An artificial intelligence-based legal text professional review and intelligent annotation system

Abstract

This invention discloses an AI-based legal text professional review and intelligent annotation system, comprising: a text unit sequence generation module for forming text unit sequences; a role identifier sequence generation module for forming role identifier sequences; a hierarchical memory state construction module for forming hierarchical memory states; a bipolar correspondence construction module for forming bipolar correspondences; an improved DeltaNet update module for forming master rule states, exception rule states, and conflict frontier indices; a rollback path generation module for forming rollback paths; an evidence strengthening module for forming evidence strengthening states; and a review annotation output module for forming professional review results, conflict location results, and intelligent annotation results. This invention improves the accuracy of legal text review, conflict location capabilities, and annotation interpretability.

Description

Technical Field

[0001] This invention relates to the field of legal artificial intelligence technology, and in particular to an artificial intelligence-based system for professional review and intelligent annotation of legal texts. Background Technology

[0002] With the increasing demand for digital rule of law, intelligent legal services, and automated legal document processing, professional review and intelligent annotation technologies for legal texts have received widespread attention. Existing contract review, policy verification, or legal document generation systems primarily rely on keyword retrieval, rule template matching, or general language models for clause identification and risk alerts. However, these systems commonly suffer from the following problems in practical applications: Legal texts are lengthy, hierarchical, and complex in their citations. Existing methods struggle to establish stable, continuous semantic memory across chapters, clauses, and paragraphs, leading to gaps between definitions, applicable conditions, and liability content. Main rules and exceptions often have overlapping constraints, coverage limitations, and scattered locations, making it difficult for existing review techniques to accurately trace conflict chains, resulting in unclear risk identification and inaccurate evidence matching. For high-risk issue units, traditional annotation methods often remain at the level of conclusion suggestions, lacking relevant evidence fragments and contextual support consistent with the issue's location, leading to insufficient interpretability, verifiability, and practical applicability of the annotation results.

[0003] Therefore, how to provide a legal text professional review and intelligent annotation system based on artificial intelligence 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 an AI-based legal text professional review and intelligent annotation system. This invention achieves intelligent identification and annotation output of target problem units and related evidence fragments in legal texts through text unit segmentation, standardized role parsing, hierarchical memory state construction, bipolar correspondence modeling, improved DeltaNet recursive update, rollback path tracing, and evidence enhancement processing. It has the advantages of high review accuracy, strong conflict location capability, and good interpretability of annotation results.

[0005] According to an embodiment of the present invention, a legal text professionalism review and intelligent annotation system based on artificial intelligence includes: The text unit sequence generation module is used to acquire the legal text to be reviewed and segment it according to chapter level to form a text unit sequence; The role identifier sequence generation module is used to perform standardized role parsing on the text unit sequence to form a role identifier sequence; The hierarchical memory state construction module is used to establish a role state channel according to the role identifier sequence and map the text unit sequence to the role state channel to form a hierarchical memory state. The bipolar correspondence construction module is used to extract the main rule unit and the exception rule unit from the text unit sequence and perform pairing to form a bipolar correspondence. The improved DeltaNet update module is used to write the hierarchical memory state and bipolar correspondence into the improved DeltaNet, forming the main rule state, the exception rule state, and the conflict front index; The rollback path generation module is used to trace the main rule state and the exception rule state backward from the conflict front index to form a rollback path; The evidence strengthening module is used to calculate the professional risk value of the text units in the rollback path, and adjust the number of update steps of the improved DeltaNet according to the professional risk value to form an evidence strengthening state. The review and annotation output module is used to extract target problem units and related evidence fragments from the evidence strengthening status, and generate professional review results, conflict location results, and intelligent annotation results.

[0006] Optionally, the text unit sequence generation module specifically comprises: Obtain the legal text to be reviewed, and perform character encoding standardization, white space regularization, removal of format marks, and numbering style standardization on the legal text to be reviewed to form a regular text; Extract title numbers, clause numbers, item numbers, and citation marks from the regular text, and build a chapter-level index based on title numbers, clause numbers, item numbers, and line break boundaries; The text is hierarchically segmented according to the chapter-level index to form chapter units, clause units, and sentence units. Sentence units with citation marks are extracted to form citation units. The chapter, clause, sentence / paragraph, and quotation units are jointly arranged according to the hierarchical order and the positional order of the original text. Unit identifiers, hierarchical position identifiers, and start and end position identifiers are written to the chapter, clause, sentence / paragraph, and quotation units to form a text unit sequence.

[0007] Optionally, the role identifier sequence generation module specifically comprises: Read clause units, sentence units, and citation units from the text unit sequence, and build a role parsing record set according to unit identifier, hierarchical position identifier, and start and end position identifier; Semantic trigger word recognition, syntactic relation extraction, and citation pointer matching are performed on the clause units, sentence units, and citation units in the role parsing record set to extract definition expressions, rule expressions, condition expressions, exception expressions, responsibility expressions, and citation expressions, forming a role determination record set; Based on the role determination record set, define the role identifier, main rule role identifier, condition role identifier, exception role identifier, responsibility role identifier, and citation role identifier into the clause unit, sentence unit, and citation unit to form a role mapping record set; The role mapping record set is arranged according to the original position order, so that the role identifiers maintain the positional correspondence with the clause units, sentence units, and citation units, forming a role identifier sequence.

[0008] Optionally, the hierarchical memory state construction module specifically comprises: Read the text unit sequence and the role identifier sequence, and perform corresponding matching according to the unit identifier, hierarchical position identifier and start and end position identifier to form a state mapping record set; Establish definition role status channels, main rule role status channels, condition role status channels, exception role status channels, responsibility role status channels, and citation role status channels based on the role identifier sequence; The clause units, sentence units, and citation units in the state mapping record set are written into the definition role state channel, main rule role state channel, condition role state channel, exception role state channel, responsibility role state channel, and citation role state channel respectively according to the role identifier, forming a channel state record set; The channel status record set is sequentially arranged and aligned according to the hierarchical position identifier and the start and end position identifier to form a hierarchical memory state.

[0009] Optionally, the bipolar correspondence construction module specifically comprises: Read the text unit sequence and role identifier sequence, and perform corresponding matching according to the unit identifier, hierarchical position identifier and start and end position identifier to form a bipolar parsing record set; Extract the segment units with the main rule role identifier from the bipolar parsing record set as the main rule units, and extract the segment units with the exception role identifier as the exception rule units to form a paired candidate record set; Extract the target object expression and applicable condition expression from the main rule unit and exception rule unit in the paired candidate record set, and perform pairing verification according to the consistency relationship of the target object expression and the association relationship of the applicable condition expression to form a paired record set; The paired record sets are arranged in the order of execution according to the hierarchical position identifier and the start and end position identifier. The main rule unit identifier, the exception rule unit identifier, the target object expression and the applicable condition expression are written to each record in the paired record set to form a bipolar correspondence.

[0010] Optionally, the improved DeltaNet update module specifically comprises: Read the definition role status channel, main rule role status channel, condition role status channel, exception role status channel, responsibility role status channel and citation role status channel in the hierarchical memory state, and read the main rule unit identifier, exception rule unit identifier, target expression and applicable condition expression in the bipolar correspondence. Perform corresponding matching according to unit identifier, hierarchical position identifier and start and end position identifier to form an updated associated record set. The update associated record set is assigned a channel according to the role status channel, so that clause units, sentence units and citation units enter the corresponding role status channel, and a channel update path is established in the definition role status channel, main rule role status channel, condition role status channel, exception role status channel, responsibility role status channel and citation role status channel according to the hierarchical position identifier and start and end position identifier. Write the state records corresponding to adjacent positions along the channel update path into the improved DeltaNet to perform differential recursive updates, so that the defined role state channel becomes the defined role state, the main rule role state channel becomes the main rule role state, the condition role state channel becomes the condition role state, the exception role state channel becomes the exception role state, the responsibility role state channel becomes the responsibility role state, and the citation role state channel becomes the citation role state. Write the main rule unit in the bipolar correspondence into the main rule role state channel, write the exception rule unit in the bipolar correspondence into the exception role state channel, and perform correspondence maintenance according to the main rule unit identifier, exception rule unit identifier, target object expression and applicable condition expression, so that the main rule role state and the exception role state are updated independently within the same correspondence range, forming the main rule state and the exception rule state. The main rule state and the exception rule state are compared and the state difference is detected by the main rule unit identifier, the exception rule unit identifier, the hierarchical position identifier and the start and end position identifier. The positions with state differences are written into the index record to form the conflict front index.

[0011] Optionally, the rollback path generation module specifically comprises: Read the conflict front index, main rule status, exception rule status and text unit sequence, and perform corresponding matching according to the main rule unit identifier, exception rule unit identifier, hierarchical position identifier and start and end position identifier to form a rollback parsing record set; Extract the position records corresponding to the conflict front index from the rollback parsing record set, and locate the clause unit, sentence unit and citation unit in the text unit sequence according to the hierarchical position identifier and the start and end position identifier to form a rollback candidate record set; Perform reverse tracing along the text unit sequence on the main rule state and exception rule state in the rollback candidate record set, and establish path connections between the current main rule state and exception rule state and the adjacent previous main rule state and exception rule state to form a rollback connection record set; The rollback connection record set is ordered according to the hierarchical location identifier and the start and end location identifier, and the main rule unit identifier and the exception rule unit identifier are written to the records in the rollback connection record set to form the rollback path.

[0012] Optionally, the evidence enhancement module specifically comprises: The rollback path and the text unit sequence are matched according to the main rule unit identifier, the exception rule unit identifier, the hierarchical position identifier, and the start and end position identifier to form an evidence analysis record set; Extract terminology standardization features, citation completeness features, subject clarity features, right and responsibility correspondence features, and conflict position features from clause units, sentence units, and citation units in the evidence analysis record set. Perform aggregation calculation according to terminology standardization features, citation completeness features, subject clarity features, right and responsibility correspondence features, and conflict position features to form a professional risk value record set; Based on the correspondence between professional risk values ​​and step ranges in the professional risk value record set, step adjustment is performed on the update steps of the improved DeltaNet to form an enhanced update record set; The enhanced update record set is written into the improved DeltaNet to perform enhanced recursive updates. Evidence aggregation is performed on the clause units, sentence units, and citation units in the rollback path according to the enhanced recursive update results to form an evidence-enhanced state.

[0013] Optionally, the review annotation output module specifically includes: The evidence strengthening status is merged according to the main rule unit identifier, the exception rule unit identifier, the hierarchical position identifier, and the start and end position identifier to form a review candidate record set; For clause units, sentence units, and citation units in the candidate record set for review, issue location is performed according to the path connection relationship in the return path and the evidence aggregation result in the evidence strengthening state, forming a target issue unit record set; Based on the main rule unit identifier, exception rule unit identifier, hierarchical position identifier, and start and end position identifier in the target problem unit record set, the corresponding clause unit, sentence unit, and quotation unit are extracted in the evidence strengthening state to form a related evidence fragment record set. The target problem unit record set and the related evidence fragment record set are arranged accordingly to form professional review results, conflict location results and intelligent annotation results.

[0014] The beneficial effects of this invention are: This invention, through the collaborative setup of a text unit sequence generation module, a role identifier sequence generation module, a hierarchical memory state construction module, and a bipolar correspondence construction module, transforms the original continuous legal text under review into a structured expression with chapter hierarchy, role affiliation, and rule correspondence. Furthermore, an improved DeltaNet update module performs multi-channel recursive updates to the definition role state, main rule role state, condition role state, exception role state, liability role state, and citation role state. This ensures that the previously scattered definitions, rules, conditions, exceptions, and liabilities within the legal text maintain positional correspondence and semantic coherence within a unified review framework. This effectively reduces the problems of semantic disjointness, complex rule relationships, and unclear clause connections in the processing of long legal texts, and improves the ability to identify conflicts between main rules and exception rules, as well as the modeling accuracy of complex legal logic chains.

[0015] This invention further utilizes a rollback path generation module, an evidence enhancement module, and a review annotation output module to perform reverse tracing of the conflict front index corresponding to the location. It also adjusts the update steps of the improved DeltaNet based on a professional risk value, thereby strengthening and aggregating clause units, sentence units, and citation units related to the conflict location, thus forming associated evidence fragments that match the target issue unit. This not only outputs professional review results and conflict location results but also intelligent annotation results consistent with the issue location, making the review results more interpretable, verifiable, and supportive of evidence. Simultaneously, it helps reduce omissions, misjudgments, and redundant comparisons during manual review, improving the accuracy and practicality of professional review and annotation processing of legal texts. Attached Figure Description

[0016] 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 structural diagram of an artificial intelligence-based legal text professional review and intelligent annotation system proposed in this invention; Figure 2 This is a flowchart of the operation of an artificial intelligence-based legal text professional review and intelligent annotation system proposed in this invention; Figure 3 This is a schematic diagram of the improved DeltaNet update process of an artificial intelligence-based legal text professional review and intelligent annotation system proposed in this invention. Detailed Implementation

[0017] 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.

[0018] refer to Figures 1-3 A legal text professional review and intelligent annotation system based on artificial intelligence, comprising: The text unit sequence generation module is used to acquire the legal text to be reviewed and segment it according to chapter level to form a text unit sequence; The role identifier sequence generation module is used to perform standardized role parsing on the text unit sequence to form a role identifier sequence; The hierarchical memory state construction module is used to establish a role state channel according to the role identifier sequence and map the text unit sequence to the role state channel to form a hierarchical memory state. The bipolar correspondence construction module is used to extract the main rule unit and the exception rule unit from the text unit sequence and perform pairing to form a bipolar correspondence. The improved DeltaNet update module is used to write the hierarchical memory state and bipolar correspondence into the improved DeltaNet, forming the main rule state, the exception rule state, and the conflict front index; The rollback path generation module is used to trace the main rule state and the exception rule state backward from the conflict front index to form a rollback path; The evidence strengthening module is used to calculate the professional risk value of the text units in the rollback path, and adjust the number of update steps of the improved DeltaNet according to the professional risk value to form an evidence strengthening state. The review and annotation output module is used to extract target problem units and related evidence fragments from the evidence strengthening status, and generate professional review results, conflict location results, and intelligent annotation results.

[0019] In this embodiment, the text unit sequence generation module specifically comprises: The process involves obtaining the legal text to be reviewed, performing a unified mapping on the different character representations in the legal text to ensure that the same character maintains the same encoding form throughout the text, performing regularization on consecutive spaces, tabs, and line breaks without content while preserving paragraph line breaks, clearing layout marks on headers, footers, page numbers, and separators, and regularizing numbering styles for Arabic numerals, Chinese characters, and parentheses to form a regularized text. Read the text content segment by segment from the regularized text, locate the occurrence of title number, clause number, item number and citation mark in the regularized text, and build a chapter-level index according to the hierarchical relationship corresponding to the title number, the sequential relationship corresponding to the clause number, the subordinate relationship corresponding to the item number and the separation relationship corresponding to the line break boundary, so that the chapter-level index can correspond to the chapter boundary, clause boundary and item boundary in the regularized text. The text is hierarchically segmented according to the chapter-level index. First, chapter content is extracted based on chapter boundaries to form chapter units. Then, clause content is extracted based on clause boundaries within the chapter content to form clause units. Next, sentence segments are extracted based on sentence boundaries and citation mark positions within the clause content to form sentence segments units. Citation extraction is performed on sentence segments with citation marks to separate the citation content from the sentence segments and form citation units. The chapter, clause, paragraph, and quotation units are jointly arranged according to the hierarchical order and the positional order of the original text. This ensures that the chapter, clause, paragraph, and quotation units maintain hierarchical and positional correspondences in the arrangement result. Unit identifiers, hierarchical position identifiers, and start and end position identifiers are written to the chapter, clause, paragraph, and quotation units. The unit identifiers are used to distinguish different units, the hierarchical position identifiers are used to indicate the chapter level position of the unit, and the start and end position identifiers are used to indicate the start and end positions of the unit in the organized text, forming a text unit sequence.

[0020] In this embodiment, the role identifier sequence generation module specifically comprises: Read the clause units, sentence units, and quotation units in the text unit sequence, locate the corresponding text content according to the unit identifier, determine the hierarchical position of the clause units, sentence units, and quotation units in the text unit sequence according to the hierarchical position identifier, determine the position range of the clause units, sentence units, and quotation units in the legal text to be reviewed according to the start and end position identifier, and write the unit identifier, hierarchical position identifier, and start and end position identifier into the same record unit to form a role resolution record set; Semantic trigger word recognition, syntactic relation extraction, and citation matching are performed on clause units, sentence units, and citation units in the role analysis record set. Definition expressions, rule expressions, condition expressions, exception expressions, responsibility expressions, and citation expressions are identified according to the word combination positions in the text content. The corresponding positions of definition expressions, rule expressions, condition expressions, exception expressions, responsibility expressions, and citation expressions in clause units, sentence units, and citation units are determined according to syntactic connection relations. Citation matching is completed according to the positional correspondence between citation tags and citation content to form a role determination record set. Based on the role determination record set, write definition role identifiers, main rule role identifiers, condition role identifiers, exception role identifiers, responsibility role identifiers, and citation role identifiers into clause units, sentence units, and citation units, so that definition expressions correspond to definition role identifiers, rule expressions correspond to main rule role identifiers, condition expressions correspond to condition role identifiers, exception expressions correspond to exception role identifiers, responsibility expressions correspond to responsibility role identifiers, and citation expressions correspond to citation role identifiers, thus forming a role mapping record set; The role mapping record set is arranged according to the original position order, and the role identifiers in the role mapping record set are arranged sequentially into the corresponding clause unit, sentence unit and quotation unit, so that the role identifiers maintain the positional correspondence with the clause unit, sentence unit and quotation unit, forming a role identifier sequence.

[0021] In this embodiment, the hierarchical memory state construction module specifically comprises: Read the text unit sequence and role identifier sequence, locate the clause unit, sentence segment unit and quotation unit in the text unit sequence according to the unit identifier, determine the position of the clause unit, sentence segment unit and quotation unit in the chapter level according to the hierarchical position identifier, determine the position range of the clause unit, sentence segment unit and quotation unit in the legal text to be reviewed according to the start and end position identifier, and then write the role identifier corresponding to the unit identifier in the role identifier sequence into the same record unit to form a status mapping record set; Based on the category distribution of each role identifier in the role identifier sequence, establish definition role status channels, main rule role status channels, condition role status channels, exception role status channels, responsibility role status channels, and citation role status channels, so that the definition role status channel corresponds to the definition role identifier, the main rule role status channel corresponds to the main rule role identifier, the condition role status channel corresponds to the condition role identifier, the exception role status channel corresponds to the exception role identifier, the responsibility role status channel corresponds to the responsibility role identifier, and the citation role status channel corresponds to the citation role identifier. The clause units, sentence units, and citation units in the state mapping record set are written into the definition role state channel, main rule role state channel, condition role state channel, exception role state channel, responsibility role state channel, and citation role state channel respectively according to the role identifier. This ensures that each record maintains the correspondence between the unit identifier, hierarchical position identifier, and start and end position identifier in the corresponding role state channel, and the channel state record set is formed according to the writing results. The channel status record set is sequentially arranged and aligned according to the hierarchical position identifier and the start and end position identifier, so that the records in the same chapter level are arranged according to the original text order, and the corresponding records in different role status channels are established according to the position range, and a hierarchical memory status is formed according to the arrangement and alignment results.

[0022] In this embodiment, the bipolar correspondence construction module is specifically as follows: Read the text unit sequence and the role identifier sequence, locate the corresponding sentence segment unit in the text unit sequence according to the unit identifier, determine the position of the sentence segment unit in the chapter level according to the hierarchical position identifier, determine the position range of the sentence segment unit in the legal text to be reviewed according to the start and end position identifier, and then write the role identifier corresponding to the unit identifier in the role identifier sequence into the same record unit, so that the sentence segment unit, hierarchical position identifier, start and end position identifier and role identifier maintain the corresponding relationship, forming a bipolar parsing record set; Segment units with main rule role identifiers are selected from the bipolar parsing record set, and the selected results are determined as main rule units. Segment units with exception role identifiers are selected from the bipolar parsing record set, and the selected results are determined as exception rule units. Then, the main rule units and exception rule units are written into the same candidate relation range according to the hierarchical position identifier and the start and end position identifier to form a paired candidate record set. The target object expression and applicable condition expression are extracted from the main rule unit and the exception rule unit in the paired candidate record set. The target object expression is formed by identifying the object referent content in the main rule unit and the exception rule unit, and the applicable condition expression is formed by identifying the condition limitation content in the main rule unit and the exception rule unit. Then, a pairing check is performed according to the consistency relationship between the target object expression and the association relationship between the applicable condition expression. The main rule unit and the exception rule unit with consistent target object expression establish an object correspondence relationship, and the main rule unit and the exception rule unit with related applicable condition expression establish a condition correspondence relationship, thus forming a paired record set. The paired record set is sequentially arranged according to the hierarchical position identifier and the start and end position identifier, so that each record in the paired record set is arranged in the order of the chapter hierarchy and the original text position. Then, the main rule unit identifier, the exception rule unit identifier, the pointing object expression, and the applicable condition expression are written to each record in the paired record set. The main rule unit identifier is used to identify the corresponding main rule unit, the exception rule unit identifier is used to identify the corresponding exception rule unit, the pointing object expression is used to represent the object correspondence between the main rule unit and the exception rule unit, and the applicable condition expression is used to represent the condition correspondence between the main rule unit and the exception rule unit, forming a bipolar correspondence.

[0023] In this embodiment, the improved DeltaNet update module is specifically as follows: The system reads the definition role status channel, main rule role status channel, condition role status channel, exception role status channel, responsibility role status channel, and citation role status channel from the hierarchical memory state. It also reads the main rule unit identifier, exception rule unit identifier, target expression, and applicable condition expression from the bipolar correspondence. Based on the unit identifier, it locates the clause unit, sentence segment unit, and citation unit in the hierarchical memory state. Based on the hierarchical position identifier, it determines the position of the clause unit, sentence segment unit, and citation unit in the chapter level. Based on the start and end position identifier, it determines the position range of the clause unit, sentence segment unit, and citation unit in the legal text to be reviewed. It writes the main rule unit identifier, exception rule unit identifier, target expression, and applicable condition expression into the record unit corresponding to the clause unit, sentence segment unit, and citation unit, so that the position record in the hierarchical memory state and the relationship record in the bipolar correspondence converge into the same record structure, forming an updated associated record set. The update associated record set is assigned a channel according to the role status channel. Records with the definition role identifier are sent to the definition role status channel, records with the main rule role identifier are sent to the main rule role status channel, records with the condition role identifier are sent to the condition role status channel, records with the exception role identifier are sent to the exception role status channel, records with the responsibility role identifier are sent to the responsibility role status channel, and records with the citation role identifier are sent to the citation role status channel. The record positions are arranged in each role status channel according to the hierarchical position identifier and the start and end position identifier. Records in the same role status channel maintain the original text expansion order, and corresponding records in different role status channels maintain the position comparison relationship, forming a channel update path. Along the channel update path, the state records corresponding to adjacent positions are written into the improved DeltaNet to perform differential recursive update. The current position state record and the adjacent previous state record are read, and the content difference is calculated on the current position state record and the adjacent previous state record. The content difference is written into the recursive update position of the improved DeltaNet, so that the current position state record completes the recursive state update based on the adjacent previous state record. This makes the defined role state channel form the defined role state, the main rule role state channel form the main rule role state, the condition role state channel form the condition role state, the exception role state channel form the exception role state, the responsibility role state channel form the responsibility role state, and the citation role state channel form the citation role state. Among them, the content difference is used to represent the changes in text content and role attribution of adjacent position state records, and the recursive update position is used to carry the state correction result of the current position state record on the adjacent previous state record. The main rule unit in the bipolar correspondence is written into the main rule role state channel, and the exception rule unit in the bipolar correspondence is written into the exception role state channel. The correspondence is maintained according to the main rule unit identifier, exception rule unit identifier, object expression, and applicable condition expression. This ensures that the update position of the main rule unit in the main rule role state channel and the update position of the exception rule unit in the exception role state channel are in the same correspondence range. The main rule role state and the exception role state in the same correspondence range are updated independently along their respective role state channels without state overwriting, forming the main rule state and the exception rule state. The correspondence maintenance is used to maintain the pairing relationship between the main rule unit and the exception rule unit in terms of object content and condition content. The same correspondence range is used to represent the corresponding position range of the main rule state and the exception rule state in the hierarchical memory state. Perform position comparison and state difference detection on the main rule state and the exception rule state according to the main rule unit identifier, exception rule unit identifier, hierarchical position identifier, and start and end position identifier. Locate the corresponding main rule state and exception rule state according to the main rule unit identifier and exception rule unit identifier. Perform position comparison on the main rule state and the exception rule state according to the hierarchical position identifier and the start and end position identifier. Perform state difference detection on the state content in the position comparison result. Write the positions where state differences exist into index records, so that the index records correspond to the position range where the main rule state and the exception rule state differ, forming a conflict front index. Among them, the state difference detection is used to identify the state change differences between the main rule state and the exception rule state within the same corresponding range, and the index records are used to record the position range corresponding to the state change differences.

[0024] In this embodiment, the rollback path generation module specifically comprises: Read the conflict front index, main rule status, exception rule status, and text unit sequence. Locate the corresponding status record in the main rule status according to the main rule unit identifier, locate the corresponding status record in the exception rule status according to the exception rule unit identifier, determine the position of the corresponding status record in the chapter level according to the hierarchical position identifier, determine the position range of the corresponding status record in the legal text to be reviewed according to the start and end position identifier, and write the index position in the conflict front index, the status record in the main rule status, the status record in the exception rule status, and the position record in the text unit sequence into the same record unit to form a rollback parsing record set. Extract the position records corresponding to the conflict front index from the rollback parsing record set, locate the clause units, sentence units, and citation units in the text unit sequence according to the hierarchical position identifier and the start and end position identifier, establish the position correspondence between the index position in the conflict front index and the clause units, sentence units, and citation units in the text unit sequence, and write the position correspondence results into the candidate record unit to form the rollback candidate record set; The main rule status and exception rule status in the rollback candidate record set are traced in reverse along the text unit sequence. The main rule status and exception rule status are read one by one from the later position to the earlier position according to the start and end position identifiers. The main rule status and exception rule status at the current position are established with the main rule status and exception rule status of the adjacent previous position. The path connection corresponds to the state transmission relationship between the position pointed to by the conflict front index and the adjacent previous position. The path connection result is written into the record unit to form the rollback connection record set. The rollback connection record set is ordered according to the hierarchical position identifier and the start and end position identifier, so that the records in the rollback connection record set are arranged in the order of chapter hierarchy and original text position. The main rule unit identifier and the exception rule unit identifier are written to the records in the rollback connection record set. The main rule unit identifier is used to represent the main rule unit corresponding to the rollback path, and the exception rule unit identifier is used to represent the exception rule unit corresponding to the rollback path, thus forming the rollback path.

[0025] In this embodiment, the evidence enhancement module specifically includes: When performing corresponding matching between the return path and the text unit sequence according to the main rule unit identifier, the exception rule unit identifier, the hierarchical position identifier, and the start and end position identifier, the rule corresponding unit in the return path is determined by the main rule unit identifier and the exception rule unit identifier, the position of the rule corresponding unit in the chapter hierarchy is determined by the hierarchical position identifier, and the position range of the rule corresponding unit in the legal text to be reviewed is determined by the start and end position identifier. Then, the path record in the return path and the clause unit, sentence unit, and quotation unit in the text unit sequence are gathered into the same record structure, so that each record simultaneously maintains the rule correspondence, hierarchical position relationship, and original text position relationship, forming an evidence analysis record set; When extracting terminology standardization features, citation completeness features, subject clarity features, right-responsibility correspondence features, and conflict location features from clause units, sentence units, and citation units in the evidence analysis record set, terminology standardization features are formed by statistically analyzing the consistency of terminology usage in clause units, sentence units, and citation units; citation completeness features are formed by statistically analyzing the correspondence between citation marks and citation content; subject clarity features are formed by statistically analyzing the clarity of subject expression in clause units, sentence units, and citation units; right-responsibility correspondence features are formed by statistically analyzing the correspondence between right expression and responsibility expression in clause units, sentence units, and citation units; and conflict location features are formed by statistically analyzing the overlap between the corresponding position of the conflict front index and the corresponding position of the rollback path. Then, aggregation calculations are performed according to terminology standardization features, citation completeness features, subject clarity features, right-responsibility correspondence features, and conflict location features, so that the same record corresponds to a professional risk value, forming a professional risk value record set. Based on the correspondence between professional risk values ​​and step ranges in the professional risk value record set, when adjusting the number of update steps for the improved DeltaNet, the professional risk values ​​in the professional risk value record set are mapped to a preset step range, so that larger professional risk values ​​correspond to larger update steps and smaller professional risk values ​​correspond to smaller update steps. The correspondence between professional risk values ​​and update steps is written into the same record unit, so that each record maintains the correspondence between professional risk values ​​and update steps, forming an enhanced update record set. When the enhanced update record set is written into the improved DeltaNet to perform enhanced recursive updates, recursive state updates are performed on the corresponding records according to the update steps in the enhanced update record set. This results in records with larger update steps forming stronger state enhancements in the improved DeltaNet, and records with smaller update steps forming weaker state enhancements in the improved DeltaNet. Then, evidence aggregation is performed on the clause units, sentence units, and citation units in the rollback path according to the enhanced recursive update results. This allows clause units, sentence units, and citation units in the same rollback path to converge into evidence expression results according to the path connection relationship, forming evidence-enhanced states.

[0026] In this embodiment, the review annotation output module specifically comprises: When merging the execution positions of evidence strengthening status according to the main rule unit identifier, exception rule unit identifier, hierarchical position identifier, and start and end position identifier, the main rule unit identifier and exception rule unit identifier determine the same rule pairing range, the hierarchical position identifier determines the arrangement position of clause unit, sentence unit, and quotation unit in the chapter hierarchy, and the start and end position identifier determines the position range of clause unit, sentence unit, and quotation unit in the legal text to be examined. Then, records that are in the same rule pairing range and are in consecutive positions are merged into the same record structure, so that the same record structure simultaneously maintains the rule pairing relationship, hierarchical position relationship, and original text position relationship, forming a candidate record set for examination. When locating issues for clause units, sentence units, and citation units in the candidate record set based on the path connection relationship in the rollback path and the evidence aggregation result in the evidence strengthening state, the connection order of clause units, sentence units, and citation units in the rollback direction is determined according to the path connection relationship in the rollback path, and the evidence aggregation position of clause units, sentence units, and citation units in the same path is determined according to the evidence aggregation result in the evidence strengthening state. Clause units, sentence units, and citation units that simultaneously meet the path connection set and evidence aggregation conditions are marked as issue positions, and the issue positions are written into the corresponding records to form a target issue unit record set. When extracting corresponding clause units, sentence units, and quotation units in the evidence strengthening state, based on the main rule unit identifier, exception rule unit identifier, hierarchical position identifier, and start-end position identifier in the target issue unit record set, the extraction scope is limited by the main rule unit identifier and exception rule unit identifier, the position of the extracted record in the chapter level is limited by the hierarchical position identifier, and the position range of the extracted record in the legal text to be examined is limited by the start-end position identifier. Clause units, sentence units, and quotation units that fall within the extraction scope and position range are written into the same evidence record, so that the same evidence record maintains the rule pairing relationship and position relationship consistent with the target issue unit, forming a set of related evidence fragment records. When performing corresponding arrangement on the target issue unit record set and the associated evidence fragment record set, the correspondence between the target issue unit record set and the associated evidence fragment record set is established according to the main rule unit identifier, the exception rule unit identifier, the hierarchical position identifier, and the start and end position identifier. The issue positions in the target issue unit record set and the evidence content in the associated evidence fragment record set are arranged in the same correspondence order, so that each output record contains both the issue position and the evidence content, forming professional review results, conflict location results, and intelligent annotation results.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a legal document intelligent review scenario. The documents to be reviewed include contract texts, policy documents, authorization documents, confidentiality agreements, and business correspondence. These texts generally suffer from complex hierarchical structures, scattered definitions, overlapping main rules and exception rules, discontinuous citations, and significant discrepancies between liability clauses and applicable conditions. Traditional keyword retrieval and rule matching methods often encounter problems when processing such texts, such as definitions not being incorporated into subsequent judgments, exception clauses failing to correctly cover main rules, incomplete extraction of evidence fragments corresponding to the same liability clause, and insufficient evidence to support annotations. This necessitates repeated manual review of the original text, resulting in unsatisfactory review efficiency and consistency. To address this issue, in this example, the legal text to be reviewed is input into a text unit sequence generation module, forming a text unit sequence composed of chapter units, clause units, sentence units, and citation units. Then, a role identifier sequence generation module identifies definition expressions, rule expressions, condition expressions, exception expressions, liability expressions, and citation expressions, forming a role identifier sequence. The hierarchical memory state construction module maps content from different roles to corresponding role state channels. The bipolar correspondence construction module pairs main rule units and exception rule units according to their target object expression and applicable condition expression. The improved DeltaNet update module performs differential recursive updates within each role state channel and forms a conflict front index between the main rule state and the exception rule state. The rollback path generation module traces the rule chain backward from the conflict front index. The evidence strengthening module adjusts the update steps based on the professional risk value, strengthening and aggregating conflict-intensive locations, missing citation locations, and locations where responsibilities do not correspond. The review and annotation output module finally provides the target problem unit, related evidence fragments, professional review results, conflict location results, and intelligent annotation results.

[0028] In practical application, a total of 1200 legal text samples of the same type were selected, including 720 contracts, 260 institutional documents, 110 authorization documents, 70 confidentiality agreements, and 40 business correspondence. The length of a single text ranged from 3200 to 18600 words, with an average length of 8420 words. Manual pre-annotation revealed 186 instances of missing definitions, 243 instances of conflicts between main rules and exception rules, 167 instances of incomplete citations, 154 instances of unclear subject descriptions, and 219 instances of mismatched rights and responsibilities, totaling 969 valid problem points. Using existing keyword rule systems as the control group and the system of this invention as the experimental group, the systems were run continuously under the same hardware environment. The average processing time per document, problem location accuracy, conflict identification recall rate, accuracy of matching related evidence fragments, average time spent on manual review, and annotation adoption rate were recorded. The results are shown in Table 1. Table 1 Comparison of Legal Text Review Effectiveness

[0029] As shown in Table 1, this invention significantly outperforms the control group in processing speed, problem localization, conflict identification, and evidence matching. The average processing time per document was reduced by 7.3 seconds, indicating that text unit sequence generation, role status channel updates, and rollback path tracking reduce invalid traversals. The conflict identification recall rate increased from 69.4% to 91.2%, demonstrating that the parallel retention of main rule states and exception rule states, along with the conflict frontier index, effectively captures rule conflicts scattered across different clause locations. The accuracy rate of matching associated evidence fragments increased from 64.9% to 92.5%, indicating that the evidence strengthening module can stably correspond the target problem unit to the original evidence fragment. The average time for manual review decreased to 3.1 minutes per document, and the annotation adoption rate increased to 87.6%, demonstrating that the professional review results, conflict localization results, and intelligent annotation results output by this invention have strong interpretability and practical usability, effectively solving the problems of inaccurate localization, unclear evidence, and heavy manual review burden in the review of long legal texts.

[0030] 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 legal text professional review and intelligent annotation system based on artificial intelligence, characterized in that, include: The text unit sequence generation module is used to acquire the legal text to be reviewed and segment it according to chapter level to form a text unit sequence; The role identifier sequence generation module is used to perform standardized role parsing on the text unit sequence to form a role identifier sequence; The hierarchical memory state construction module is used to establish a role state channel according to the role identifier sequence and map the text unit sequence to the role state channel to form a hierarchical memory state. The bipolar correspondence construction module is used to extract the main rule unit and the exception rule unit from the text unit sequence and perform pairing to form a bipolar correspondence. The improved DeltaNet update module is used to write the hierarchical memory state and bipolar correspondence into the improved DeltaNet, forming the main rule state, the exception rule state, and the conflict front index; The rollback path generation module is used to trace the main rule state and the exception rule state backward from the conflict front index to form a rollback path; The evidence strengthening module is used to calculate the professional risk value of the text units in the rollback path, and adjust the number of update steps of the improved DeltaNet according to the professional risk value to form an evidence strengthening state. The review and annotation output module is used to extract target problem units and related evidence fragments from the evidence strengthening status, and generate professional review results, conflict location results, and intelligent annotation results.

2. The legal text professionalism review and intelligent annotation system based on artificial intelligence according to claim 1, characterized in that, The text unit sequence generation module is specifically as follows: Obtain the legal text to be reviewed, and perform character encoding standardization, white space regularization, removal of format marks, and numbering style standardization on the legal text to be reviewed to form a regular text; Extract title numbers, clause numbers, item numbers, and citation marks from the regular text, and build a chapter-level index based on title numbers, clause numbers, item numbers, and line break boundaries; The text is hierarchically segmented according to the chapter-level index to form chapter units, clause units, and sentence units. Sentence units with citation marks are extracted to form citation units. The chapter, clause, sentence / paragraph, and quotation units are jointly arranged according to the hierarchical order and the positional order of the original text. Unit identifiers, hierarchical position identifiers, and start and end position identifiers are written to the chapter, clause, sentence / paragraph, and quotation units to form a text unit sequence.

3. The legal text professionalism review and intelligent annotation system based on artificial intelligence according to claim 1, characterized in that, The specific function of the role identifier sequence generation module is as follows: Read clause units, sentence units, and citation units from the text unit sequence, and build a role parsing record set according to unit identifier, hierarchical position identifier, and start and end position identifier; Semantic trigger word recognition, syntactic relation extraction, and citation pointer matching are performed on the clause units, sentence units, and citation units in the role parsing record set to extract definition expressions, rule expressions, condition expressions, exception expressions, responsibility expressions, and citation expressions, forming a role determination record set; Based on the role determination record set, define the role identifier, main rule role identifier, condition role identifier, exception role identifier, responsibility role identifier, and citation role identifier into the clause unit, sentence unit, and citation unit to form a role mapping record set; The role mapping record set is arranged according to the original position order, so that the role identifiers maintain the positional correspondence with the clause units, sentence units, and citation units, forming a role identifier sequence.

4. The legal text professionalism review and intelligent annotation system based on artificial intelligence according to claim 1, characterized in that, The hierarchical memory state construction module is specifically as follows: Read the text unit sequence and the role identifier sequence, and perform corresponding matching according to the unit identifier, hierarchical position identifier and start and end position identifier to form a state mapping record set; Establish definition role status channels, main rule role status channels, condition role status channels, exception role status channels, responsibility role status channels, and citation role status channels based on the role identifier sequence; The clause units, sentence units, and citation units in the state mapping record set are written into the definition role state channel, main rule role state channel, condition role state channel, exception role state channel, responsibility role state channel, and citation role state channel respectively according to the role identifier, forming a channel state record set; The channel status record set is sequentially arranged and aligned according to the hierarchical position identifier and the start and end position identifier to form a hierarchical memory state.

5. The legal text professionalism review and intelligent annotation system based on artificial intelligence according to claim 1, characterized in that, The bipolar correspondence construction module is specifically as follows: Read the text unit sequence and role identifier sequence, and perform corresponding matching according to the unit identifier, hierarchical position identifier and start and end position identifier to form a bipolar parsing record set; Extract the segment units with the main rule role identifier from the bipolar parsing record set as the main rule units, and extract the segment units with the exception role identifier as the exception rule units to form a paired candidate record set; Extract the target object expression and applicable condition expression from the main rule unit and exception rule unit in the paired candidate record set, and perform pairing verification according to the consistency relationship of the target object expression and the association relationship of the applicable condition expression to form a paired record set; The paired record sets are arranged in the order of execution according to the hierarchical position identifier and the start and end position identifier. The main rule unit identifier, the exception rule unit identifier, the target object expression and the applicable condition expression are written to each record in the paired record set to form a bipolar correspondence.

6. The legal text professionalism review and intelligent annotation system based on artificial intelligence according to claim 1, characterized in that, The improved DeltaNet update module specifically comprises: Read the definition role status channel, main rule role status channel, condition role status channel, exception role status channel, responsibility role status channel and citation role status channel in the hierarchical memory state, and read the main rule unit identifier, exception rule unit identifier, target expression and applicable condition expression in the bipolar correspondence. Perform corresponding matching according to unit identifier, hierarchical position identifier and start and end position identifier to form an updated associated record set. The update associated record set is assigned a channel according to the role status channel, so that clause units, sentence units and citation units enter the corresponding role status channel, and a channel update path is established in the definition role status channel, main rule role status channel, condition role status channel, exception role status channel, responsibility role status channel and citation role status channel according to the hierarchical position identifier and start and end position identifier. Write the state records corresponding to adjacent positions along the channel update path into the improved DeltaNet to perform differential recursive updates, so that the defined role state channel becomes the defined role state, the main rule role state channel becomes the main rule role state, the condition role state channel becomes the condition role state, the exception role state channel becomes the exception role state, the responsibility role state channel becomes the responsibility role state, and the citation role state channel becomes the citation role state. Write the main rule unit in the bipolar correspondence into the main rule role state channel, write the exception rule unit in the bipolar correspondence into the exception role state channel, and perform correspondence maintenance according to the main rule unit identifier, exception rule unit identifier, target object expression and applicable condition expression, so that the main rule role state and the exception role state are updated independently within the same correspondence range, forming the main rule state and the exception rule state. The main rule state and the exception rule state are compared and the state difference is detected by the main rule unit identifier, the exception rule unit identifier, the hierarchical position identifier and the start and end position identifier. The positions with state differences are written into the index record to form the conflict front index.

7. The legal text professionalism review and intelligent annotation system based on artificial intelligence according to claim 1, characterized in that, The rollback path generation module specifically comprises: Read the conflict front index, main rule status, exception rule status and text unit sequence, and perform corresponding matching according to the main rule unit identifier, exception rule unit identifier, hierarchical position identifier and start and end position identifier to form a rollback parsing record set; Extract the position records corresponding to the conflict front index from the rollback parsing record set, and locate the clause unit, sentence unit and citation unit in the text unit sequence according to the hierarchical position identifier and the start and end position identifier to form a rollback candidate record set; Perform reverse tracing along the text unit sequence on the main rule state and exception rule state in the rollback candidate record set, and establish path connections between the current main rule state and exception rule state and the adjacent previous main rule state and exception rule state to form a rollback connection record set; The rollback connection record set is ordered according to the hierarchical location identifier and the start and end location identifier, and the main rule unit identifier and the exception rule unit identifier are written to the records in the rollback connection record set to form the rollback path.

8. The legal text professionalism review and intelligent annotation system based on artificial intelligence according to claim 1, characterized in that, The evidence enhancement module specifically includes: The rollback path and the text unit sequence are matched according to the main rule unit identifier, the exception rule unit identifier, the hierarchical position identifier, and the start and end position identifier to form an evidence analysis record set; Extract terminology standardization features, citation completeness features, subject clarity features, right and responsibility correspondence features, and conflict position features from clause units, sentence units, and citation units in the evidence analysis record set. Perform aggregation calculation according to terminology standardization features, citation completeness features, subject clarity features, right and responsibility correspondence features, and conflict position features to form a professional risk value record set; Based on the correspondence between professional risk values ​​and step ranges in the professional risk value record set, step adjustment is performed on the update steps of the improved DeltaNet to form an enhanced update record set; The enhanced update record set is written into the improved DeltaNet to perform enhanced recursive updates. Evidence aggregation is performed on the clause units, sentence units, and citation units in the rollback path according to the enhanced recursive update results to form an evidence-enhanced state.

9. The legal text professionalism review and intelligent annotation system based on artificial intelligence according to claim 1, characterized in that, The review annotation output module is specifically as follows: The evidence strengthening status is merged according to the main rule unit identifier, the exception rule unit identifier, the hierarchical position identifier, and the start and end position identifier to form a review candidate record set; For clause units, sentence units, and citation units in the candidate record set for review, issue location is performed according to the path connection relationship in the return path and the evidence aggregation result in the evidence strengthening state, forming a target issue unit record set; Based on the main rule unit identifier, exception rule unit identifier, hierarchical position identifier, and start and end position identifier in the target problem unit record set, the corresponding clause unit, sentence unit, and quotation unit are extracted in the evidence strengthening state to form a related evidence fragment record set. The target problem unit record set and the related evidence fragment record set are arranged accordingly to form professional review results, conflict location results and intelligent annotation results.

Images