cluster of expert opinions.

[0074] After the evidence consistency pooling mask is completed, the mask positions are read one by one. The row and column positions corresponding to the admission markers are submitted to the evidence chain attribution matrix, while the row and column positions corresponding to the prohibition markers are not created as optional objects in the evidence chain attribution matrix.

[0075] The evidence consistency pooling mask also records the serial number consistency results. These results include case number consistency, evidence number consistency, and sample number consistency. If all three are consistent, the serial number consistency result is entered into the admission criteria. If any one is inconsistent, the serial number consistency result is entered into the prohibition criteria. The serial number consistency results are output synchronously during the output of structured forensic information.

[0076] In this embodiment, the operation of using the expert opinion role cluster as the anchor column of the evidence chain attribution matrix includes: Read the identification role clusters marked as admission from the evidence consistency pooling mask, configure the admission identification role cluster numbers as row indices of the evidence chain attribution matrix, and configure the identification opinion role cluster numbers as column indices of the evidence chain attribution matrix. Each row of the evidence chain attribution matrix corresponds to an admission-level expert role cluster number, and each column corresponds to an expert opinion role cluster number. The row index comes from the admission marker in the evidence consistency pooling mask, and the column index comes from the expert opinion role cluster number. After the matrix rows and columns are configured, each matrix position corresponds to a candidate attribution relationship of "expected expert role cluster - expert opinion role cluster".

[0077] The identification process support edge and occupied chain number between the identification role cluster and the identification opinion role cluster are compared line by line. During row-by-row comparison, the admission identification role cluster number is read according to the row index, and the identification opinion role cluster number is read according to the column index. The system then searches the judicial identification text evidence diagram to see if there is a supporting edge for the identification process between the two role clusters. If a direct supporting edge for the identification process exists, the existence of the edge is recorded. If no direct supporting edge for the identification process exists, the system continues to check if there are continuous supporting edges for the identification process between the identification matter role cluster, the sample role cluster, the testing method role cluster, the identification process role cluster, and the identification opinion role cluster. During the occupancy chain number comparison, the system reads the occupancy chain number corresponding to the admission identification role cluster and the occupancy chain number corresponding to the identification opinion role cluster from the evidence occupancy conservation ledger and performs a consistency comparison.

[0078] When there is a supporting edge for the identification process and the occupied chain number is consistent, the attribution value of the corresponding row and column position in the evidence chain attribution matrix is ​​set to 1; When an edge exists and the occupied chain number is consistent, the corresponding matrix position is confirmed as a valid attribution position. After the attribution value is set to 1, the identification role cluster number in the corresponding row index is bound to the identification opinion role cluster number in the corresponding column index.

[0079] If there is no supporting edge for the identification process or the occupied chain number is inconsistent, the attribution value of the corresponding row and column position in the evidence chain attribution matrix is ​​set to 0; When there is no supporting edge for the authentication process or the occupied chain number is inconsistent, the corresponding matrix position is determined to be an invalid attribution position. After the attribution value is set to 0, the authentication role cluster number in the corresponding row index is not bound to the authentication opinion role cluster number in the corresponding column index.

[0080] The expert opinion role cluster number corresponding to the column index with an attribution value of 1 is used as the main index of the evidence chain cluster. The expert opinion evidence chain cluster is generated by summarizing the expert opinion role cluster numbers corresponding to the row index with an attribution value of 1.

[0081] When generating an evidence chain cluster for expert opinions, the evidence chain attribution matrix is ​​read column by column. The column index corresponds to the expert opinion role cluster number. All row indices with an attribution value of 1 in the same column are read, and the expert opinion role cluster number, the evidence sample role cluster number, the testing method role cluster number, and the expert process role cluster number corresponding to the row index are combined with the expert opinion role cluster number corresponding to the column index to form an expert opinion evidence chain cluster.

[0082] After the evidence chain attribution matrix is ​​completed, an attribution process record is generated. This record includes the cluster number of the role to be attributed, the cluster number of the expert opinion, the evidence consistency pooling mask marker, the results of the expert process support edge check, the consistency result of the occupied chain number, and the attribution value. The attribution process record explains the basis for the expert opinion evidence chain cluster to be included in the expert opinion evidence chain cluster.

[0083] In this embodiment, the operation of updating the occupancy status according to the evidence occupancy conservation ledger includes: Read the expert opinion evidence chain cluster number and the expert role cluster number from the expert opinion evidence chain cluster; When updating the occupancy status, each expert opinion evidence chain cluster is read sequentially. Each expert opinion evidence chain cluster includes an expert opinion evidence chain cluster number, an expert opinion role cluster number, and several expert role cluster numbers. After reading, the corresponding record within the cluster is retrieved according to the expert role cluster number to obtain the corresponding text fragment number.

[0084] In the evidence occupancy conservation ledger, search for the corresponding text fragment number, sharing tag, occupancy chain number, and occupancy status by the identification role cluster number; During the search, the corresponding record within the cluster is first located using the cluster number of the identified role, and then the ledger line in the evidence occupancy conservation ledger is located using the text fragment number in the corresponding record within the cluster. Each ledger line reads the shared token, occupancy chain number, and occupancy status as the basis for updating the occupancy status.

[0085] When the shared token is set to the exclusive token and the occupancy status is unoccupied, the occupancy status is changed to the exclusive status. After being rewritten to an exclusive status, the expert opinion evidence chain cluster number is registered in the occupied chain number field of the corresponding ledger line. The exclusive status indicates that the same text fragment has been occupied by one expert opinion evidence chain cluster and will no longer participate in the role allocation of other expert opinion evidence chain clusters.

[0086] When the shared flag is set to shared and the occupied status is unoccupied, the occupied status is changed to shared occupied status. After being rewritten to a shared occupancy status, the current expert opinion evidence chain cluster number is registered in the occupancy chain number field. A shared occupancy status indicates that the same text fragment has entered one expert opinion evidence chain cluster and allows it to enter other expert opinion evidence chain clusters within the scope of the evidence consistency pooling mask admission.

[0087] When the shared marker is a shared marker and the occupancy status is a shared occupancy status, the evidence chain cluster number of the expert opinion will be appended to the occupancy chain number; During appending, the existing expert opinion evidence chain cluster number in the occupied chain number field is read, and the current expert opinion evidence chain cluster number is appended to the same occupied chain number field. After appending, the same text fragment number corresponds to multiple expert opinion evidence chain cluster numbers.

[0088] When the shared token is an exclusive token and the occupancy status is exclusive, the occupancy status remains unchanged.

[0089] While maintaining the existing occupancy status, no new expert opinion evidence chain cluster numbers are added, and the existing occupancy chain numbers are not changed. During the subsequent generation of the role allocation matrix, the matrix position corresponding to the same text fragment number is restricted from selection based on the already exclusively occupied state.

[0090] After the occupancy status update is completed, an occupancy status update record is generated. The occupancy status update record includes the text fragment number, sharing tag, occupancy status before update, occupancy status after update, original occupancy chain number, new occupancy chain number, and the evidence chain cluster number of the expert opinion that triggered the update. The occupancy status update record is used to show whether an exclusive text fragment has been repeatedly occupied and whether a shared text fragment has been reused within the access scope.

[0091] In this embodiment, the operation of back-mapping the evidence chain cluster of the expert opinion to the original text index and performing chain deletion verification includes: Read the evidence chain cluster number, the expert role cluster number, and the corresponding text fragment number of the expert opinion; At the start of the backmapping process, each evidence chain cluster of expert opinions is read sequentially. For each evidence chain cluster of expert opinions, the cluster number of the expert opinion evidence chain, the cluster number of the expert opinion role, the cluster number of the expert opinion matter role, the cluster number of the evidence sample role, the cluster number of the testing method role, the cluster number of the expert opinion process role, and the text fragment number corresponding to each role cluster are extracted.

[0092] Search the original text index by text fragment number in the judicial appraisal text fragment node table to generate the original text mapping result; When searching the original text index, the forensic identification text fragment node table is accessed according to the text fragment number, and the corresponding document page number, area type, paragraph number, table row and column position, character start position, and character end position are read. The read original text index is then used to establish a correspondence with the evidence chain cluster number of the expert opinion, forming the original text mapping result.

[0093] The role cluster is deleted from the evidence chain cluster of the expert opinion, resulting in the evidence chain to be verified. The deletion of role clusters is performed in three rounds. In the first round, the "Examined Sample" role cluster is deleted from the "Expert Opinion" evidence chain cluster, while the "Expert Item" role cluster, "Testing Method" role cluster, "Expert Process" role cluster, and "Expert Opinion" role cluster are retained, resulting in the first evidence chain to be verified. In the second round, the "Testing Method" role cluster is deleted from the "Expert Opinion" evidence chain cluster, while the "Examined Item" role cluster, "Examined Sample" role cluster, "Expert Process" role cluster, and "Expert Opinion" role cluster are retained, resulting in the second evidence chain to be verified. In the third round, the "Expert Opinion" evidence chain cluster is deleted from the "Expert Opinion" evidence chain cluster, while the "Examined Item" role cluster, "Examined Sample" role cluster, "Testing Method" role cluster, and "Expert Opinion" role cluster are retained, resulting in the third evidence chain to be verified.

[0094] Check the integrity of the supporting edges of the identification process between the identification matter role cluster and the identification opinion role cluster in the evidence chain to be verified. If the edges are complete, the evidence chain is marked as closed; if they are missing, the evidence chain is marked as unclosed. During the inspection, the supporting edges of the identification process between the role clusters in the forensic evidence text diagram are read. When the supporting edge paths between the identification matter role cluster, the sample role cluster, the testing method role cluster, the identification process role cluster, and the identification opinion role cluster are continuous, the evidence chain is recorded as closed. If the supporting edge path is broken after deleting any role cluster, the evidence chain is recorded as unclosed.

[0095] Generate deletion verification results based on the closed state of the evidence chain, and output structured forensic identification information based on the original text mapping results, deletion verification results, occupancy status and number consistency results.

[0096] The deletion chain verification result includes the expert opinion evidence chain cluster number, the type of the deleted role cluster, the evidence chain number to be verified, and the evidence chain closure status. The number consistency result is generated by the admission and prohibition flags in the evidence consistency pooling mask. When outputting structured forensic information, the expert opinion evidence chain cluster number, the expert opinion matter field, the evidence sample field, the testing method field, the expert opinion process field, the expert opinion field, the original text mapping result, the deletion chain verification result, the occupancy status, and the number consistency result are organized into a structured record.

[0097] After the chain deletion verification results are generated, review prompts are generated according to the closure status of the evidence chain. When the first evidence chain to be verified is not closed, the review prompt field records that the evidence sample is missing or the evidence sample is insufficient; when the second evidence chain to be verified is not closed, the review prompt field records that the testing method is missing or the testing method is insufficient; when the third evidence chain to be verified is not closed, the review prompt field records that the identification process is missing or the identification process is insufficient. After the structured forensic identification information is output, the reviewers can locate the corresponding forensic identification data according to the original text mapping results and check the corresponding evidence chain according to the chain deletion verification results.

[0098] Example 1: To verify the feasibility of this invention in practice, it was applied to a scenario involving the batch extraction of information from a certain type of forensic identification data. Input data included commission information, acceptance records, sample registration pages, testing process pages, result analysis pages, expert opinion pages, and signature pages; all data underwent anonymization. In this scenario, multiple expert opinions shared testing method descriptions, inspection standards, and instrument descriptions, and multiple sample numbers and identification numbers were recorded interchangeably in the same case. Manual processing easily leads to shared text being categorized as incorrect expert opinions and also easily results in exclusive text being repeatedly written into other expert opinion evidence chains; keyword-based methods struggle to record the occupancy status of text fragments, and ordinary graph pooling methods struggle to eliminate role clusters with conflicting numbers.

[0099] In this scenario, the system segments the forensic identification data into text fragments, extracts the case number, commission number, evidence number, sample number, and identification matter number, and writes them into the original text index to generate a forensic identification text fragment node table; then, it converts the text fragment numbers into graph nodes, establishes text position edges, number consistency edges, and identification process support edges, and generates a forensic identification text evidence graph.

[0100] The system establishes an evidence occupancy conservation ledger, registering fields such as role, occupancy chain number, sharing flag, and occupancy status. The improved DiffPool structure reads the forensic evidence text diagram and the evidence occupancy conservation ledger, writes occupancy constraints into the role allocation matrix, marks exclusively occupied text fragments as prohibited, shares as shareable, and unoccupied text fragments as optional, and pools them to form role clusters for identification matters, evidence samples, testing methods, identification processes, and identification opinions.

[0101] After the expert opinion role cluster is generated, the system compares the case number, evidence number, and sample number using a sample consistency pooling mask. Positions with matching numbers are registered as admitted, while positions with conflicting numbers are registered as prohibited. The evidence chain attribution matrix only reads the admitted positions and uses the expert opinion role cluster as the anchor column to generate the expert opinion evidence chain cluster. The system then checks the original text index and performs a chain deletion check. If the supporting edges are continuous, it is recorded as closed; if the supporting edges are broken, it is recorded as unclosed, and a review prompt is recorded.

[0102] To verify the effectiveness of the implementation, 160 samples of anonymized forensic identification data were selected for comparative testing. The data samples totaled 1842 pages, containing 16780 text fragments, 1126 expert opinions, 3248 sample numbers, 4091 specimen numbers, and 43620 manually labeled standard structured fields. Comparison methods included manual processing, keyword rule-based methods, ordinary graph pooling, and the method of this invention. The comparison results are shown in the table below: Table 1: Comparison of the Implementation Effects of Forensic Expertise Information Extraction

[0103] The data in the table shows that the method of this invention outperforms manual processing, keyword rule-based methods, and ordinary graph pooling methods in all core indicators. Regarding the accuracy of structured fields, the method of this invention achieves 97.4%, higher than 83.6% for manual processing, 88.9% for keyword rule-based methods, and 91.7% for ordinary graph pooling methods, indicating that this invention can more accurately extract fields from forensic identification materials. In terms of evidence chain closure rate, the method of this invention achieves 96.8%, 12.5 percentage points higher than 84.3% for ordinary graph pooling methods, indicating a higher degree of organizational integrity in the evidence chain clusters of the expert opinions.

[0104] Regarding original text tracing and number consistency, the method of this invention achieves a 98.2% original text index hit rate, a 97.9% accuracy rate for number consistency identification, and a 98.9% screening rate for sample number consistency, all significantly higher than the other three methods. The corresponding values ​​for ordinary graph pooling are 82.1%, 88.6%, and 77.6%, respectively, indicating that simply relying on graph node aggregation is insufficient to handle scenarios where sample and sample numbers overlap. This invention effectively reduces incorrect attribution by using a sample consistency pooling mask to screen out conflicting numbering clusters.

[0105] Regarding text fragment occupancy control, the method of this invention achieves a 96.1% accuracy rate for shared text fragment reuse and a 98.0% interception rate for exclusive text fragment conflicts, significantly higher than the 85.9% and 73.4% of the ordinary graph pooling method, respectively. In terms of the number of misattributed numbering conflicts, the method of this invention results in only 11 such instances, lower than the 146 instances of manual sorting, the 98 instances of keyword rule-based methods, and the 64 instances of ordinary graph pooling methods. Regarding processing efficiency, the method of this invention takes an average of 3.2 minutes to process a single data set and 2.7 minutes for verification, both lower than other methods. The above data demonstrate that this invention exhibits good implementation results in terms of extraction accuracy, evidence chain closure, numbering conflict screening, text fragment occupancy control, and verification efficiency.

[0106] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for extracting forensic identification information based on machine learning, characterized in that, Includes the following steps: Read forensic identification data, extract text fragments, numbering information, and original text index, and generate a forensic identification text fragment node table; Using text fragments as graph nodes, graph edges are established according to text positional relationships, numbering consistency relationships, and identification process support relationships to generate a judicial identification text evidence graph. Establish an evidence occupancy conservation ledger using text fragments, and register the fields of role, occupancy chain number, sharing marker, and occupancy status; The judicial appraisal text evidence diagram and evidence occupancy conservation ledger are input into the improved DiffPool structure. An occupancy constraint term is introduced in the generation of the role allocation matrix. The text fragments are constrained into prohibited selection, shareable selection and optional selection according to the state of exclusive, shared and unoccupied. The constrained role allocation matrix is ​​pooled into appraisal role clusters. In the evidence chain attribution stage of the improved DiffPool structure, a sample consistency pooling mask is introduced to retain the identification role clusters with consistent numbers as optional objects of the evidence chain attribution matrix, and to exclude the identification role clusters with conflicting numbers from the same identification opinion evidence chain cluster. Using the expert opinion role cluster as the anchor column of the evidence chain attribution matrix, the attribution position of the expert opinion role cluster is determined within the limit of the sample consistency pooling mask. The pooling generates the expert opinion evidence chain cluster, and the occupancy status is updated according to the evidence occupancy conservation ledger. The evidence chain clusters of expert opinions are back-mapped to the original text index. The chain deletion verification is performed on the evidence chain clusters of expert opinions. Based on the original text mapping results, chain deletion verification results, occupancy status and number consistency results, structured judicial appraisal information is output.

2. The method for extracting forensic identification information based on machine learning according to claim 1, characterized in that, The operation of generating the forensic identification text fragment node table includes: The forensic identification data was segmented according to page order to obtain text fragments; Extract the numbering information from the text fragments, and configure the text fragment numbering and original text index; Write the text fragment number, text fragment, number information, and original text index into the same row of records to generate a forensic identification text fragment node table.

3. The method for extracting forensic identification information based on machine learning according to claim 2, characterized in that, The operation of generating forensic evidence text diagrams includes: Read the text fragment number, number information, and original text index from the forensic text fragment node table; Create text position edges based on the text segments adjacent to the original index; Create edges with the same numbering information for text segments; Establish a supporting framework for the identification process in the order of identification matters, sample materials, testing methods, identification process, and identification opinions; Write the text location edge, the numbering consistency edge, and the identification process support edge into the graph edge record to generate a judicial identification text evidence graph.

4. The method for extracting forensic identification information based on machine learning according to claim 3, characterized in that, The operation of establishing an evidence-based conservation ledger using text fragments includes: Read the text fragment number, text fragment, and number information from the forensic identification text fragment node table; Configure field roles based on the identification matter, sample, testing method, identification process, and identification opinion to which the text fragment belongs; Configure the occupancy chain number for the text fragment according to the numbering information; Text fragments that can be shared by several expert opinions will be registered as shared tags, and text fragments that correspond to only a single expert opinion will be registered as exclusive tags; Text fragments not included in the evidence chain cluster of the expert opinion are registered as unoccupied, and an evidence occupancy conservation ledger containing text fragment number, field role, occupancy chain number, sharing tag and occupancy status is generated.

5. The method for extracting forensic identification information based on machine learning according to claim 4, characterized in that, The operation of introducing an occupancy constraint term in the generation of the role allocation matrix includes: Read the occupancy status and sharing marker in the evidence occupancy conservation ledger according to the text fragment number; Mark the corresponding matrix position of the text fragment whose occupancy status is exclusive in the role allocation matrix as unselectable; Mark the corresponding matrix position of the text fragment with the shared tag in the role assignment matrix as shareable; Mark the corresponding matrix position of the text fragment with an unoccupied status in the role allocation matrix as optional; Based on the marked results of prohibited, shareable, and optional roles, occupancy constraints are generated and written into the corresponding matrix positions of the role allocation matrix to obtain the constrained role allocation matrix.

6. The method for extracting forensic identification information based on machine learning according to claim 5, characterized in that, The operation of pooling the constrained role allocation matrix into an identification role cluster includes: Read the role assignment matrix after constraints and mark the matrix positions that are shareable and optional. Merge the text fragment numbers according to the field role to form the role clusters of identification matters, evidence samples, testing methods, identification process, and identification opinions. Exclude text fragments marked as forbidden from the identification role cluster; Create corresponding records within the same identification role cluster by assigning text fragment numbers, field roles, number information, and original text indexes to the same identification role cluster. Write the identification role cluster number and the corresponding record within the cluster into the evidence occupancy conservation ledger.

7. The method for extracting forensic identification information based on machine learning according to claim 6, characterized in that, The operation of introducing a sample consistency pooling mask in the evidence chain attribution stage of the improved DiffPool structure includes: Read the text fragment number and number information from the corresponding record within the identified role cluster; Extract the case number, evidence number, and sample number from the numbering information; Using the expert opinion role cluster as the target chain index, the case number, sample number and specimen number of the expert role cluster to be assigned are compared with the expert opinion role cluster item by item; When all three numbers are consistent, the cluster of roles to be identified is marked as admitted in the sample consistency pooling mask. If any number is inconsistent, the role cluster to be identified will be marked as prohibited in the sample consistency pooling mask; Clusters of expert witness roles marked as admitted are retained as optional objects in the evidence chain attribution matrix, while clusters of expert witness roles marked as prohibited are excluded from the same evidence chain cluster of expert opinions.

8. The method for extracting forensic identification information based on machine learning according to claim 7, characterized in that, The operation of using the expert opinion role cluster as the anchor column of the evidence chain attribution matrix includes: Read the identification role clusters marked as admission from the evidence consistency pooling mask, configure the admission identification role cluster numbers as row indices of the evidence chain attribution matrix, and configure the identification opinion role cluster numbers as column indices of the evidence chain attribution matrix. The identification process support edge and occupied chain number between the identification role cluster and the identification opinion role cluster are compared line by line. When there is a supporting edge for the identification process and the occupied chain number is consistent, the attribution value of the corresponding row and column position in the evidence chain attribution matrix is ​​set to 1; If there is no supporting edge for the identification process or the occupied chain number is inconsistent, the attribution value of the corresponding row and column position in the evidence chain attribution matrix is ​​set to 0; The expert opinion role cluster number corresponding to the column index with an attribution value of 1 is used as the main index of the evidence chain cluster. The expert opinion evidence chain cluster is generated by summarizing the expert role cluster numbers corresponding to the row index with an attribution value of 1.

9. The method for extracting forensic identification information based on machine learning according to claim 8, characterized in that, The operation of updating the occupancy status according to the evidence-based occupancy conservation ledger includes: Read the expert opinion evidence chain cluster number and the expert role cluster number from the expert opinion evidence chain cluster; In the evidence occupancy conservation ledger, search for the corresponding text fragment number, sharing tag, occupancy chain number, and occupancy status by the identification role cluster number; When the shared token is set to the exclusive token and the occupancy status is unoccupied, the occupancy status is changed to the exclusive status. When the shared flag is set to shared and the occupied status is unoccupied, the occupied status is changed to shared occupied status. When the shared marker is a shared marker and the occupancy status is a shared occupancy status, the evidence chain cluster number of the expert opinion will be appended to the occupancy chain number; When the shared token is an exclusive token and the occupancy status is exclusive, the occupancy status remains unchanged.

10. The method for extracting forensic identification information based on machine learning according to claim 9, characterized in that, The operation of mapping the evidence chain cluster of the expert opinion back to the original text index and performing chain deletion verification includes: Read the evidence chain cluster number of the expert opinion, the cluster number of the expert role, and the corresponding text fragment number; Search the original text index by text fragment number in the judicial appraisal text fragment node table to generate the original text mapping result; The role cluster is deleted from the evidence chain cluster of the expert opinion, resulting in the evidence chain to be verified. Check the integrity of the supporting edges of the identification process between the identification matter role cluster and the identification opinion role cluster in the evidence chain to be verified. If the edges are complete, the evidence chain is marked as closed; if they are missing, the evidence chain is marked as unclosed. Generate deletion verification results based on the closed state of the evidence chain, and output structured forensic identification information based on the original text mapping results, deletion verification results, occupancy status and number consistency results.

Images