A complaint-oriented element structured extraction and automatic backfilling method

Abstract

The application discloses a kind of element structured extraction and automatic backfilling method for complaint, carries out fact proposition analysis to original complaint material, extracts fact element, identifies fact limiting element and its limiting object, constructs litigation element structure;According to the field semantic attribute of target complaint template and the field mapping relationship of the fact behavior, semantic role and limiting state corresponding to candidate litigation element, candidate backfilling content is generated and automatic backfilling is executed;Again carry out fact proposition analysis to backfilling content, establish the proposition level mapping relationship of original fact proposition and backfilling fact proposition, identify fact limiting state change;When satisfying preset fact change condition, determine abnormal backfilling field or target text range, execute field rollback or local regeneration again and verify again.The application can keep the fact relationship and limiting state before and after automatic backfilling, improve the accuracy of complaint element extraction, field backfilling and fact semantic consistency.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and intelligent processing of legal documents, specifically a method for the structured extraction and automatic backfilling of elements in a complaint. Background Technology

[0002] As the level of informatization in judicial operations continues to improve, the generation of indictments is gradually shifting from manual editing to template-based and automated processes. In the process of compiling indictment materials, it is typically necessary to extract litigation-related elements such as parties, claims, factual acts, time, place, and amount from party statements, case factual materials, identity information, and other textual materials, and then fill them in according to preset field positions or text areas in the indictment. Existing technologies usually employ keyword matching, named entity recognition, preset rules, or large language models to extract relevant information from the original materials, and then complete the indictment content filling based on field names, template variables, or preset mapping relationships.

[0003] However, the information in the indictment materials does not all appear in a standard form that directly corresponds to the fields in the complaint. The same facts of the case may be recorded in different statements or different materials, and there may be limiting relationships such as negation, speculation, vague quantity, and uncertainty of the subject among the elements of the facts, such as the subject, the act, and the object. For example, expressions such as "may have been committed by the defendant," "approximately 30,000 yuan was paid," and "the specific time of occurrence cannot be determined" in the original materials not only contain the subject, amount, or time elements to be extracted, but also limit the degree of certainty or scope of the corresponding facts. If element extraction is based solely on keywords, entity categories, or semantic similarity of fields, it is easy to directly convert the limited candidate elements into certain elements, or to separate the limiting information from the actual factual object it limits.

[0004] Furthermore, when automatically backfilling the extracted results into the complaint, the factual expressions in the original materials do not always correspond one-to-one with the target fields of the complaint. There may be situations where one factual proposition corresponds to multiple fields, multiple factual fragments together form a single target field, and the same element has different semantic roles in different factual relationships. If replacement and backfilling are performed directly based on template variables or field names, errors in element-field correspondence, loss of factual constraints, or mixing of information between different factual relationships may occur, resulting in the generated complaint content exceeding the scope of factual semantics supported by the original materials.

[0005] In view of this, the present invention provides a method for structured extraction and automatic backfilling of elements for indictments, in order to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide a method for structured extraction and automatic backfilling of elements for indictments, which solves the technical problems of lost factual constraints, mismatched fields, and distorted factual semantics caused by the scattered expression of litigation elements, complex limiting relationships, and inaccurate correspondence between elements and target fields in the existing automatic generation of indictments.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for structured extraction and automatic backfilling of elements for indictments, comprising the following steps:

[0008] Obtain the original complaint materials and the target complaint template including multiple target fields; perform factual proposition parsing on the original complaint materials; extract the factual elements corresponding to each original factual proposition; identify the factual limiting elements associated with the factual elements and determine the limiting objects of each factual limiting element; and construct the litigation element structure based on the factual elements, factual limiting elements, limiting objects, and the semantic relationships between the factual elements.

[0009] Obtain the semantic attributes of each target field, filter candidate litigation elements from the litigation element structure based on the semantic attributes of the fields, establish a field mapping relationship between the litigation elements and the target fields based on the factual behavior, semantic role and limiting state in the original factual proposition of the candidate litigation elements, generate candidate backfill content based on the field mapping relationship and perform automatic backfilling on the target fields to form a candidate complaint;

[0010] The backfilled content in the candidate complaint is parsed for factual propositions. The backfilled factual propositions are matched with the original factual propositions and a proposition-level mapping relationship is established. The limitation states of the corresponding factual limiting elements in the mutually mapped factual propositions are compared, and the changes in factual limiting states caused by automatic backfilling are identified.

[0011] When the change in the factual constraint state meets the preset factual change conditions, the abnormal backfill field or target text range is determined according to the changed factual constraint element and its constraint object. Field rollback or partial regeneration is performed on the abnormal backfill field or target text range. The corrected backfill content is then subjected to factual proposition parsing and constraint state verification again. Based on the verification results, the backfill content of the corresponding target field is determined and the complaint is output.

[0012] Furthermore, the factual elements include at least one core factual element among factual subject, factual act, and factual object, as well as at least one related element among time element, place element, monetary element, and quantity element;

[0013] For each original factual proposition, determine the semantic role of each factual element relative to the factual behavior, and save the factual proposition identifier, factual element and corresponding semantic role association to distinguish factual elements of the same element type in different original factual propositions.

[0014] Furthermore, the factual limiting elements include at least one of negative limiting, speculative limiting, quantitative limiting, and subject certainty limiting, wherein the speculative limiting is used to characterize the degree of cognitive certainty regarding factual behavior or factual relationship, and the subject certainty limiting is used to characterize the degree of certainty regarding the identity of the factual subject;

[0015] Determine the limiting object of each fact limiting element, wherein the limiting object includes the core fact element, the factual relationship between the core fact elements, or the cognitive state formed in response to the corresponding fact, and establish an association between the limiting state of the fact limiting element and the corresponding limiting object;

[0016] When standardizing the data format of fact elements, the limiting state associated with the fact elements is preserved.

[0017] Furthermore, the target fields are configured with field identifiers and field semantic attributes, and the field semantic attributes include at least the feature types that the target field is allowed to carry and their corresponding semantic roles;

[0018] When establishing the field mapping relationship, the litigation elements in the litigation element structure are screened as candidates according to the element types allowed to be carried by the target field, and matched according to the factual behavior in the original factual proposition where the candidate litigation elements are located and the semantic role relative to the factual behavior. When there are multiple candidate litigation elements of the same type, the candidate litigation elements corresponding to the target field are determined according to the factual behavior and semantic role.

[0019] Furthermore, when establishing the field mapping relationship, the limiting state corresponding to the candidate litigation element is associated with the target field, so that the field mapping relationship at least represents the correspondence between the target field, the original factual proposition, the candidate litigation element, the semantic role and the limiting state;

[0020] When a target field corresponds to multiple litigation elements, the multiple litigation elements are grouped into a field mapping group corresponding to the target field, and the corresponding candidate backfill content is generated according to the field mapping group.

[0021] Furthermore, for target fields that can be filled in a structured manner, the corresponding candidate litigation elements are standardized according to the data format of the target fields, while maintaining the limited state of the candidate litigation elements;

[0022] For a target field that needs to generate continuous text, a generation constraint is constructed based on the original factual proposition and litigation element structure corresponding to the target field. The generation constraint includes a factual relationship maintenance constraint for maintaining the factual subject, factual behavior, factual object and their semantic roles, and a limitation state maintenance constraint for maintaining the state of negation, speculation, quantity and subject certainty. Candidate backfill content is generated based on the generation constraint.

[0023] Furthermore, when performing automatic backfilling, the system determines whether the corresponding target field forms a unique mapping relationship based on the element type of the candidate litigation element, the factual behavior in the original factual proposition, the semantic role, and the limiting state.

[0024] When a unique mapping relationship is formed and the candidate backfill content meets the data type and semantic requirements of the target field, the candidate backfill content is written into the corresponding target field; when there are multiple candidate litigation elements, and a unique mapping relationship cannot be formed based on element type, factual behavior, semantic role and limiting status, each candidate litigation element and its field mapping relationship are retained, and the corresponding target field is set to the pending confirmation status.

[0025] Furthermore, when establishing the proposition-level mapping relationship, candidate proposition pairs are determined based on the factual behaviors in the original factual propositions and the backfilled factual propositions, and the corresponding original factual propositions and backfilled factual propositions are determined from the candidate proposition pairs based on the semantic roles of the factual subject, factual object, and related factual elements relative to the corresponding factual behaviors.

[0026] When one original fact proposition corresponds to multiple backfilled fact propositions, a one-to-many proposition-level mapping relationship is established; when multiple original fact propositions correspond to one backfilled fact proposition or a group of backfilled fact propositions, a many-to-one proposition-level mapping relationship is established.

[0027] Furthermore, the changes in the factual constraint state include constraint retention, constraint weakening, constraint strengthening, constraint deletion, and constraint addition;

[0028] The factual limiting elements and their limiting objects in the original factual propositions and backfilled factual propositions that are mapped to each other are compared. When the change in the factual limiting state causes the factual validity state, factual certainty degree or factual scope of the backfilled factual proposition to change relative to the original factual proposition, it is determined that the corresponding backfilled factual proposition satisfies the preset factual change condition.

[0029] Furthermore, when the backfilled factual proposition meets the preset factual change conditions, the corresponding target field is determined based on the changed factual limiting elements and their limiting objects; for continuous text fields, the target text range that can fully carry the corresponding factual proposition is determined based on the factual limiting elements, limiting objects, and the factual relationship between the two.

[0030] When the original content can meet the requirements of the target field and direct replacement does not affect the semantic integrity of adjacent content, field rollback is performed on the abnormal backfill field or the target text range; otherwise, local generation constraints are constructed based on the corresponding original factual proposition, and only the target text range is locally regenerated.

[0031] The corrected content is parsed again to establish a proposition-level mapping relationship between the corrected factual propositions and the corresponding original factual propositions, and a state-limited verification is performed. If the verification passes, the corrected content is retained; if the verification fails, the content is reverted to the original content or the corresponding target field is set to a pending confirmation state.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] This invention analyzes the factual propositions of the original indictment materials, extracting the factual subject, factual act, factual object, and related elements while identifying factual limiting elements and determining their limiting objects. This allows litigation elements to retain the original factual relationships and limiting states such as negation, speculation, quantity range, and subject certainty during the structuring process. By combining the semantic attributes of the target field with the factual act, semantic role, and limiting state in the factual proposition of the candidate litigation element, a field mapping relationship is established. This allows similar litigation elements to be matched with the corresponding target field based on their actual factual relationships, reducing the risk of element mismatch and the definitive backfilling of uncertain information. Furthermore, by establishing a proposition-level mapping relationship between the original factual proposition and the backfilled factual proposition, and comparing changes in the corresponding factual limiting state, the invention can identify changes in the factual establishment state, factual certainty, or factual scope caused during the automatic backfilling process. For abnormal backfilled fields or target text ranges, field rollback or partial regeneration and re-verification are performed, thus forming a closed-loop process from element extraction, field mapping, automatic backfilling to factual consistency verification and local correction, improving the accuracy of automatic complaint backfilling, factual semantic consistency, and reliability of the generated results. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the structured extraction and automatic backfilling method for elements of the present invention.

[0035] Figure 2 This is a schematic diagram illustrating the structure of the litigation elements and the factual limitations of this invention;

[0036] Figure 3 This is a schematic diagram illustrating the mapping and automatic backfilling of litigation elements and target fields in this invention;

[0037] Figure 4 This is a schematic diagram illustrating the mapping between original factual propositions and backfilled factual propositions, as well as the identification of changes in limited states, in this invention.

[0038] Figure 5 This is a schematic diagram illustrating the local correction and re-verification of abnormal backfill content in this invention. Detailed Implementation

[0039] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only used to illustrate the technical principles and specific implementation processes of the present invention and are not intended to limit the scope of protection of the present invention. Those skilled in the art can make adaptive adjustments to the specific implementation methods according to the type of prosecution materials, case type, complaint template type, target field setting method, and natural language processing model type without departing from the technical concept of the present invention.

[0040] This invention addresses the issue that during the automatic generation of complaints, the original complaint materials are usually in an unstructured or semi-structured form. The litigation elements, such as parties, factual acts, factual objects, time, amount, and quantity, are scattered across different text fragments. Furthermore, the corresponding litigation elements may be limited by information such as negation, speculation, quantity range, and the degree of certainty of the subject. This makes it easy to convert litigation elements with uncertain states into certain content or to fill in the wrong fields when information is extracted and templates are filled in based solely on entity type, keywords, or field names.

[0041] To this end, this invention first performs factual proposition analysis on the original complaint materials, using factual propositions as the organizational unit of litigation elements, extracting the factual subject, factual behavior, factual object, and related elements, while identifying factual limiting elements and their actual limiting objects, thereby forming a litigation element structure that retains factual relationships and limiting states. Then, based on the semantic attributes of the target fields in the complaint template, candidate litigation elements are screened from the litigation element structure, and a field mapping relationship is established by combining factual behavior, semantic roles, and limiting states, thereby generating candidate backfill content and performing automatic backfilling. For the candidate complaints generated by automatic backfilling, factual proposition analysis is performed again, matching the obtained backfilled factual propositions with the original factual propositions and establishing a proposition-level mapping relationship. By comparing the state changes of the corresponding factual limiting elements, it is determined whether the automatic backfilling has changed the factual establishment state, degree of factual certainty, or factual scope expressed by the original materials. For backfill content with abnormal changes, the corresponding abnormal backfilled field or target text range is located, and field rollback or partial regeneration is performed, followed by factual proposition analysis and limiting state verification, thus forming a closed-loop processing process for automatic complaint backfilling.

[0042] Example 1

[0043] like Figure 1 As shown, this embodiment provides a method for structured extraction and automatic backfilling of elements for indictments, including the following steps:

[0044] S101. Obtain the original complaint materials corresponding to the case to be processed, and obtain the target complaint template corresponding to the case to be processed. The original complaint materials may include at least one of the following: party statements, identity documents, evidence, statements of claims, and other relevant textual materials.

[0045] For materials provided in electronic text form, the corresponding text data can be obtained directly; for materials provided in the form of images or scanned documents, text recognition can be performed first to obtain the corresponding text content; when the original materials come from different documents, material identifiers and source location identifiers can be configured for different documents and their text fragments to preserve the source correspondence between the subsequently extracted litigation elements and the original materials.

[0046] The target complaint template includes multiple target fields, each with its own field identifier and semantic attributes.

[0047] The semantic attributes of the fields are used to describe the element type and semantic role of the information carried by the corresponding target field. For example, the target field may include the plaintiff's name, the defendant's name, the amount involved, the payment time, the number of payments, the payment method, and the facts and reasons, etc.

[0048] For structured target fields, the corresponding data type can be further set; for target fields that need to form continuous text, their corresponding fact description type can be set.

[0049] This enables subsequent automatic backfilling to not simply perform text replacement based on field names, but to determine the backfill content based on the feature type required by the field and its semantic role in the corresponding factual relationship.

[0050] S102. After obtaining the original indictment materials, perform factual proposition analysis on the text content, and determine the text content that can express a relatively complete factual relationship as the original factual propositions. For example... Figure 2 As shown, for each original factual proposition, the core factual element and other elements associated with the core factual element are extracted. The core factual element includes at least one of the following: factual subject, factual action, and factual object.

[0051] The subject of a fact is used to represent the subject that implements or participates in the corresponding factual act; the factual act is used to represent the behavior, state, or factual relationship in the corresponding factual proposition; the object of a fact is used to represent the person, item, money, or other object to which the factual act is directed.

[0052] The associated elements may include time, location, amount, quantity, and other elements related to the factual action. For each original factual proposition, the semantic role of each extracted factual element relative to the factual action is further determined, and the factual proposition identifier, factual element, and corresponding semantic role are associated and saved. The semantic role is used to characterize the role relationship of the corresponding factual element in the corresponding factual action, so that factual elements of the same type can be distinguished according to their respective factual propositions and semantic roles.

[0053] For example, the sentence "The defendant may have paid the plaintiff approximately 30,000 yuan around March 2026" can be parsed to obtain the subject of the fact "defendant," the act of the fact "payment," the object of the fact or the recipient "plaintiff," the time element "around March 2026," and the amount element "approximately 30,000 yuan." Here, the semantic role of "defendant" relative to the act of "payment" is that of the paying party, the semantic role of "plaintiff" is that of the receiving party, and the semantic role of "approximately 30,000 yuan" is that of the payment amount.

[0054] When multiple factual acts exist in the original complaint materials, corresponding original factual propositions are formed based on each factual act and its related subjects, objects, and other elements, avoiding the indiscriminate merging of similar elements throughout the text simply according to entity type. For example, if the materials simultaneously contain "contractually agreed amount," "actually paid amount," and "outstanding amount," even if all three belong to the monetary type, their corresponding factual propositions and factual relationships are retained separately. Furthermore, the semantic roles of the three monetary elements relative to their respective factual acts are preserved for subsequent field mapping.

[0055] S103. Identify the factual limiting elements in each original factual proposition. These elements include at least one of the following: negation limiting, conjecture limiting, quantity limiting, and subject-specific limiting. Wherein:

[0056] Negation is used to indicate that the corresponding factual action did not occur, the factual relationship is not established, or the corresponding fact is negated.

[0057] The term "speculation" is used to characterize the degree of certainty in one's cognition regarding a factual action or relationship. For example, phrases like "may happen," "probably happen," and "it seems that payment has already been made" express the state of speculation.

[0058] Quantity limits are used to indicate that numerical information such as amount, quantity, and frequency belongs to a definite value, a range of values, or a fuzzy value.

[0059] Subject certainty is used to characterize the degree of certainty of the identity of a factual subject, including whether the factual subject is a certain subject, a candidate subject, or a subject that cannot be determined.

[0060] For the identified factual limiting elements, not only are the limiting elements themselves recorded, but the corresponding limiting objects are further determined. The limiting objects include the core factual elements, the factual relationships between the core factual elements, or the cognitive state formed in response to the corresponding facts. For example:

[0061] Regarding the statement "the defendant did not pay the second installment," the word "did not" refers to the factual act of "paying the second installment."

[0062] Regarding the statement "it cannot be determined whether the payment was made by the defendant himself," "cannot be determined" refers to the state of understanding of the identity correspondence between the payer and "the defendant himself," and does not mean that the payment itself did not occur.

[0063] The word "approximately 30,000 yuan" limits the monetary element "30,000 yuan" to indicate that the amount is a vague quantity rather than a precise amount that can be directly determined from the original material.

[0064] Regarding the statement "the defendant may have paid the corresponding amount to the plaintiff," the word "may" refers to the degree of certainty in the understanding of the corresponding payment fact, and is thus defined as a speculative limitation, limiting the cognitive state corresponding to the payment fact. Conversely, regarding the statement "it is possible that the defendant paid the corresponding amount to the plaintiff," the word "may" applies to the identity of the paying entity, and is thus defined as a subject-specific limitation, limiting the cognitive state of the identity correspondence between the paying entity and the "defendant." This distinguishes between speculative states regarding factual acts or relationships and definitive states regarding the identity of the factual subjects.

[0065] Therefore, by defining the limiting object, the limiting information is kept related to the factual elements or factual relationships that actually play a role, thus reducing the change in the original factual semantics caused by the mechanical binding of limiting words to adjacent entities during subsequent structured processing.

[0066] S104. Based on the aforementioned results of factual proposition analysis, core elements of facts, related elements, limiting elements of facts and their limiting objects, construct the corresponding litigation element structure. The litigation element structure shall at least include factual proposition identifiers, factual elements, limiting elements of facts, limiting objects, and semantic relationships between different factual elements.

[0067] In one implementation, a structured record is established for each factual proposition, and elements such as the corresponding factual subject, factual action, factual object, time, place, amount, and quantity are used as element nodes of the structured record. The factual limiting elements are associated with their corresponding limiting objects. Each factual element stores its semantic role relative to the factual action, and each factual limiting element stores its limiting state and limiting object, thus enabling each structured record to represent the factual elements, semantic roles, and limiting relationships in the corresponding original factual proposition.

[0068] For example, a structured record for the statement "The defendant may pay the plaintiff approximately 30,000 yuan around March 2026" could include:

[0069] The subject of the facts is the "defendant";

[0070] The act in question is "payment";

[0071] The recipient is the "plaintiff";

[0072] The time element is "March 2026", and its limitation status is range limitation;

[0073] The monetary element is "thirty thousand yuan", and its limitation state is an fuzzy quantity;

[0074] "Possible" corresponds to the speculative limitation, and its object of limitation is the cognitive state of the corresponding payment fact.

[0075] When the original expression is "it is possible that the defendant paid the plaintiff approximately 30,000 yuan", the "possible" in the context of the payment entity's identity is a certainty limitation, and the limitation applies to the cognitive state of the identity correspondence between the payment entity and the "defendant".

[0076] It should be noted that when standardizing the format of elements such as amount and time, the corresponding limiting states are retained. For example, "thirty thousand yuan" can be converted into a uniform numerical representation, but the word "approximately" in the original expression is still preserved through the limiting state identifier, and the ambiguous amount is not converted into a definite amount due to data format standardization.

[0077] When the same entity assumes different semantic roles in multiple factual propositions, the factual proposition in which it is located and its corresponding semantic role are retained respectively.

[0078] S105, such as Figure 3 As shown, each target field and its corresponding semantic attribute in the target complaint template are obtained, and candidate litigation elements are filtered from the litigation element structure based on the field semantic attributes.

[0079] First, candidate elements are filtered based on the element type corresponding to the target field. For example, the "amount involved" field prioritizes filtering litigation elements of the amount involved, while the "payment time" field filters litigation elements of the time involved. When multiple candidate litigation elements of the same type exist, they are further matched based on the factual behavior and semantic role in the factual proposition in which the candidate litigation elements are located.

[0080] For example, when the original complaint materials include the total contract amount, the amount already paid, and the amount not yet paid, for the target field of "amount already paid", instead of directly selecting any one of the above elements as they are all amount types, we further determine the "payment" factual act corresponding to the target field and determine the candidate litigation elements from the amount elements associated with the payment act.

[0081] Furthermore, matching is performed based on the limiting status corresponding to the candidate litigation elements, and the limiting status is retained as part of the field mapping relationship.

[0082] Specifically, when establishing field mapping relationships, the candidate range is first narrowed down according to the element types that the target field is allowed to carry. Then, the semantic correspondence is determined according to the factual behavior in the original factual proposition where the candidate litigation element is located and the semantic role relative to the factual behavior. Finally, the limited state of the candidate litigation element is associated with the target field.

[0083] Through the above processing, the field mapping relationship can at least represent the correspondence between: target field—original factual proposition—candidate litigation element—semantic role—limited state. When a target field requires multiple litigation elements to form a co-occurrence, the multiple litigation elements are established as a field mapping group for the corresponding field. For example, the "facts and reasons" field can simultaneously map the factual subject, factual act, factual object, time, amount, and corresponding limited state, and generate corresponding text content based on multiple litigation elements.

[0084] S106. After establishing the field mapping relationship, generate candidate backfill content according to the type of the target field, such as... Figure 3 As shown, for target fields such as name, date, amount, and number of times that can be filled in a structured manner, the corresponding litigation elements are standardized according to the data format required by the target fields.

[0085] Standardization only changes the data representation format and does not remove the already recorded constraints of the corresponding litigation elements. When the constraints cannot be fully expressed through the data format of the target field, or when the original materials cannot uniquely determine the content of the corresponding field, no artificial increase in factual certainty is made through format conversion. For continuous text fields such as "Facts and Reasons," constraints are generated based on the original factual propositions and litigation element structure mapped to the field, and candidate backfill text is generated using a natural language processing model. The generated constraints include at least factual relationship preservation constraints and constraint preservation of constraints.

[0086] The factual relation preservation constraint is used to ensure that the factual subjects, factual actions, factual objects and their semantic roles in the candidate backfill text remain consistent with the original factual proposition; the limited state preservation constraint is used to ensure that the candidate backfill text retains the negation, speculation, quantity range and subject certainty in the original factual proposition.

[0087] For example, if the original factual proposition is: "The defendant may pay the plaintiff approximately 30,000 yuan," then the candidate backfill text cannot be generated simply by language standardization as: "The defendant paid the plaintiff 30,000 yuan." This is because the latter expression removes the speculative state regarding the payment fact and transforms the amount from a vague quantity to a definite quantity, thereby increasing the certainty of the corresponding fact.

[0088] S107. For target fields that have already formed candidate backfill content, determine whether the corresponding target fields form a unique mapping relationship based on the element type of the candidate litigation element, the factual behavior in the original factual proposition, the semantic role, and the limiting state.

[0089] When the target field has a unique candidate litigation element or a unique field mapping group, and the candidate backfill content meets the data type and semantic requirements of the corresponding field, the candidate backfill content is written to the corresponding target field.

[0090] When multiple candidate litigation elements exist, and the target litigation element cannot be uniquely determined based on element type, factual behavior, semantic role, and limited status, one candidate element is not forced to be selected. Instead, the candidate litigation elements and their field mapping relationships are retained, and the corresponding target field is set to the pending confirmation status.

[0091] For multiple target fields that can be automatically filled, candidate content is written according to the field positions in the target complaint template to form a candidate complaint. The automatic filling process does not simply perform a direct "variable-text" replacement, but uses the aforementioned litigation element structure and field mapping relationship as the basis for filling.

[0092] S108. After automatic backfilling is completed, as follows: Figure 4 As shown, the backfilled content in the candidate complaint is analyzed again to form one or more backfilled factual propositions. For each backfilled factual proposition, the core elements, related elements, and limiting elements of the backfilled fact are extracted, and the corresponding limiting objects are determined to form the backfilled fact structure. Then, the obtained backfilled factual propositions are matched with the original factual propositions to establish a proposition-level mapping relationship.

[0093] Specifically, candidate proposition pairs are first determined based on the factual actions in the original factual propositions and the backfilled factual propositions; then, based on the semantic roles of the factual subject, factual object and other related elements in the corresponding factual actions, the corresponding original factual propositions and backfilled factual propositions are determined from the candidate proposition pairs.

[0094] When an original factual proposition is split into multiple target fields during the automatic backfilling process, a one-to-many propositional mapping relationship can be established between the original factual proposition and the multiple backfilled factual propositions.

[0095] When multiple original factual propositions are merged to form a continuous text such as "facts and reasons", a many-to-one propositional mapping relationship can be established between multiple original factual propositions and a backfilled factual proposition or a backfilled proposition group.

[0096] Through the proposition-level mapping relationship, subsequent state comparisons are performed between factual propositions that have corresponding factual behaviors and semantic roles, avoiding the need to judge whether the semantics of facts are consistent solely based on the similarity of strings before and after backfilling or the same independent entities.

[0097] S109, such as Figure 4 As shown, for the original factual propositions and the backfilled factual propositions that establish propositional-level mapping relationships, the corresponding factual limiting elements and their limiting objects are compared. Based on the comparison results, the type of limiting state change is determined. The limiting state change types include limiting retention, limiting weakening, limiting strengthening, limiting deletion, and limiting addition. Among them:

[0098] The constraint "maintain" indicates that the corresponding constraint state before and after backfilling did not change in terms of the semantics of the facts.

[0099] Restricted reinforcement indicates that the backfilled factual proposition has a higher degree of certainty or a narrower scope of facts compared to the original factual proposition;

[0100] The weakening of the constraint indicates that the backfilled factual proposition reduces the certainty of the original fact or expands its uncertainty.

[0101] Limited deletion indicates that the limiting information present in the original factual proposition is not retained in the backfilled factual proposition;

[0102] The "limited addition" option indicates a situation where the original factual proposition does not exist in the backfilled factual proposition.

[0103] For example, such as Figure 4 As shown, when the amount in the original material is "approximately 30,000 yuan" and the backfill content is "30,000 yuan", the original fuzzy quantity state is transformed into a definite quantity state, which can be identified as a limited reinforcement.

[0104] When the original material states that "the specific amount of the breach needs to be further verified", but the backfilled content directly forms a definite amount of the breach or the corresponding limiting information completely disappears, it can be identified as a strengthening or deletion of the limitation.

[0105] Further determine whether the corresponding change in the limiting state leads to a change in the factual validity, degree of factual certainty, or scope of the factual proposition to be filled in.

[0106] When a change occurs, determine whether the corresponding backfilled factual proposition meets the preset factual change conditions; otherwise, exclude cases where only the form of expression has changed without altering the semantics of the fact.

[0107] For example, if the original material states "the defendant has not paid the remaining payment", and the automatic backfilling results in "the defendant has not yet paid the remaining payment", although the textual expression has changed, the negation limitation and the payment fact it limits remain consistent. Therefore, it is determined to be a limitation that is maintained and does not meet the presupposed condition of factual change.

[0108] S110. When at least one backfilled factual proposition is detected to meet the preset factual change conditions, such as Figure 5 As shown, the corresponding target field is determined in reverse based on the changed factual elements and their defined objects; for structured fields, if the changed litigation elements correspond to only one target field that can be modified independently, then that target field is determined as an abnormal backfill field.

[0109] For continuous text fields, the target text range is determined based on the position of the changed factual defining elements and their defined objects within the backfilled text. The target text range is a range of text capable of fully encompassing the changed factual defining elements, their defined objects, and the corresponding factual relationship between them.

[0110] Preferably, a minimum semantically complete text range that meets the above conditions is determined to reduce the impact on other backfilled content that has already passed the verification during the correction process.

[0111] When the target text ranges corresponding to multiple abnormal limiting elements overlap or have direct semantic relationships that cannot be processed independently, the corresponding target text ranges are merged and then processed.

[0112] S111. Based on the content type of the abnormal backfill field or the target text range, determine the corresponding correction method. First, determine whether the original content corresponding to the original complaint material can be directly used to perform the rollback. When the original litigation element corresponding to the abnormal backfill field can meet the data type and semantic requirements of the target field, and direct replacement will not destroy the semantic integrity of the adjacent content, execute the field rollback.

[0113] For example, when "approximately 30,000 yuan" is converted to a specific amount "30,000 yuan" during the automatic backfilling process, and the corresponding target field allows for the retention of ambiguous monetary expressions, the corresponding content can be rolled back to "approximately 30,000 yuan" as supported by the original material.

[0114] When the target text is formed by multiple litigation elements, and directly reverting to the original text would result in syntactic incompleteness or abnormal contextual coherence, a partial regeneration of the target text is performed. During partial regeneration, the local generation constraints are reconstructed based on the corresponding original factual propositions.

[0115] Locally generated constraints include at least:

[0116] Preserve the original factual subject, factual act, factual object, and their semantic roles;

[0117] Maintain the original state of denial, speculation, quantity, and subject definiteness;

[0118] The regeneration will be limited to the target text.

[0119] Maintain semantic coherence between the target text range and adjacent content that has already passed validation.

[0120] Under the above constraints, only the target text range is regenerated, while the backfilled content outside the target text range that has passed the factual consistency check remains unchanged. Compared to regenerating the entire complaint, this process can focus the corrective effect on the local content where the facts have changed.

[0121] S112. Perform fact proposition parsing again on the candidate corrected content after field rollback or partial regeneration. In accordance with the method described in S108, establish a proposition-level mapping relationship between the corrected fact proposition and the corresponding original fact proposition, and compare the corresponding fact constraint states in accordance with the method described in S109.

[0122] Specifically, the factual proposition after correction is compared again with the factual limiting elements and their limiting objects in the corresponding original factual proposition, and it is determined whether the factual proposition after correction still changes the factual validity, degree of factual certainty, or factual scope of the original factual proposition.

[0123] When the corrected factual proposition no longer meets the preset factual change conditions, the corresponding candidate correction content is determined to pass the factual consistency check and written into the corresponding target field.

[0124] If the regenerated content still meets the preset fact change conditions, the regenerated content will not be used.

[0125] At this point, for target fields where the original content can be directly used, the process reverts to the content supported by the original complaint materials; for cases where the original materials themselves have multiple candidate elements or cannot uniquely determine the content of the target field, the corresponding target field is set to a pending confirmation state, while retaining the corresponding candidate litigation elements and source information. This avoids generating definitive facts that the original materials cannot support in order to achieve a complete formality in the complaint.

[0126] S113. For target fields that have passed the factual consistency check, retain the corresponding backfilled content in the target complaint template; for target fields that require manual confirmation, keep the corresponding pending confirmation status, combine the target fields according to the field positions and document structure in the target complaint template, form the final complaint and output it.

[0127] Example 2

[0128] The following is combined Figure 2 and Figure 3 The document further explains the construction of the litigation elements structure and the automatic backfilling process; it assumes that the original complaint materials include the party's statement, identity documents, evidence materials, and a statement of the litigation claims. Among them, the party's statement states: "Around March 2026, the other party transferred money to me in two installments. I remember the total was about 30,000 yuan, but I cannot be completely sure whether one of the payments was transferred from the defendant's own account."

[0129] The system first parses the text into raw factual propositions related to the payment transaction; from these raw factual propositions, the following can be extracted:

[0130] The actual act is "transferring or making payment";

[0131] The recipient is the plaintiff;

[0132] The time element is "around March 2026";

[0133] The frequency element is "twice";

[0134] The amount is listed as "approximately 30,000 yuan";

[0135] It is uncertain whether the payer for one of the payments is the defendant himself.

[0136] Further identification revealed:

[0137] "Left and right" correspond to a time range limitation;

[0138] "Approximately" corresponds to a vague quantitative limit on the amount;

[0139] "Cannot be completely determined" corresponds to the subject's certainty in the identity relationship of the payment entity, and its object of limitation is the cognitive state of the identity correspondence between the corresponding payment entity and the defendant.

[0140] Therefore, when constructing the litigation element structure, it is not a simple set of fields such as "Payment time = March 2026; Number of payments = 2; Payment amount = 30,000; Payment subject = Defendant", but rather the factual proposition, semantic role and limiting state of the corresponding elements are further preserved.

[0141] like Figure 3 As shown, when the target complaint template includes target fields such as "amount involved", "payment time", "number of payments", "payment method" and "facts and reasons", candidate litigation elements are first filtered according to the semantic attributes of the fields.

[0142] For "Payment Time", select the time element associated with the actual payment transaction: "around March 2026".

[0143] For "Number of payments", select the quantity element "twice" that is associated with the actual payment transaction;

[0144] Regarding the "amount involved," the element of "approximately 30,000 yuan" is selected as the amount associated with the actual payment, while retaining its vague quantitative limitation;

[0145] For “Facts and Reasons”, a field mapping group is constructed based on the multiple litigation elements corresponding to the payment fact proposition and their limiting status.

[0146] If the original materials also record a total contract price of 50,000 yuan, then although both "50,000 yuan" and "30,000 yuan" are monetary elements, the former corresponds to the contract price relationship, while the latter corresponds to the actual payment. For the "Amount Paid" field, after further filtering based on factual behavior and semantic roles, "30,000 yuan," which is associated with the actual payment, should be selected, rather than simply matching based on the fact that both are monetary types.

[0147] Example 3

[0148] The following is combined Figure 4 The process of verifying the consistency of facts is explained below. Assume that the original complaint stated: "Between June and August 2023, the defendant paid the plaintiff approximately RMB 30,000 in two installments, but some payment for goods is still outstanding. The specific amount of the default needs to be further verified based on the account statement."

[0149] After automatic backfilling, the "Facts and Reasons" section reads: "Between June and August 2023, the defendant paid the plaintiff 30,000 yuan in two installments, and some of the payment is still outstanding."

[0150] The system provides factual proposition analysis for both the original materials and the backfilled content:

[0151] For payment facts, firstly, candidate proposition pairs are determined based on the factual act of "payment". Then, based on the semantic roles of the payer, receiver, time, frequency and amount relative to the payment act, the propositional mapping relationship between the original payment fact propositions and the backfilled payment fact propositions is determined.

[0152] The monetary element in the original payment proposition is "30,000 yuan", and its limiting state is "approximately"; the monetary element in the backfilled payment proposition is "30,000 yuan", and there is no corresponding fuzzy quantity limitation.

[0153] Therefore, it can be determined that the state corresponding to this amount has changed from an ambiguous state to a definite state, which is a form of constraint reinforcement.

[0154] For the original proposition that "the specific amount of default needs to be further verified based on the statement of account", if the automatic backfilling process directly generates a definite amount of default or completely omits the corresponding uncertain state, it can be further identified as a limited reinforcement or limited deletion based on the proposition-level mapping result.

[0155] When the aforementioned changes cause the certainty of the backfilled facts to exceed the certainty that the original indictment materials can support, the corresponding backfilled fact proposition is determined to meet the preset fact change conditions.

[0156] Conversely, if the original material states "the defendant has not paid the remaining payment," and the automatic backfill changes it to "the defendant has not yet paid the remaining payment," although the wording changes, the negative status and the payment behavior it specifies remain consistent. Therefore, it is determined to be a limitation and is not treated as an abnormal backfill.

[0157] Example 4

[0158] The following is combined Figure 5 This section explains the corrective action taken after an anomaly is filled in. For example, the phrase "approximately 30,000 yuan" was automatically filled in as "30,000 yuan". Once the determination of the amount limit status is strengthened, the corresponding target field or the target text range in "Facts and Reasons" is determined in reverse based on the field mapping relationship and the changed factual limiting elements and objects.

[0159] If the amount is in the "Amount Involved" field that can be processed independently, and the field allows the original ambiguous expression to be retained, then the field rollback is performed directly to restore "30,000 yuan" to "approximately 30,000 yuan" that the original material can support.

[0160] If the amount is located in the continuously generated "facts and reasons" text, and direct replacement may affect the contextual expression, then determine the minimum semantically complete text range that includes the corresponding amount element, payment behavior, and necessary related subjects.

[0161] Construct local generation constraints based on the original payment fact proposition to maintain the following during the regeneration process:

[0162] The payer and the recipient remain unchanged;

[0163] The factual relationship of payment remains unchanged;

[0164] The amount remains ambiguous.

[0165] Other established factual elements, such as time and frequency, remain unchanged.

[0166] For example, it can be regenerated:

[0167] "Between June and August 2023, the defendant paid the plaintiff approximately 30,000 yuan in two installments."

[0168] After regeneration, the payment fact propositions are parsed again, and a mapping and state comparison are performed with the original payment fact propositions.

[0169] When the payment subject, payment behavior, receiving subject, time range, number of times, and amount limit are all within the factual semantic range that the original material can support, the regenerated result is determined to pass the verification and written back to the corresponding location.

[0170] If the regenerated result again produces a definite amount or a new fact arises that the original material does not exist, the regenerated result will not be used. Instead, the original content will be reverted or the corresponding target field will be set to a pending confirmation status.

[0171] Through the above implementation methods, this invention does not merely complete the one-way information extraction and template filling from the indictment materials to the complaint fields, but uses factual propositions and their litigation element structure as a unified semantic benchmark for pre- and post-processing: before automatic backfilling, factual behavior, semantic roles, and limiting states are used to constrain the mapping between litigation elements and target fields; after automatic backfilling, the proposition-level mapping between the original factual propositions and the backfilled factual propositions is used to detect changes in factual limiting states, and the detection results are used in reverse to locally correct abnormal backfilled content. Thus, the factual propositions, semantic roles, and limiting states used in the field mapping stage can continue to serve as the comparison basis for factual consistency verification after backfilling, and the anomalies found in the verification can be reversed to specific target fields or target text ranges, forming a closed-loop processing relationship.

[0172] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for structured extraction and automatic backfilling of elements for indictments, characterized in that, Includes the following steps: Obtain the original complaint materials and the target complaint template including multiple target fields; perform factual proposition parsing on the original complaint materials; extract the factual elements corresponding to each original factual proposition; identify the factual limiting elements associated with the factual elements and determine the limiting objects of each factual limiting element; and construct the litigation element structure based on the factual elements, factual limiting elements, limiting objects, and the semantic relationships between the factual elements. Obtain the semantic attributes of each target field, filter candidate litigation elements from the litigation element structure based on the semantic attributes of the fields, establish a field mapping relationship between the litigation elements and the target fields based on the factual behavior, semantic role and limiting state in the original factual proposition of the candidate litigation elements, generate candidate backfill content based on the field mapping relationship and perform automatic backfilling on the target fields to form a candidate complaint; The backfilled content in the candidate complaint is parsed for factual propositions. The backfilled factual propositions are matched with the original factual propositions and a proposition-level mapping relationship is established. The limitation states of the corresponding factual limiting elements in the mutually mapped factual propositions are compared, and the changes in factual limiting states caused by automatic backfilling are identified. When the change in the factual constraint state meets the preset factual change conditions, the abnormal backfill field or target text range is determined according to the changed factual constraint element and its constraint object. Field rollback or partial regeneration is performed on the abnormal backfill field or target text range. The corrected backfill content is then subjected to factual proposition parsing and constraint state verification again. Based on the verification results, the backfill content of the corresponding target field is determined and the complaint is output.

2. The method for structured extraction and automatic backfilling of elements for indictments according to claim 1, characterized in that, The factual elements include at least one core factual element among factual subject, factual act, and factual object, as well as at least one related element among time element, place element, monetary element, and quantity element; For each original factual proposition, determine the semantic role of each factual element relative to the factual behavior, and save the factual proposition identifier, factual element and corresponding semantic role association to distinguish factual elements of the same element type in different original factual propositions.

3. The method for structured extraction and automatic backfilling of elements for indictments according to claim 1, characterized in that, The factual limiting elements include at least one of negative limiting, speculative limiting, quantitative limiting, and subject certainty limiting, wherein the speculative limiting is used to characterize the degree of cognitive certainty regarding factual behavior or factual relationship, and the subject certainty limiting is used to characterize the degree of certainty regarding the identity of the factual subject; Determine the limiting object of each fact limiting element, wherein the limiting object includes the core fact element, the factual relationship between the core fact elements, or the cognitive state formed in response to the corresponding fact, and establish an association between the limiting state of the fact limiting element and the corresponding limiting object; When standardizing the data format of fact elements, the limiting state associated with the fact elements is preserved.

4. The method for structured extraction and automatic backfilling of elements for indictments according to claim 1, characterized in that, The target fields are configured with field identifiers and field semantic attributes, and the field semantic attributes include at least the feature types that the target field is allowed to carry and their corresponding semantic roles; When establishing the field mapping relationship, the litigation elements in the litigation element structure are screened as candidates according to the element types allowed to be carried by the target field, and matched according to the factual behavior in the original factual proposition where the candidate litigation elements are located and the semantic role relative to the factual behavior. When there are multiple candidate litigation elements of the same type, the candidate litigation elements corresponding to the target field are determined according to the factual behavior and semantic role.

5. The method for structured extraction and automatic backfilling of elements for indictments according to claim 4, characterized in that, When establishing the field mapping relationship, the limiting state corresponding to the candidate litigation element is associated with the target field, so that the field mapping relationship at least represents the correspondence between the target field, the original factual proposition, the candidate litigation element, the semantic role and the limiting state. When a target field corresponds to multiple litigation elements, the multiple litigation elements are grouped into a field mapping group corresponding to the target field, and the corresponding candidate backfill content is generated according to the field mapping group.

6. The method for structured extraction and automatic backfilling of elements for indictments according to claim 1, characterized in that, For target fields that can be filled in a structured manner, the corresponding candidate litigation elements are standardized according to the data format of the target fields, while maintaining the limited state of the candidate litigation elements; For a target field that needs to generate continuous text, a generation constraint is constructed based on the original factual proposition and litigation element structure corresponding to the target field. The generation constraint includes a factual relationship maintenance constraint for maintaining the factual subject, factual behavior, factual object and their semantic roles, and a limitation state maintenance constraint for maintaining the state of negation, speculation, quantity and subject certainty. Candidate backfill content is generated based on the generation constraint.

7. The method for structured extraction and automatic backfilling of elements for indictments according to claim 1, characterized in that, When performing automatic backfilling, the system determines whether the corresponding target field forms a unique mapping relationship based on the element type of the candidate litigation element, the factual behavior in the original factual proposition, the semantic role, and the limiting state. When a unique mapping relationship is formed and the candidate backfill content meets the data type and semantic requirements of the target field, the candidate backfill content is written into the corresponding target field; when there are multiple candidate litigation elements, and a unique mapping relationship cannot be formed based on element type, factual behavior, semantic role and limiting status, each candidate litigation element and its field mapping relationship are retained, and the corresponding target field is set to the pending confirmation status.

8. The method for structured extraction and automatic backfilling of elements for indictments according to claim 1, characterized in that, When establishing the proposition-level mapping relationship, candidate proposition pairs are determined based on the factual behaviors in the original factual propositions and the backfilled factual propositions. Based on the semantic roles of the factual subject, factual object, and related factual elements relative to the corresponding factual behaviors, the corresponding original factual propositions and backfilled factual propositions are determined from the candidate proposition pairs. When one original fact proposition corresponds to multiple backfilled fact propositions, a one-to-many proposition-level mapping relationship is established; when multiple original fact propositions correspond to one backfilled fact proposition or a group of backfilled fact propositions, a many-to-one proposition-level mapping relationship is established.

9. The method for structured extraction and automatic backfilling of elements for a complaint according to claim 1, characterized in that, The aforementioned factual state changes include constraint retention, constraint weakening, constraint strengthening, constraint deletion, and constraint addition; The factual limiting elements and their limiting objects in the original factual propositions and backfilled factual propositions that are mapped to each other are compared. When the change in the factual limiting state causes the factual validity state, factual certainty degree or factual scope of the backfilled factual proposition to change relative to the original factual proposition, it is determined that the corresponding backfilled factual proposition satisfies the preset factual change condition.

10. The method for structured extraction and automatic backfilling of elements for a complaint according to claim 1, characterized in that, When the backfilled factual proposition meets the preset factual change conditions, the corresponding target field is determined according to the changed factual limiting elements and their limiting objects; for continuous text fields, the target text range that can fully carry the corresponding factual proposition is determined according to the factual limiting elements, limiting objects and the factual relationship between the two. When the original content can meet the requirements of the target field and direct replacement does not affect the semantic integrity of adjacent content, field rollback is performed on the abnormal backfill field or the target text range; otherwise, local generation constraints are constructed based on the corresponding original factual proposition, and only the target text range is locally regenerated. The corrected content is parsed again to establish a proposition-level mapping relationship between the corrected factual propositions and the corresponding original factual propositions, and a state-limited verification is performed. If the verification passes, the corrected content is retained; if the verification fails, the content is reverted to the original content or the corresponding target field is set to a pending confirmation state.

Images