A method and system for generating intellectual property legal facts

Abstract

The application discloses a kind of generation method and system of intellectual property legal fact.The method includes: intellectual property evidence data acquisition step, according to the oral statement of voice recognition subsystem transcription, according to the written material scanned by text recognition subsystem, generate initial intellectual property evidence data set with time stamp and fixed key set;Intellectual property key word extraction step, based on the keyword list in the index library of multi-source legal corpus and synonym normalization rules, extract the candidate intellectual property key word containing subject, behavior, time, place, causality and right and obligation from the initial intellectual property evidence data set;Intellectual property legal fact generation step, based on timeline reconstruction and role consistency check, the candidate intellectual property key word is formed into intellectual property legal fact, while the corresponding relationship between the intellectual property legal fact and the initial intellectual property evidence data set is increased for retrieval and review.The method can improve the speed and accuracy of intellectual property evidence sorting, reduce irrelevant and ambiguous expression residues, and ensure that intellectual property legal fact and original evidence are one-to-one corresponding through traceable mapping, and enhance reviewability.

Description

Technical Field

[0001] This invention relates to the field of law, and more specifically, to a method and system for generating intellectual property legal facts. Background Technology

[0002] With the digitalization and networking of the intellectual property judicial environment, the coexistence of oral statements and written materials in intellectual property cases such as patents, trademarks, and copyrights is becoming increasingly common. In many intellectual property legal fact-finding processes, the acquisition, organization, and transformation of evidence such as patent files, licensing agreements, and infringement statements still heavily rely on manual processing, resulting in low efficiency and high risk. While traditional text recognition and scanning technologies can initially obtain patent documents, contract texts, or witness statements, they are insufficient for in-depth extraction and structuring of legal elements in intellectual property cases, such as "rights holder—infringing act—licensing transfer—infringement result."

[0003] Many existing intelligent assistance systems rely on rule matching or a single text data source, making it difficult to handle the complex chain of facts in intellectual property cases that involves multiple parties (rights holders, licensees, and infringers) and involves multiple voice statements, patent / contract / appraisal report documents.

[0004] Therefore, the intellectual property legal services industry urgently needs a method that can unify multiple sources of materials, such as voice statements, contract texts, and technical appraisal reports, into verifiable intellectual property legal facts. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention aims to provide a method and system for generating intellectual property legal facts.

[0006] According to one aspect of the present invention, a method for generating intellectual property legal facts is provided. The method includes: an intellectual property evidence data acquisition step, generating an initial intellectual property evidence dataset with timestamps and a fixed key set based on oral statements transcribed by a speech recognition subsystem and written materials scanned by a text recognition subsystem; an intellectual property keyword extraction step, extracting candidate intellectual property keywords containing subject, behavior, time, place, causal relationship, and rights and obligations from the initial intellectual property evidence dataset based on a keyword list and synonym normalization rules in a multi-source legal corpus index; and an intellectual property legal fact generation step, assembling intellectual property legal facts from the candidate intellectual property keywords based on timeline reconstruction and role consistency verification, while simultaneously adding a correspondence between the intellectual property legal facts and the initial intellectual property evidence dataset for retrieval and verification. This method can improve the speed and accuracy of evidence processing, reduce irrelevant and ambiguous residual expressions, and ensure a one-to-one correspondence between intellectual property legal facts and original evidence through traceable mapping, thereby enhancing verifiability.

[0007] According to another aspect of the present invention, a system for generating intellectual property legal facts is provided. The system includes: an intellectual property evidence data acquisition module, used to generate an initial intellectual property evidence dataset with timestamps and a fixed key set based on oral statements transcribed by a speech recognition subsystem and written materials scanned by a text recognition subsystem, wherein the fixed key set includes the name of the speaker of the oral statement, the location of formation and collection of the written materials, the collection time, the carrier type, and the evidence number; an intellectual property keyword extraction module, used to extract candidate intellectual property keywords containing subject, behavior, time, place, causal relationship, and rights and obligations from the initial intellectual property evidence dataset based on a keyword list and synonym normalization rules in a multi-source legal corpus index, wherein the multi-source legal corpus index is pre-established based on an intellectual property judicial case database and an intellectual property law and regulation provisions database; and an intellectual property legal fact generation module, used to assemble the candidate intellectual property keywords into intellectual property legal facts based on timeline reconstruction and role consistency verification, while adding a correspondence between the intellectual property legal facts and the initial intellectual property evidence dataset for retrieval and verification.

[0008] The method and system for generating intellectual property legal facts according to embodiments of the present invention can improve the speed and accuracy of intellectual property evidence organization, reduce the residue of irrelevant and vague expressions, and ensure that intellectual property legal facts correspond one-to-one with original evidence through traceable mapping, thereby enhancing verifiability. Attached Figure Description

[0009] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0010] Figure 1 This is a flowchart of a method for generating intellectual property legal facts according to an embodiment of the present invention;

[0011] Figure 2 This is a flowchart of a method for generating intellectual property legal facts according to Embodiment 1 of the present invention; and

[0012] Figure 3 This is a schematic diagram of a system for generating intellectual property legal facts according to an embodiment of the present invention. Detailed Implementation

[0013] The following embodiments are merely examples to clearly illustrate the present invention and are not intended to limit the implementation of the invention. Those skilled in the art will recognize that other variations or modifications can be made based on the following description, and these variations, modifications, substitutions, and alterations arising from the principles and spirit of the present invention still fall within the protection scope of the present invention.

[0014] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0015] According to an embodiment of the present invention, a method for generating intellectual property legal facts is provided. Figure 1 This is a flowchart of a method for generating intellectual property legal facts according to an embodiment of the present invention, such as... Figure 1 As shown, the steps include S102-S106.

[0016] Step S102, Intellectual Property Evidence Data Acquisition Step: Based on the oral statements transcribed by the speech recognition subsystem and the written materials scanned by the text recognition subsystem, an initial intellectual property evidence dataset with timestamps and a fixed key set is generated. The fixed key set includes the name of the speaker of the oral statement, the location where the written materials were formed and collected, the collection time, the carrier type, and the evidence number.

[0017] Step S104, the intellectual property keyword extraction step, extracts candidate intellectual property keywords containing subject, behavior, time, place, causal relationship and rights and obligations from the initial intellectual property evidence dataset based on the keyword list and synonym normalization rules in the multi-source legal corpus index. The multi-source legal corpus index is pre-trained and established based on the intellectual property judicial case database and the intellectual property law and regulation provisions database.

[0018] Step S106, the intellectual property legal fact generation step, based on timeline reconstruction and role consistency verification, assembles the candidate intellectual property keywords into intellectual property legal facts, and adds the correspondence between the intellectual property legal facts and the initial intellectual property evidence dataset for retrieval and verification.

[0019] In related technologies, intelligent auxiliary systems often rely on rule matching or a single text data source, making it difficult to handle the complex chain of facts in intellectual property cases, which involves multiple stakeholders (rights holders, licensees, and infringers) and involves multiple voice statements, patent / contract / appraisal report documents. In this invention, the three steps of acquiring intellectual property evidence data, extracting intellectual property keywords, and generating intellectual property legal facts are linked together to achieve automatic conversion from natural language and written materials to intellectual property legal facts. This improves the speed and accuracy of evidence processing, reduces irrelevant and ambiguous expressions, and ensures a one-to-one correspondence between intellectual property legal facts and original evidence through traceable mapping, enhancing verifiability.

[0020] To illustrate the implementation of steps S102-S106, the present invention also provides the following specific embodiment 1.

[0021] Example 1

[0022] This embodiment provides a method for generating legal facts related to intellectual property rights, applicable to scenarios such as judicial case handling, arbitration hearings, and administrative review, and can run on computer systems or server clusters with voice and text recognition capabilities.

[0023] Figure 2 This is a flowchart of a method for generating intellectual property legal facts according to Embodiment 1 of the present invention. Figure 2 As shown, the method includes the following steps:

[0024] (I) Steps for Obtaining Intellectual Property Evidence Data

[0025] (1) Speech recognition processing

[0026] During judicial hearings or questioning, spoken language is input into a speech recognition subsystem. This subsystem employs a recognition algorithm combining acoustic and language models to transcribe spoken language into a written record. The transcription result includes a timestamp and speaker identification, forming a text of spoken evidence.

[0027] (2) Text recognition processing

[0028] Written materials (including paper files, electronic documents, contracts, invoices, etc.) are scanned and recognized using a text recognition subsystem. The recognition system extracts information such as the document's paragraph structure, page numbers, and the location where it was created, and adds the collection location and collection time identifiers to the output results.

[0029] (3) Generation of the initial intellectual property evidence dataset

[0030] Oral and written evidence are stored together to form an initial intellectual property evidence dataset with timestamps and a fixed key set (including an evidence source identifier). The evidence source identifier includes the name of the person making the statement, the location where the written material was created, and the location where it was collected. Each piece of intellectual property evidence data has a unique number, source_id, used for subsequent retrieval and comparison.

[0031] Through the above steps, standardized digital collection of multi-source evidence can be achieved, providing a unified input for subsequent keyword extraction and the formation of intellectual property legal facts.

[0032] (II) Steps for Extracting Intellectual Property Keywords

[0033] (1) Establishment of multi-source corpus

[0034] A multi-source legal corpus index was established based on the intellectual property judicial case database and the intellectual property laws and regulations database. This index stores standard legal terms used for keyword comparison, covering types such as subjects, behaviors, time, place, causation, and rights and obligations.

[0035] (2) Extraction of candidate intellectual property keywords

[0036] The text in the initial intellectual property evidence dataset is segmented, part-of-speech tagging, and entity recognition are performed to generate several candidate words. When a candidate word matches the keyword list in the index and satisfies the synonym normalization rules, the word is recorded as a candidate intellectual property keyword.

[0037] (3) Keyword selection and labeling

[0038] For the extracted candidate intellectual property keywords, the system records their location and corresponding timestamp in the evidence text, and generates a set of candidate intellectual property keywords. Each keyword carries the source evidence number and its context, which can be used for subsequent fact reconstruction.

[0039] This step transforms unstructured natural language evidence into structured legal language elements, achieving semantic unity.

[0040] (III) Steps for Generating Intellectual Property Legal Facts

[0041] (1) Timeline reconstruction

[0042] The system reconstructs the sequence of events based on time-related elements in the keyword set. Keywords from the same source of evidence are sorted by timestamp to form a sequence of semantic fragments arranged chronologically.

[0043] (2) Role consistency verification

[0044] Based on the timeline, consistency checks are performed on the actions of different entities. When the same entity appears in multiple pieces of evidence, the system confirms whether they are the same legal entity by comparing names, citizen ID numbers, and semantic similarity.

[0045] (3) Fact output and traceable mapping

[0046] The system combines the results of timeline reconstruction and role consistency verification into "intellectual property legal facts." Each intellectual property legal fact retains its corresponding original intellectual property evidence data pointer for subsequent review and comparison.

[0047] By following the steps above, the automatic conversion from natural language evidence to intellectual property legal facts can be achieved, significantly improving the efficiency and accuracy of fact-finding.

[0048] (IV) Implementation Results

[0049] The technical solution provided in Embodiment 1 forms a traceable chain of facts through three stages: standardized evidence collection, structured keyword extraction, and timeline and role verification. This method can automatically generate structured intellectual property legal facts from a large amount of raw audio and document evidence without relying on manual analysis, significantly improving the efficiency of case fact formation. System verification results show that in 500 case file samples, the fact recognition accuracy rate reaches 92%, and the average fact generation time is reduced to 1 / 10 of manual analysis. The system has good scalability and can be interfaced with case management systems, electronic case file systems, and voice acquisition systems.

[0050] According to an embodiment of the present invention, the intellectual property keyword extraction step S104 includes:

[0051] S1042, Construct a hierarchical keyword list And generate a standard token code id(v) for each token v;

[0052] S1044, perform page layout hierarchy reconstruction and speech wheel annotation on the initial intellectual property evidence dataset to form a set of structural anchor points A = {a k Each anchor point contains fixed anchor point fields: hierarchical position, sentence boundary, speaker, pause duration, speech rate, legal citation mark, and page / line number;

[0053] S1046, Define the candidate set for intellectual property keyword extraction Keywords containing a source pointer and a timestamp, 'v' appears in the rights and obligations section of an intellectual property case or intellectual property provision.

[0054] In this embodiment of the invention, a “lexicon-anchor point” binary pair C0 is established as a traceable starting point, substantially coupling the lexicon with the layout / turn-of-speech structure, ensuring the step-by-step inheritance and backtracking of the source, position and time in subsequent stages.

[0055] According to an embodiment of the present invention, S1046 defines the intellectual property keyword crawling candidate set C0 as including:

[0056] S10461, Perform synonym merging mapping π on C0: Fold regional styles and colloquial variations into standard word units. And maintain the merged traceability chain For traceability;

[0057] S10462, based on semantic proximity, temporal proximity, and source credibility, respectively, for each pair Calculate the ternary score (sem, tmp, src) ∈ [0, 1] 3 And record the calculation results on the intellectual property keyword crawling candidate set C0;

[0058] S10463, Let the gate threshold be θ=(θ s ,θ t ,θ r ), gate controller Output the candidate set after gating only if all three scores are not lower than the corresponding gating threshold.

[0059] In this embodiment of the invention, semantic, temporal, and source three-dimensional gating is applied to C0 of the intellectual property keyword crawling candidate set to obtain the purified C1.

[0060] According to an embodiment of the present invention, defining the intellectual property keyword crawling candidate set COS1046 further includes performing matrix fusion and graph regularization on C1, which includes:

[0061] S10464, gating the candidate set Construct the characteristic matrix with numbers i = 1…N. The three columns are semantic proximity, temporal proximity, and source credibility, respectively, and all three have been normalized to be dimensionless.

[0062] S10465, Set Case Type Condition Weights And ||w||1 = 1, calculate the fusion score vector.

[0063] S10466, Construct a similarity matrix based on anchor point adjacency. Where S ij ≥0, usually let S = S T And S ii =0;

[0064] S10467, define the degree matrix D = diag(S1). Then the Laplace matrix L = DS;

[0065] S10468, by minimizing the objective Learn w, where y is the historical review score and λ>0 is the dimensionless regularization coefficient;

[0066] S10469, sorting by s in descending order yields an ordered candidate column C2 = (c (1) ,…,c (N) It also preserves the source pointer and parent-child synonym chain for each item as a basis for traceable mapping of evidence to facts when generating subsequent intellectual property legal facts.

[0067] In this embodiment of the invention, the above C1 is transformed into a globally consistent sorted sequence C2 by matrix fusion using graph regularization, thereby improving the stability of high-value elements and fully transmitting the source and synonym chain to support subsequent context modeling and tracing.

[0068] According to an embodiment of the present invention, defining the intellectual property keyword crawling candidate set COS1046 further includes:

[0069] S10470, along C2 for each candidate c (i) Extract the window context and construct a third-order tensor. element Let be the embedding value of the i-th candidate u-th context in the e-th dimension, which is dimensionless;

[0070] S10471, Set the attention matrix With dimensionality reduction matrix Using Kronek kernel First-order expansion Projecting, we obtain the context enhancement vector c = softmax(T) (1) K1), where the attention matrix and the dimensionality reduction matrix are learnable parameters fitted offline or fine-tuned online from the historical case review set.

[0071] S10472, will As the new sorting criterion, output the weighted sequence. The significantly weakened samples, along with their context weights, are written into the conflict sample library Ξ for relearning.

[0072] In this embodiment of the invention, semantic density and contextual evidence are explicitly injected into the score through tensor attention, making the above ranking more robust at the discourse level, and the output C3 provides a higher confidence input for subsequent filtering and fact generation.

[0073] To further illustrate the implementation of the intellectual property keyword extraction step S104 in this invention, the present invention also provides the following specific embodiment 2.

[0074] Example 2

[0075] This embodiment 2, based on embodiment 1, provides a more detailed explanation of the steps for extracting intellectual property keywords. This step utilizes a multi-source legal corpus index to perform unified semantic processing on the initial intellectual property evidence data obtained from speech recognition and text recognition, in order to generate a structured set of candidate intellectual property keywords.

[0076] (I) Keyword List and Anchor Point Construction

[0077] The system first extracts the word set V = {v1, v2, ..., v} from the corpus based on a multi-source legal corpus index. NEach word element v corresponds to a legal subject, behavior, or rights and obligations element appearing in the case library or legal provisions. The system generates a unique code id(v) for each word element and establishes a set of structural anchor points A = {a} based on information such as the layout hierarchy, turn-taking boundaries, and sentence structure of the identified text. k}, where each anchor point ak contains a hierarchical position index, a sentence boundary, and a turn annotation.

[0078] The system forms an initial candidate set by pairing the word 'v' with its corresponding anchor 'a'.

[0079] C0 = {(v,a) ||v∈V,a∈A,v appears in the intellectual property evidence dataset}

[0080] Each candidate record contains a source pointer src_id and a timestamp t, which are used for subsequent tracing.

[0081] (II) Synonym Mapping and Gating Filtering

[0082] To reduce semantic redundancy, the system performs a synonym mapping π on C0:

[0083] Fold synonyms or near-synonyms into a unified standard word unit.

[0084] Then each candidate Calculate the ternary score (sem, tmp, src) ∈ [0, 1] 3 , representing semantic similarity, temporal proximity, and source credibility, respectively.

[0085] Let the gate threshold vector be θ = (θ s ,θ t ,θ r )

[0086] When all three scores are not lower than the corresponding threshold, let the gating function... otherwise

[0087] The candidate is retained only when G=1, and the output set is...

[0088] The typical threshold value is θ s =0.6, θ t =0.5, θ r =0.7.

[0089] (III) Weighted Calculation and Sorting Output

[0090] The system uses the historical review scoring vector y and the feature matrix X to...

[0091] Minimize objective function

[0092] Where λ>0 is the dimensionless regularization coefficient, L is the Laplacian matrix, and the learning parameter w is used to balance semantic and structural weights.

[0093] After obtaining the optimal weight w*, calculate the comprehensive score s = Xw*.

[0094] Sort the candidates in descending order of s to obtain an ordered candidate column C2 = (c (1) ,c (2) ,…,c (N) )

[0095] The system retains a source pointer and a parent-child synonym chain for each item so that the mapping can be traced when subsequent facts are generated.

[0096] (iv) Context Tensor and Attention Weighting

[0097] To characterize the support of keywords in the evidentiary context, the system performs a step along C2 for each candidate c. (i) Extract the window context and construct a third-order tensor. element Let represent the e-th dimension embedding (both are dimensionless) of the i-th candidate in the u-th context. Let the attention matrix be... With dimensionality reduction matrix (Both are learnable parameters), construct the Kronecker kernel

[0098] Will First-order expansion matrix Multiply by K and then pass through softmax to obtain the context weight vector.

[0099] Where 1 represents an all-1 vector. c is a dimensionless probability distribution used to weight the base score s.

[0100] (V) Weighted Sorting and Output

[0101] Calculate the weighted score according to Descending order yields a weighted sequence

[0102]

[0103] Each item retains its source pointer and parent-child synonym chain as a basis for tracing the "evidence → fact" mapping in subsequent intellectual property legal fact generation steps.

[0104] (VI) Implementation Results

[0105] This embodiment achieves automatic conversion from original intellectual property evidence data to a standard keyword set through a unified lexical set V, an anchor point set A, and a mapping π.

[0106] By employing gated threshold filtering and regularization learning algorithms, the system effectively reduces redundant terms and improves semantic consistency. In actual testing, keyword filtering accuracy improved by approximately 10%, and irrelevant term retention decreased by approximately 15%, providing high-quality input for subsequent generation of intellectual property legal facts.

[0107] According to an embodiment of the present invention, the intellectual property legal fact generation step S106 includes:

[0108] S1061, using the conflict sample library Ξ and historical verification, pseudo-labels are generated, and the recoverability probability p of each weakened sample is estimated using the trained discriminator h. i ∈[0,1];

[0109] S1062, for C + middle Lower but p i Samples greater than the threshold κ are included in the enhancement set E. + The rest remain in their original order, where the threshold κ∈(0,1);

[0110] S1063, regarding the enhancement set E + The internal samples undergo only a one-time position boosting to form the final grab sequence C4, where the boosting set E + ={i|p i ≥κ},

[0111] S1064, write back the error statistics of h to the gate threshold θ and the weight file w, and declare C4 as the unique keyword input for the fact generation step.

[0112] In this embodiment of the invention, semi-supervised rescreening, while ensuring overall accuracy, moderately recovers high-value long-tail segments, significantly improves the quality and adaptability of the output C4, and drives the front-end threshold and weight to be updated adaptively.

[0113] According to an embodiment of the present invention, the timeline reconstruction of the intellectual property legal fact generation step S106 includes:

[0114] S1065 maps the case timeline into a measurement space. Construct an indicator function χ for each segment in C4 i (t) and with smooth kernel g σ Smoothing Forming candidate time tensors

[0115] S1066, the index set S of candidate fact k k Calculate integrals with consistent order

[0116]

[0117] Let the threshold τ∈[0,1] and the penalty coefficient α>0 be defined, and the time criterion Φ be defined. k =I k -αΩ k ;

[0118] Only when Φ k When ≥τ, S k The submission role consistency optimization ensures that the candidate facts entering the generation of intellectual property legal facts are self-consistent in the time dimension, and the reasons for non-compliance are recorded and written back to the conflict sample library Ξ.

[0119] In this embodiment of the invention, C4 is projected onto a continuous time measure space and the overlapping intervals are penalized, which effectively constrains the sequential relationship and time density, ensuring that the fact candidates entering the role optimization have temporal self-consistency.

[0120] According to an embodiment of the present invention, the intellectual property legal fact generation step S106 further performs role consistency / evidence mapping optimization on the time-consistent candidate set, including:

[0121] S1067, Construction of the Main Character Diagram: Based on the Time Criterion Φ k For candidate sets ≥τ, extract the relevant subjects and their evidence, and construct a weighted adjacency matrix. The rows and columns correspond to the subject, and the edge weights are dimensionless scores that combine the strength of evidence citation and consistency. The degree matrix D = diag(A1) and the Laplace matrix L = DA are defined.

[0122] S1068, Semi-supervised spectral optimization: Set the role label matrix It is known prior that the set of subjects Γ has a label R0;

[0123] by Solve Where P is the projection matrix for selecting the Γ rows, λ,β>0 are dimensionless hyperparameters, and ||·|| * For nuclear norm;

[0124] S1069, Role Uncertainty and Reduction: Defining Uncertainty for Each Agent m And calculate the role reduction coefficient for each candidate fact k, where Ent(k) is the set of subjects involved in fact k;

[0125]

[0126] S1070, ρ k The subsequent evidence mapping and causal chaining are used as the role quality weights.

[0127] In this embodiment of the invention, through joint optimization of graph smoothing, prior anchoring, and low-rank constraints, the role labels are consistent across the "evidence-keyword-fact" chain; the reduction coefficient ρ k By explicitly injecting the risk of uncertain subjects into downstream matching, the subject drift of intellectual property legal facts caused by pronouns and aliases can be suppressed.

[0128] According to an embodiment of the present invention, the intellectual property legal fact generation step S106 further includes:

[0129] S1071, Compatibility Fraction Construction: Constructing a candidate set of evidence for each candidate k of intellectual property legal facts. And based on the context tensor in claim 5 Compatibility fractions defined with Kronecker kernel K

[0130]

[0131] Where e i ,q k For dimensionless embedding, γ>0 is the weighting coefficient;

[0132] S1072, Global Gain Objective: Define a one-to-one mapping M: And maximize

[0133]

[0134] Subject to uniqueness and credibility constraints:

[0135]

[0136] in,

[0137] η1,η2,η3>0 indicates a dimensionless weighted average.

[0138] W i The source credibility score is a composite of the evidence source channel, the proximity of the collection time, and the source level, and is normalized to [0,1].

[0139] S1073, Tensor Kernel Relaxation and Iterative Solution: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] Rewritten in matrix-tensor simultaneous form, let Z∈{0,1} |E|×|K| To select the matrix, g = {Φ k}、r={ρ k}、φ={φ ik}、 but

[0140]

[0141] Where R ik =r k Φik =φ ik , ‖·‖ * Let λ be the nuclear norm. c >0;

[0142] S1074 is approximately optimal by using augmented Lagrange multiplication and alternating minimization iterations, thereby restoring M.

[0143] In this embodiment of the invention, "time consistency Φ" is used. k "and "the main body is reliable ρ" k "and evidence compatibility" ik , "Merge into an optimizable global objective, and obtain a stable one-to-one mapping under tensor kernel relaxation and uniqueness constraints, ensuring that every intellectual property legal fact is supported by real evidence fragments and is traceable."

[0144] To illustrate weighted sequences In the application of step S106 of generating intellectual property legal facts, the following embodiments 3 to 7 provide a complete process for generating intellectual property legal facts. This process takes C3, the output of embodiment 2, as input, wherein each sample... It has context weights, source pointers, and confidence scores. i All of these are dimensionless quantities.

[0145] Example 3: Semi-supervised rescreening and enhanced set formation

[0146] In this embodiment 3, through semi-supervised rescreening and probability enhancement, the system recovers high-value long-tail fragments while maintaining overall accuracy, improves sample diversity, and drives adaptive updates of the front-end threshold.

[0147] The system first establishes a conflict sample database Ξ, storing samples that have been significantly weakened or whose boundaries are uncertain. A pseudo-label vector y is then generated using historical verification samples. ′ And train the discriminator h ψ (x) Estimate the recoverable probability of each weakened sample:

[0148] p i =σ(h) ψ (x i ))∈(0,1), where σ(·) is the Sigmoid function.

[0149] when And p i >κ (where the threshold κ∈[0.4,0.6]), If the average score is used, then sample i is included in the enhancement set:

[0150]

[0151] The augmented set of samples undergoes only a one-time positional boost, resulting in the rearranged sequence C4. The output C4 of this process is declared as the unique keyword input for the fact generation step.

[0152] Error statistics of eliminated samples It is written back to the gating threshold file and weight file for front-end adaptive updates.

[0153] Example 4: Time Consistency Verification

[0154] In this embodiment 4, facts are projected onto a continuous time measurement space through a time consistency mechanism, with overlapping penalty intervals, ensuring that the candidate facts entering the role optimization stage are consistent in both time order and density.

[0155] The system normalizes the case timeline to the interval [0,1], and defines a time indication function χ for each candidate segment i in C4. i (t)∈{0,1}, and a smoothing kernel g is used. σ (t) (such as a Gaussian kernel) yields the smoothing indicator function.

[0156]

[0157] For any fact candidate k, the index set S k Calculate integrals in the same order

[0158]

[0159] Based on this definition, the time consistency criterion Φ is... k =I k -αΩ k

[0160] Where α>0 is the overlap penalty coefficient. Only when Φ k When the value is greater than or equal to τ (τ∈[0,1] is the criterion threshold), the fact candidate k enters the role consistency optimization; otherwise, the reason is recorded and Ξ is written back.

[0161] Example 5: Main Role Consistency Optimization

[0162] In this embodiment 5, the role labels are kept consistent along the "evidence-keyword-fact" chain through joint optimization of graph smoothing and low-rank constraints. The reduction coefficient explicitly injects the risk of uncertain subjects into downstream matching, suppressing subject drift caused by pronouns and aliases.

[0163] For satisfying Φ k The candidate set ≥τ is used to extract the subjects involved and their evidence, construct a weighted adjacency matrix A (the edge weight is the dimensionless score of "evidence citation strength and consistency fusion"), and define the degree matrix D = diag(A1) and the Laplace L = DA.

[0164] Set up a character tag matrix Given a set of subjects Γ with prior labels R0, solve using semi-supervised spectral optimization:

[0165] by Solve

[0166] Where P is the projection matrix for selecting the Γ rows, and λ,β>0.

[0167] get Then, define the subject uncertainty.

[0168] And based on this, the character reduction factor is calculated: ρ k =∏ m∈Ent(k) (1-u m )

[0169] Example 6: Evidence-Fact Mapping and Global Gain Optimization

[0170] The optimization process in this embodiment 6 unifies time consistency, subject reliability, and evidence compatibility under the same optimizable framework, and obtains a stable mapping under tensor kernel relaxation and uniqueness constraints, ensuring that each intellectual property legal fact is supported by real evidence fragments and is traceable.

[0171] Evidence candidate set for each fact candidate k Above, we define the compatibility fraction.

[0172]

[0173] Where e i ,q k For dimensionless embedding, Let K be the context tensor, K be the Kronecker kernel, and γ > 0.

[0174] Let M be a unique mapping: Maximize the global gain target given by weight 9:

[0175]

[0176] Satisfy constraints:

[0177]

[0178] Where η1, η2, η3 > 0 are weighting coefficients, W i ω represents the source credibility, and ω is its lower limit.

[0179] The system ultimately outputs a fact-evidence mapping table, recording (k, M(k)) and corresponding Φ. k ,ρ k ,φM(k),k , In addition to evidence source identification and timestamps, it achieves a one-to-one correspondence between "intellectual property legal facts and original evidence" and full-chain verifiability.

[0180] According to embodiments of the present invention, a system for generating intellectual property legal facts is also provided. Figure 3 This is a schematic diagram of a system for generating intellectual property legal facts according to an embodiment of the present invention. Figure 3 As shown, it includes: intellectual property evidence data acquisition module 32, intellectual property keyword extraction module 34, and intellectual property legal fact generation module 36.

[0181] The intellectual property evidence data acquisition module 32 is used to generate an initial intellectual property evidence dataset with timestamps and a fixed key set based on the oral statements transcribed by the speech recognition subsystem and the written materials scanned by the text recognition subsystem. The fixed key set includes the name of the speaker of the oral statement, the location of formation and collection of the written materials, the collection time, the carrier type, and the evidence number.

[0182] The intellectual property keyword extraction module 34 is used to extract candidate intellectual property keywords containing subject, behavior, time, place, causal relationship and rights and obligations from the initial intellectual property evidence dataset based on the keyword list and synonym normalization rules in the multi-source legal corpus index library. The multi-source legal corpus index library is pre-established based on the intellectual property judicial case library and the intellectual property law and regulation provisions library.

[0183] The intellectual property legal fact generation module 36 is used to reconstruct the timeline and verify the consistency of roles, and to form intellectual property legal facts from the candidate intellectual property keywords. At the same time, it adds the correspondence between the intellectual property legal facts and the initial intellectual property evidence dataset for retrieval and review.

[0184] The unified declaration of symbols and units in this invention is as follows.

[0185] General and Set:

[0186] · A hierarchical keyword vocabulary set, where each element is a lexical unit v, which is dimensionless.

[0187] ·id(v): Standardized token encoding (string / number), dimensionless.

[0188] ·A={a k}: A set of structural anchor points (derived from the layout / turn-of-speech location of the evidence), dimensionless, but each anchor point contains the following fields: hierarchical position (integer / string), sentence boundary (character range), speaker (string), pause duration (seconds), and speech rate (words / second). -1), legal citation markers (Boolean).

[0189] • C0, C1, C2, C3, C4: Candidate pairs or their sequences, all of which are sets / sequences of indices and are dimensionless.

[0190] ·S k : A set of candidate indices that constitutes a candidate k of a certain intellectual property legal fact, dimensionless.

[0191] Grab and Gating:

[0192] ·π: Synonym mapping folds regional styles / colloquial variations into standard word units. Dimensionless.

[0193] ·(sem,tmp,src)∈[0,1] 3 : These are the scores for semantic proximity, temporal proximity, and source credibility, respectively, all of which are dimensionless.

[0194] ·θ=(θ s ,θ t ,θ r )∈[0,1] 3 : Three-valued gate threshold, dimensionless.

[0195] · The gated Boolean output is dimensionless.

[0196] · The merged original word set is used for evidence → keyword etymology tracing and is dimensionless.

[0197] Matrix fusion and graph regularization:

[0198] · The candidate feature matrix (three columns: semantic, time, and source) is dimensionless.

[0199] · ‖w‖1=1: Case type condition weight, dimensionless.

[0200] · The fusion score vector prioritizes the elements of intellectual property legal facts and is dimensionless.

[0201] · Similarity matrix, S ij ≥0, S ii =0, dimensionless.

[0202] ·D=diag(S1): a degree matrix, with diagonal elements being row sums, dimensionless.

[0203] ·L=DS: Graph Laplace matrix, dimensionless.

[0204] •λ>0: Graph smoothing regularity intensity, dimensionless.

[0205] · Historical review scoring (annotated results), dimensionless.

[0206] ·J(w): Objective function, dimensionless.

[0207] ·C2=(c (1) ,…,c (N) Evidence in descending order of s: Keywords are sequential and dimensionless.

[0208] Tensor context:

[0209] · Candidate × Context × Embedding third-order tensor, dimensionless.

[0210] · Context attention matrix, dimensionless.

[0211] · Embedded dimension-reduced matrix, dimensionless.

[0212] · Kronecker nucleus, dimensionless.

[0213] ·c = softmax(T) (1) K1): Contextual reinforcement probability, dimensionless.

[0214] · Enhanced integration score, priority of legal factual elements in intellectual property (updated version), dimensionless.

[0215] ·Ξ: Conflict sample library (a set of weakened but potentially valuable pieces of evidence), dimensionless.

[0216] Semi-supervised rescreening:

[0217] •h: Discriminator model (estimated "recoverable"), dimensionless output.

[0218] ·p i ∈[0,1]: The recoverable probability of sample i, dimensionless.

[0219] ·κ∈(0,1): The threshold for adding to the enhancement set, dimensionless.

[0220] ·E + : Enhancement set (used for one-time position lifting), dimensionless.

[0221] • Final output C4: The unique keyword input sequence for entering the "Intellectual Property Legal Fact Generation Steps", dimensionless.

[0222] Time-measured space:

[0223] ·t∈[0,1]: Normalized time, dimensionless.

[0224] ·χ i (t)∈{0,1}: The time indicator function of candidate segment i, dimensionless.

[0225] ·g σ (t): smooth kernel (such as Gaussian kernel), dimensionless;

[0226] · Candidate time tensor entries, dimensionless.

[0227] ·I k The sequential consistency integral of candidate fact k is dimensionless.

[0228] ·Ω k Interval overlap penalty, dimensionless.

[0229] ·Φ k =I k -αΩ k Time consistency criterion, dimensionless.

[0230] • α>0: Overlap penalty coefficient, dimensionless; τ∈[0,1]: Time criterion threshold, dimensionless.

[0231] Character illustrations and matrix:

[0232] · The adjacency matrix of the main roles consists of evidence citation strength and consistency scores, which have been normalized and are dimensionless.

[0233] ·D=diag(A1): Role degree matrix, the diagonal elements are the total strength of evidence connections for each subject, dimensionless.

[0234] ·L=DA: Laplace matrix of the character diagram, dimensionless.

[0235] Labels and Priors:

[0236] · The character tag matrix has rows corresponding to the subject, columns corresponding to the character category, and entries representing dimensionless probability values.

[0237] ·Γ: The set of known character tags, dimensionless.

[0238] ·R0: The known character label matrix in Γ, with entries being Boolean or probability values, dimensionless.

[0239] • P: Selection matrix used for projecting R to Γ, dimensionless.

[0240] ·λ,β>0: Hyperparameters, which are the prior constraint weights and low-rank constraint weights, respectively, and are dimensionless.

[0241] Uncertainty and Reduction:

[0242] · The uncertainty of subject m is dimensionless.

[0243] ·ρ k =∏ m∈Ent(k) (1-u m ): The role reduction coefficient of fact k, with a value range of (0,1) and dimensionless.

[0244] ·Ent(k): The set of subjects involved in fact k, dimensionless.

[0245] In summary, according to the above embodiments of the present invention, a method and system for generating intellectual property legal facts are provided. The method includes: an intellectual property evidence data acquisition step, generating an initial intellectual property evidence dataset with timestamps and a fixed key set based on oral statements transcribed by a speech recognition subsystem and written materials scanned by a text recognition subsystem, wherein the fixed key set includes the name of the speaker of the oral statement, the location of formation and collection of the written materials, the collection time, the carrier type, and the evidence number; an intellectual property keyword extraction step, extracting candidate intellectual property keywords containing subject, behavior, time, location, causal relationship, and rights and obligations from the initial intellectual property evidence dataset based on a keyword list and synonym normalization rules in a multi-source legal corpus index, wherein the multi-source legal corpus index is pre-trained and established based on an intellectual property judicial case database and an intellectual property legal regulations database; and an intellectual property legal fact generation step, assembling intellectual property legal facts from the candidate intellectual property keywords based on timeline reconstruction and role consistency verification, while simultaneously adding a correspondence between the intellectual property legal facts and the initial intellectual property evidence dataset for retrieval and verification. This method achieves automatic conversion from natural language and written materials to intellectual property legal facts by linking three steps: intellectual property evidence data acquisition, intellectual property keyword extraction, and intellectual property legal fact generation. This improves the speed and accuracy of evidence processing, reduces irrelevant and vague expressions, and ensures a one-to-one correspondence between intellectual property legal facts and original evidence through traceable mapping, thereby enhancing verifiability.

[0246] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating legal facts related to intellectual property rights, characterized in that, include: The steps for acquiring intellectual property evidence data are as follows: Based on the oral statements transcribed by the speech recognition subsystem and the written materials scanned by the text recognition subsystem, an initial intellectual property evidence dataset with timestamps and a fixed key set is generated. The fixed key set includes the name of the speaker of the oral statements, the location of formation and collection of the written materials, the collection time, the carrier type, and the evidence number. The intellectual property keyword extraction step involves extracting candidate intellectual property keywords containing subject, behavior, time, place, causal relationship and rights and obligations from the initial intellectual property evidence dataset based on the keyword list and synonym normalization rules in the multi-source legal corpus index. The multi-source legal corpus index is pre-trained and established based on the intellectual property judicial case database and the intellectual property law and regulation provisions database. The steps for generating intellectual property legal facts, based on timeline reconstruction and role consistency verification, combine the candidate intellectual property keywords into intellectual property legal facts, and simultaneously add the correspondence between the intellectual property legal facts and the initial intellectual property evidence dataset for retrieval and verification. The intellectual property keyword extraction step includes: Build a hierarchical keyword list and for each word element Generate standard lexical codes ; The initial intellectual property evidence dataset is subjected to page hierarchy reconstruction and turn annotation respectively, forming a set of structural anchor points A={a k Each anchor point contains fixed anchor point fields: hierarchical position, sentence boundary, speaker, pause duration, speech rate, legal citation mark, and page / line number; Define the candidate set of intellectual property keywords This includes keyword pairs with accompanying source pointers and timestamps. Appearing in the facts of intellectual property cases or the rights and obligations section of intellectual property provisions, where a candidate set of intellectual property keywords is defined. include: exist Perform synonym merging mapping π on top: → Folding regional styles and colloquial variations into standard word units and maintain the merged traceability chain. For traceability; For each pair, based on semantic similarity, temporal proximity, and source credibility respectively... Calculate the ternary score The results are then recorded in the intellectual property keyword crawling candidate set based on the calculation results. superior; Set the threshold gate controller The gating candidate set is output if and only if all three ternary scores are not lower than the corresponding gating threshold. ={( a) | G=1}; The timeline reconstruction of the intellectual property legal fact generation steps includes: Mapping the case timeline into a measure space ,right Construct indicator functions for each segment and with smooth kernel Smoothing Forming candidate time tensors ,in Enter the unique keyword for the fact generation step; For candidate facts index set Calculate integrals with consistent order Set threshold With penalty coefficient Define time criteria ; Only when When The submission role consistency optimization ensures that candidate facts entering the generation of intellectual property legal facts are self-consistent in the time dimension, and reasons for non-compliance are recorded and written back to the conflict sample database. .

2. The method for generating intellectual property legal facts according to claim 1, characterized in that, Define the candidate set of intellectual property keywords Also includes the Perform matrix fusion and graph regularization, which includes: Gated candidate set Number Construct the feature matrix The three columns are semantic proximity, temporal proximity and source credibility, respectively, and the three have been normalized to be dimensionless. Set case type condition weights and Calculate the fusion score vector ; Construct a similarity matrix based on the adjacency of anchor points S ij ≥0, usually let S=S ⊤ And S ii =0; Define degree matrix ; Then the Laplace matrix ; By minimizing the target study ,in Scoring for historical review These are dimensionless canonical coefficients; according to Descending order yields ordered candidate columns Furthermore, it retains the source pointer and parent-child synonym chain for each item as a basis for traceable mapping of evidence to facts when generating subsequent intellectual property legal facts.

3. The method for generating intellectual property legal facts according to claim 2, characterized in that, Define the candidate set of intellectual property keywords Also includes: along For each candidate Extract the window context and construct a third-order tensor. ,element For the first Candidate No. The context of the article in the first The embedded values ​​of each dimension are dimensionless; Set attention matrix With dimensionality reduction matrix Using Kronek nucleus First-order expansion Projection yields the context enhancement vector. The attention matrix and the dimensionality reduction matrix are learnable parameters that are fitted offline or fine-tuned online from a historical case review set. ; Will As the new sorting criterion, output the weighted sequence. The significantly weakened samples, along with their context weights, are then written into the conflict sample library. In preparation for further study.

4. The method for generating intellectual property legal facts according to claim 3, characterized in that, The steps for generating intellectual property legal facts include: Using the aforementioned conflict sample library And generate pseudo-labels through historical verification, and use them to train the discriminator. Estimate the recoverable probability of each weakened sample. ; right and The samples were included in the enhancement set. The rest remain in their original order, with the threshold... ; For the enhancement set The internal samples undergo only a one-time position boost to form the final grab sequence. Among them, the enhanced set , Will Error statistics are written back to the gate threshold. With the weighted archive .

5. The method for generating intellectual property legal facts according to claim 4, characterized in that, The intellectual property legal fact generation step further implements role consistency / evidence mapping optimization on the time-consistent candidate set, including: Main character diagram construction: based on time criterion From the candidate set, extract the relevant subjects and their evidence, and construct a weighted adjacency matrix. In this model, rows and columns correspond to the main body, and edge weights are dimensionless scores that combine evidence citation strength and consistency. A degree matrix is ​​also defined. With Laplace matrix ; Semi-supervised spectrum optimization: setting up a role label matrix A priori known set of subjects With tags ; by Solve ,in For selection The projection matrix of the row, It is a dimensionless hyperparameter. For nuclear norm; Role uncertainty and reduction: for each subject Define uncertainty and for each candidate fact Calculate the character reduction factor in For the facts Involves a set of subjects; Will The subsequent evidence mapping and causal chaining are used as the role quality weights.

6. The method for generating intellectual property legal facts according to claim 5, characterized in that, The steps for generating intellectual property legal facts also include: Compatibility score construction: For each candidate intellectual property legal fact Establish a candidate set of evidence And based on the context tensor in claim 3 With Kronek nucleus Define compatibility fractions in For dimensionless embedding, For the weighting factor; Global gain target: Define one-to-one mapping And maximize Subject to uniqueness and credibility constraints: in, For dimensionless weighted average, The source credibility score is a composite of the evidence source channel, the proximity of the collection time, and the source level, and is normalized to [0,1]. Tensor kernel relaxation and iterative solution: Rewritten in matrix-tensor simultaneous form, let ? To select a matrix, , , , ,but in , , , For nuclear norm, ; Approximate optimality is obtained through iterative augmented Lagrange multiplication and alternation minimization, thereby restoring... .

7. A system for generating legal facts related to intellectual property rights, characterized in that, include: The intellectual property evidence data acquisition module is used to generate an initial intellectual property evidence dataset with timestamps and a fixed key set based on oral statements transcribed by the speech recognition subsystem and written materials scanned by the text recognition subsystem. The fixed key set includes the name of the speaker of the oral statement, the location of formation and collection of the written materials, the collection time, the carrier type, and the evidence number. The intellectual property keyword extraction module is used to extract candidate intellectual property keywords containing subject, behavior, time, place, causal relationship and rights and obligations from the initial intellectual property evidence dataset based on the keyword list and synonym normalization rules in the multi-source legal corpus index library. The multi-source legal corpus index library is pre-established based on the intellectual property judicial case library and the intellectual property law and regulation provisions library. The intellectual property legal facts generation module is used to reconstruct the timeline and verify the consistency of roles, and to form intellectual property legal facts from the candidate intellectual property keywords. At the same time, it adds the correspondence between the intellectual property legal facts and the initial intellectual property evidence dataset for retrieval and review. The intellectual property keyword extraction module is used for: Build a hierarchical keyword list and for each word element Generate standard lexical codes ; The initial intellectual property evidence dataset is subjected to page hierarchy reconstruction and turn annotation respectively, forming a set of structural anchor points A={a k Each anchor point contains fixed anchor point fields: hierarchical position, sentence boundary, speaker, pause duration, speech rate, legal citation mark, and page / line number; Define the candidate set of intellectual property keywords This includes keyword pairs with accompanying source pointers and timestamps. Appearing in the facts of intellectual property cases or the rights and obligations section of intellectual property provisions, where a candidate set of intellectual property keywords is defined. include: exist Perform synonym merging mapping π on top: → Folding regional styles and colloquial variations into standard word units and maintain the merged traceability chain. For traceability; For each pair, based on semantic similarity, temporal proximity, and source credibility respectively... Calculate the ternary score The results are then recorded in the intellectual property keyword crawling candidate set based on the calculation results. superior; Set the threshold gate controller The gating candidate set is output if and only if all three ternary scores are not lower than the corresponding gating threshold. ={( a) | G=1}; The timeline reconstruction of the intellectual property legal fact generation steps includes: Mapping the case timeline into a measure space ,right Construct indicator functions for each segment and with smooth kernel Smoothing Forming candidate time tensors ,in Enter the unique keyword for the fact generation step; For candidate facts index set Calculate integrals with consistent order Set threshold With penalty coefficient Define time criteria ; Only when When The submission role consistency optimization ensures that candidate facts entering the generation of intellectual property legal facts are self-consistent in the time dimension, and reasons for non-compliance are recorded and written back to the conflict sample database. .

Images