Patent retrieval method based on feature-effect matrix and large language model

Abstract

The application discloses a patent retrieval method based on a feature-effect matrix and a large language model, and feature-effect matrix analysis and construction are performed on a patent text set to be processed by using the text generalization and downstream task potential of the large language model, without manual marking, so that the cost is reduced. In the analysis and construction step of the large language model, at least one feature-effect matrix contained in the patent text to be processed and a patent summary are obtained. The application constructs a vector, an effect interval mixed retrieval mode based on the feature-effect matrix, accurately matches the complex query requirements of the patent, and further improves the retrieval effect of the related patent.

Description

Technical Field

[0001] This invention relates to the field of patent search technology, and in particular to a patent search method based on feature-efficacy matrix and large language model. Background Technology

[0002] Patents are an important form of intellectual property protection, closely related to scientific, technological, and economic development. With the massive increase in the number of patents, patent analysis, retrieval, and evaluation have become exceptionally demanding tasks. The feature-efficacy matrix is ​​a core component of a patent, directly impacting the analysis, retrieval, and evaluation process. Analyzing, retrieving, and evaluating the feature-efficacy matrix can help quickly capture the core content of a patent, significantly improving patent analysis efficiency. Since the release of ChatGPT, Large Language Models (LLMs) have widely penetrated the public eye, becoming a core tool of general artificial intelligence, capable of efficiently handling diverse tasks such as text summarization, information extraction, and question answering. However, LLMs still face challenges in the vertical field of patent retrieval, analysis, and evaluation, including knowledge gaps and insufficient professional adaptation, resulting in a scarcity of existing research. This patent focuses on filling this gap, exploring the deep application and optimization paths of LLMs in patents. In existing technologies, pre-trained models can be used to analyze feature-efficacy matrices. However, this method requires manual annotation of a large number of sample texts, leading to increased costs. Furthermore, patent efficacy is multi-dimensionally coupled, and traditional retrieval methods, relying on only a few keywords, struggle to accurately match complex needs, further limiting retrieval effectiveness and quality. Summary of the Invention

[0003] Therefore, it is necessary to propose a patent retrieval method based on feature-efficacy matrix and large language model to address the above problems.

[0004] A patent retrieval method based on feature-efficacy matrix and large language model, the method comprising:

[0005] S1: Obtain the first set of patent texts to be processed, and divide the first set of patent texts to be processed into a target training set and a target verification set according to different levels of patent quality labels;

[0006] S2: Extract features from the target training set to obtain training text feature information, input the training text feature information into the teacher large language model, the teacher large language model outputs the first feature-efficacy matrix and the first patent summary for each patent text in the target training set; use the first feature-efficacy matrix and the first patent summary as the final label and fuse them with the corresponding target training set to obtain the first labeled training set;

[0007] S3: The first student large language model is obtained by performing full-parameter pre-training, supervised fine-tuning, reinforcement learning or knowledge editing training on the base large language model through the first labeled training set;

[0008] S4: Periodically extract a second set of patent texts to be processed from newly published patents. After being annotated by the teacher model in step S2, a second annotated training set is formed. When the distribution drift index D_KL of the second annotated training set exceeds a preset threshold δ, the first student large language model is updated by training the second annotated training set through step S3 to obtain the second student large language model. In step S3, an elastic weight consolidation or memory replay mechanism is introduced to prevent the second student large language model from experiencing performance regression on the first annotated training set during the training of the second annotated training set.

[0009] S5: Extract features from the target verification set to obtain verification text feature information, input the verification text feature information into the second student big language model, and the second student big language model outputs the second feature-efficacy matrix and the second patent summary for each patent text in the target verification set;

[0010] S6: Compare each item of the different second feature-efficacy matrices of the same patent text in the target validation set; if one or more items are not equal, call the teacher large language model. Each teacher large language model outputs a triplet and the confidence score of each item. Sort the triplets in descending order of their corresponding confidence scores, select the triplet with the highest confidence score as the final authority value of the target validation set, and fuse the final authority value with the corresponding patent text to obtain the authoritative patent text. Write the authoritative patent text back to the target training set and the target validation set.

[0011] S7: Determine the third feature-efficacy matrix of the second set of patent texts to be processed through steps S1 and S2. Multiple third feature-efficacy matrices constitute a retrieval library, which includes a feature-efficacy vector retrieval library and a feature-efficacy numerical retrieval library.

[0012] S8: Receive search keywords in natural language input, call the second student large language model to calculate the keyword feature-efficacy matrix of the search keywords, perform semantic matching between the keyword feature-efficacy matrix and the third feature-efficacy matrix in the feature-efficacy vector retrieval library, and output multiple third feature-efficacy matrices with a similarity greater than a first threshold to the keyword feature-efficacy matrix. Multiple patents corresponding to multiple third feature-efficacy matrices constitute a first candidate patent list. Receive the specified feature name, numerical range, and unit of measurement, convert the numerical range to a numerical range under a standard unit system according to unit conversion relationships, filter the range in the execution interval of the feature-efficacy numerical retrieval library corresponding to the third feature-efficacy matrix, and output a second candidate patent list that meets the conditions.

[0013] In one embodiment, the process of obtaining the graded patent quality label is as follows:

[0014]

[0015]

[0016]

[0017]

[0018] in, c_i represents the citation score for each patent text in the first set of patent texts to be processed; c_i represents the number of patent citations for each patent text. and These represent the average and standard deviation of patent citations for the first set of patent texts to be processed; recency_score is the timeliness score for each patent text; time_now is the evaluation benchmark date for each patent text; time_grant is the grant announcement date for each patent text; authorize_score is the grant rate score for each patent text; g_i is the number of patents granted for each patent text; and g_total is the number of patents granted in the same patent family within the first set of patent texts to be processed. A graded patent quality label for each patent text; , , All are fusion coefficients.

[0019] In one embodiment, dividing the first set of patent texts to be processed into a target training set and a target verification set according to different levels of patent quality labels includes:

[0020] S101: Multiple graded patent quality labels constitute a graded patent quality label set;

[0021] S102: Discretize multiple levels of patent quality labels into several level label ranges according to manually defined threshold ranges;

[0022] S103: Calculate the proportion of each grade label range in the patent quality label set of the stated grade;

[0023] S104: For each level label interval, determine the first number of patent texts corresponding to it in the first set of patent texts to be processed, randomly select a% of patent texts from the first number of patent texts to form a target verification set, and randomly select b% of patent texts from 1-a% of the patent texts in the first set of patent texts to be processed to form the target training set, where a and b are agreed constants.

[0024] In one embodiment, dividing the first set of patent texts to be processed into a target training set and a target verification set according to different levels of patent quality labels includes:

[0025] S105: Discretize multiple levels of patent quality labels into several level label ranges according to manually defined threshold ranges;

[0026] S106: Each level label interval corresponds to a patent text sample set in the first set of patent texts to be processed. The patent text sample set with the smallest proportion of patent texts in the first set of patent texts to be processed is the rare level sample. Set an amplification factor ri, and extract ri*a% of patent texts from the rare level sample and put them into the first training set; where, if ri*α%>1, then ri*α% is set to 1.

[0027] S107: For the patent text sample set corresponding to each level label interval except for rare level samples, set a reduction coefficient ti, and extract ti*a% patent texts from each patent text sample set to form a second training set. The first training set and the second training set constitute the target training set.

[0028] S108: Randomly select b% patent texts from the 1-ri*a%-ti*a% patent texts in the first set of patent texts to be processed to form a target verification set. The target verification set is not scaled up or down synchronously with the target training set.

[0029] In one embodiment, the final label is obtained by selecting the first feature-efficacy matrix with the highest number of votes and the first patent summary with the highest number of votes as the final label; if there is a tie, random and uniform sampling is performed.

[0030]

[0031] in, Let i be the maximum number of votes for the first feature-efficacy matrix or the maximum number of votes for the first patent summary, where i is the index of the teacher's large language model. Let L be the i-th teacher's large language model, and L be the set of teacher's large language models. For the target training set, argmax is the first feature-efficacy matrix or first patent summary output by the i-th teacher's large language model, and argmax is the majority vote operation.

[0032] In one embodiment, the final label is obtained by selecting the first feature-efficacy matrix with the highest weighted votes and the first patent summary with the highest weighted votes as the final label;

[0033]

[0034] in, The maximum weighted vote count for the first feature-efficacy matrix or the maximum weighted vote count for the first patent summary, where i is the teacher's large language model index. Let L be the i-th teacher's large language model, and L be the set of teacher's large language models. For the target training set, The first feature-efficacy matrix and the first patent summary are output by the i-th teacher's large language model. argmax represents the confidence score output by the i-th teacher's large language model, and argmax represents the majority vote operation.

[0035] In one embodiment, the step of obtaining the first student large language model by performing full-parameter pre-training, supervised fine-tuning, reinforcement learning, or knowledge editing training on the base large language model using the first labeled training set includes:

[0036]

[0037] in, Let L be the knowledge distillation loss function; α', β', and γ' are hyperparameters balancing the three losses, satisfying α'+β'+γ'=1; L_KL is the KL divergence between the teacher model and the student model on the first feature-power matrix dimension; L_reg is the graph regularization loss based on the first feature-power matrix. For full-parameter pre-training loss function, supervised fine-tuning loss function, reinforcement learning loss function, or knowledge editing training loss function;

[0038]

[0039] in, For the full-parameter pre-training loss function, Let x be a token sequence pair in the unsupervised patent corpus, where x represents the context, y represents the next token, and D represents the first labeled training set. The conditional probabilities given for the model;

[0040]

[0041] Where L_SFT is the supervised fine-tuning loss function; D_SFT represents a training sample in the first labeled training set; D_SFT represents the first labeled training set.

[0042]

[0043] Where L_RL is the reinforcement learning loss function, x is the cue word, D_RL is the cue set used for RLHF; y~π_θ is the student's large language model's sampled response on cue x; r(x,y) is the human feedback reward score; β_RL is the KL penalty coefficient in RL; KL(π_θ||π_ref) is the KL distance between the current policy and the frozen reference policy π_ref; Clip is the pruning operation to ensure that the policy update magnitude is controlled.

[0044]

[0045] Where L_KE is the loss function for editing training, (x_k, y_k) is a patent text to be edited, a target feature power matrix, and a summary pair; f_θ(x_k) is the output vector of the student model on the input x_k; ‖·‖ 2 The squared L2 norm measures the difference between the prediction and the target.

[0046]

[0047] Wherein, L_KL is the KL divergence between the teacher model and the student model on the first feature-power matrix dimension, which is used to convey the structured knowledge of the teacher model; ∈[0,1] is the prediction probability of the teacher's large language model for the i-th patent text in the first labeled training set in terms of feature information f and efficacy information e; ∈[0,1] is the predicted probability of the same position of the same patent text in the first labeled training set by the student large language model; ln represents the natural logarithm; N is the number of patent texts in the first labeled training set; C is the complete set of feature information of the patent texts in the first labeled training set; E is the complete set of efficacy information of the patent texts in the first labeled training set.

[0048]

[0049] Where L_reg is the graph regularization loss based on the first feature-power matrix, used to constrain the consistency of feature predictions for students at the same power nodes. It is the edge set of an undirected graph composed of functional nodes; (m,m')∈ Indicates adjacent functional nodes; | | represents the total number of edges in the graph; and Let m and m' represent the prediction probabilities of the student's large language model for the same feature information f at adjacent power nodes m and m', respectively; C is the complete set of feature information.

[0050] student_model=F(Transformer_θ(X;θ_s+θ_task[q]),L)

[0051] Where student_model is the first student large language model; Transformer_θ(·) represents the complete Transformer module with parameter θ; θ_s represents the parameters of the base large language model; θ_task[q] represents the task embedding vector indexed by patent quality level information q; X is the embedding matrix of the target training set; F is the optimization function; and L is the knowledge distillation loss function.

[0052] In one embodiment,

[0053] The teacher's large language model can be the FLAN-T5 model with an Encoder-Decoder architecture or the Qwen model with a Decoder-Only architecture, and the number of parameters can be 7B or 13B.

[0054] The base language model can be the FLAN-T5 model with an Encoder-Decoder architecture, or the Qwen model with a Decoder-Only architecture. The number of parameters can be any value smaller than that of the teacher language model.

[0055] In one embodiment, the construction of the feature-efficacy vector retrieval library and the feature-efficacy numerical retrieval library includes:

[0056] S701: Map the PMU items in the third feature-efficacy matrix to basic units or industry-standard units to obtain the fourth feature-efficacy matrix, and then obtain the updated feature-efficacy vector.

[0057] S702: The updated feature-efficacy vector <patent-patent abstract-fourth feature efficacy matrix> is vectorized and compressed using the base large language model to obtain a compressed feature-efficacy vector; a feature-efficacy vector retrieval library corresponding to the compressed feature-efficacy vector is constructed based on the K-means method; wherein, Euclidean distance is used in the iteration of K-means;

[0058] S703: Using the fourth feature-efficacy matrix as a sample, based on patent S and unit of measurement. The database is partitioned and constructed using an SQL database to retrieve feature-efficacy numerical values.

[0059] This application leverages the text generalization and downstream task potential of large language models to construct feature-efficacy matrices from a set of patent texts to be processed, eliminating the need for manual annotation and thus reducing costs. Specifically, the analysis and construction steps of the large language model yield at least one feature-efficacy matrix and a patent summary contained within the patent texts to be processed. Based on the feature-efficacy matrix, this application constructs a hybrid vector and efficacy interval retrieval method to accurately match complex patent query requirements, further improving the relevant patent retrieval performance. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] in:

[0062] Figure 1 This is a flowchart of a patent retrieval method based on a feature-efficacy matrix and a large language model in one embodiment;

[0063] Figure 2 This is a schematic diagram of the feature-efficacy matrix structure in one embodiment;

[0064] Figure 3 This is a structural block diagram of a computer device in one embodiment. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Patents are an important form of intellectual property protection, closely related to scientific, technological, and economic development. With the massive increase in the number of patents, patent analysis, retrieval, and evaluation have become exceptionally demanding tasks. The feature-efficacy matrix is ​​a core component of a patent, directly impacting the analysis, retrieval, and evaluation process. Analyzing, retrieving, and evaluating the feature-efficacy matrix can help quickly capture the core content of a patent, significantly improving patent analysis efficiency. Since the release of ChatGPT, Large Language Models (LLMs) have widely penetrated the public eye, becoming a core tool of general artificial intelligence, capable of efficiently handling diverse tasks such as text summarization, information extraction, and question answering. However, LLMs still face challenges in the vertical field of patent retrieval, analysis, and evaluation, including knowledge gaps and insufficient professional adaptation, resulting in a scarcity of existing research. This patent focuses on filling this gap, exploring the deep application and optimization paths of LLMs in patents. In existing technologies, pre-trained models can be used to analyze feature-efficacy matrices. However, this method requires manual annotation of a large number of sample texts, leading to increased costs. Furthermore, patent efficacy is multi-dimensionally coupled, and traditional retrieval methods using only a few keywords struggle to accurately match complex needs. Meanwhile, the lack of patent rating indicators that can comprehensively depict the multidimensional feature-efficacy relationship further restricts the effectiveness and quality of search results. A patent is a document issued by a government agency or a regional organization representing several countries based on an application. This document records the content of the invention or creation and creates a legal status for a certain period. A feature-efficacy matrix refers to the fact that a patent text, after analysis, can be decomposed into multiple interrelated feature components that possess efficacy. These feature components have clear hierarchical relationships, such as inventor information, applicant information, and agent information within the patent document.

[0067] To address the aforementioned technical problems, this application provides a patent retrieval method based on a feature-efficacy matrix and a large language model.

[0068] like Figure 1 As shown, in one embodiment, a patent retrieval method based on a feature-efficacy matrix and a large language model is provided, specifically including the following steps:

[0069] S1: Obtain the first set of patent texts to be processed, and divide the first set of patent texts to be processed into a target training set and a target verification set according to different levels of patent quality labels;

[0070] S2: Extract features from the target training set to obtain training text feature information, input the training text feature information into the teacher large language model, the teacher large language model outputs the first feature-efficacy matrix and the first patent summary for each patent text in the target training set; use the first feature-efficacy matrix and the first patent summary as the final label and fuse them with the corresponding target training set to obtain the first labeled training set;

[0071] S3: The first student large language model is obtained by performing full-parameter pre-training, supervised fine-tuning, reinforcement learning or knowledge editing training on the base large language model through the first labeled training set;

[0072] S4: Periodically extract a second set of patent texts to be processed from newly published patents. After being annotated by the teacher model in step S2, a second annotated training set is formed. When the distribution drift index D_KL of the second annotated training set exceeds a preset threshold δ, the first student large language model is updated by training the second annotated training set through step S3 to obtain the second student large language model. In step S3, an elastic weight consolidation or memory replay mechanism is introduced to prevent the second student large language model from experiencing performance regression on the first annotated training set during the training of the second annotated training set.

[0073] S5: Extract features from the target verification set to obtain verification text feature information, input the verification text feature information into the second student big language model, and the second student big language model outputs the second feature-efficacy matrix and the second patent summary for each patent text in the target verification set;

[0074] S6: Compare each item of the different second feature-efficacy matrices of the same patent text in the target validation set; if one or more items are unequal, call the teacher large language model, and each teacher large language model outputs a triplet and the confidence score of each item: sort the triplets in descending order of their corresponding confidence scores, select the triplet with the highest confidence score as the final authority value of the target validation set, and fuse the final authority value with the corresponding patent text to obtain the authoritative patent text, and write the authoritative patent text back to the target training set and the target validation set; specifically, if the same patent appears "<S,O,PV110%,PMU120%,PR110%> "and"<S,O,PV130%,PMU210%,PR240%> "If one or more of PV, PMU, or PR are not equal, then the teacher's large language model is invoked to select the PV with the highest confidence level (130%), PMU (120%), and PR (240%) using a majority voting mechanism."

[0075] S7: Determine the third feature-efficacy matrix of the second set of patent texts to be processed through steps S1 and S2. Multiple third feature-efficacy matrices constitute a retrieval library, which includes a feature-efficacy vector retrieval library and a feature-efficacy numerical retrieval library.

[0076] S8: Receive search keywords in natural language input, call the second student large language model to calculate the keyword feature-efficacy matrix of the search keywords, perform semantic matching between the keyword feature-efficacy matrix and the third feature-efficacy matrix in the feature-efficacy vector retrieval library, and output multiple third feature-efficacy matrices with a similarity greater than a first threshold to the keyword feature-efficacy matrix. Multiple patents corresponding to multiple third feature-efficacy matrices constitute a first candidate patent list; receive the specified feature name, numerical range, and unit of measurement, convert the numerical range into a numerical range under the standard unit system according to the unit conversion relationship, filter the range in the execution interval in the feature-efficacy numerical retrieval library corresponding to the third feature-efficacy matrix, and output a second candidate patent list that meets the conditions.

[0077] This application leverages the text generalization and downstream task potential of large language models to construct feature-efficacy matrices from a set of patent texts to be processed, eliminating the need for manual annotation and thus reducing costs. Specifically, the analysis and construction steps of the large language model yield at least one feature-efficacy matrix and a patent summary contained within the patent texts to be processed. Based on the feature-efficacy matrix, this application constructs a hybrid vector and efficacy interval retrieval method to accurately match complex patent query needs. Simultaneously, a multi-dimensional patent quality evaluation method based on the feature-efficacy matrix, comprehensively evaluating from dimensions such as novelty, efficacy, and reputation, is applied to patent retrieval, further improving the analysis and retrieval effectiveness of relevant patents.

[0078] In one embodiment, the process of obtaining the graded patent quality label is as follows:

[0079] (1)

[0080] (2)

[0081] (3)

[0082] (4)

[0083] in, c_i represents the citation score for each patent text in the first set of patent texts to be processed; c_i represents the number of patent citations for each patent text. and These represent the average and standard deviation of patent citations for the first set of patent texts to be processed; recency_score is the timeliness score for each patent text; time_now is the evaluation benchmark date for each patent text; time_grant is the grant announcement date for each patent text; authorize_score is the grant rate score for each patent text; g_i is the number of patents granted for each patent text; and g_total is the number of patents granted in the same patent family within the first set of patent texts to be processed. A graded patent quality label for each patent text; , , All are fusion coefficients.

[0084] In one embodiment, dividing the first set of patent texts to be processed into a target training set and a target verification set according to different levels of patent quality labels includes:

[0085] S101: Multiple graded patent quality labels constitute a graded patent quality label set;

[0086] S102: Discretize multiple levels of patent quality labels into several level label intervals according to manually defined threshold intervals, denoted as Li (i=1,...,k);

[0087] S103: Calculate the proportion of each level label interval in the patent quality label set of the stated level, denoted as pi (i=1,...,k);

[0088] S104: For each level label interval Li (i=1,...,k), determine the first number of patent texts corresponding to it in the first set of patent texts to be processed. Randomly select a% of the patent texts from the first number of patent texts to form a target verification set. Randomly select b% of the patent texts from 1-a% of the patent texts in the first set of patent texts to be processed to form the target training set, where a and b are agreed constants.

[0089] In one embodiment, dividing the first set of patent texts to be processed into a target training set and a target verification set according to different levels of patent quality labels includes:

[0090] S105: Discretize multiple levels of patent quality labels into several level label intervals according to manually defined threshold intervals, denoted as Li (i=1,...,k);

[0091] S106: Each level label interval corresponds to a patent text sample set in the first set of patent texts to be processed. The patent text sample set with the smallest proportion of patent texts in the first set of patent texts to be processed is the rare level sample. Set an amplification factor ri, and extract ri*a% of patent texts from the rare level sample and put them into the first training set; where, if ri*α%>1, then ri*α% is set to 1.

[0092] S107: For the patent text sample set corresponding to each level label interval except for rare level samples, set a reduction coefficient ti, and extract ti*a% patent texts from each patent text sample set to form a second training set. The first training set and the second training set constitute the target training set.

[0093] S108: Randomly select b% patent texts from the 1-ri*a%-ti*a% patent texts in the first set of patent texts to be processed to form a target verification set. The target verification set is not scaled up or down synchronously with the target training set.

[0094] In one embodiment, the first feature-efficacy matrix is ​​a graphical data structure used to describe the patent itself, patent features and efficacy, and their interrelationships. In the first feature-efficacy matrix, each node represents a patent or a patent feature, or a patent efficacy, and edges represent the relationships between nodes. The first feature-efficacy matrix = [G, V], where G is the set of nodes: G = [patent, patent features, patent efficacy]; V is the set of edges: the set of edges formed by every two nodes; the first feature-efficacy matrix for a patent consists of one or more attribute quintuples, each quintuple including a main entity, a subordinate entity of the main entity, and the corresponding attribute of the subordinate entity, i.e., <main entity (automobile engine patent), subordinate entity (engine speed), attribute value (6000), attribute unit (r / min), attribute relationship (equality)> (<Subject,Object,Predicate-Value, Predicate-Measurement Unit,Predicate-Relation> abbreviation<S,O,PV,PMU,PR> For example, such as Figure 2 As shown, Figure 2 A schematic diagram illustrating the structure of a feature-effect provided in an embodiment of this application is shown. (Regarding...) Figure 2There are 22 nodes in total: S1, S2, O11, O12, O21, O22, PV11, PV12, PV21, PV22, PV23, PMU11, PMU12, PMU21, PMU22, PMU23, PR11, PR12, PR21, PR22, PR23. There are five knowledge points, namely five feature-efficacy quintuples, which are:<S1,O11,PV11,PMU11,PR11> ,<S1,O12,PV12,PMU12,PR12> ,<S2,O21,PV21,PMU21,PR21> ,<S2,O22,PV22,PMU22,PR22> as well as<S2,O23,PV23,PMU23,PR23> The first feature-efficacy matrix includes feature information and efficacy information. The first set of patent texts to be processed may contain the first feature-efficacy matrix itself, synonyms and subset elements of the feature information, and equivalent values ​​and calculation formulas of the efficacy information, etc., without limitation. For example, when the feature information is "car engine", the elements under its subset include "cylinder", "piston", etc.; when the efficacy information is "100 revolutions per second", the equivalent value can be "6000 r / min". In this embodiment, the first set of patent texts to be processed can obtain the original full text of the patent through web crawling or API interface, which includes at least the specification, claims and abstract; the teacher's large language model can be called in parallel according to computing power, and the first annotation training set can be trained through majority voting or confidence weighting to ensure annotation robustness.

[0095] The final label is obtained as follows: the first feature-efficacy matrix with the highest number of votes and the first patent summary with the highest number of votes are selected as the final label; if there is a tie, random and uniform sampling is performed.

[0096] (5)

[0097] in, Let i be the maximum number of votes for the first feature-efficacy matrix or the maximum number of votes for the first patent summary, where i is the index of the teacher's large language model. Let L be the i-th teacher's large language model, and L be the set of teacher's large language models. For the target training set, argmax is the first feature-efficacy matrix or first patent summary output by the i-th teacher's large language model, and argmax is the majority vote operation.

[0098] The final label can also be obtained by selecting the first feature-efficacy matrix with the highest weighted votes and the first patent summary with the highest weighted votes as the final label;

[0099] (6)

[0100] in, The maximum weighted vote count for the first feature-efficacy matrix or the maximum weighted vote count for the first patent summary, where i is the teacher's large language model index. Let L be the i-th teacher's large language model, and L be the set of teacher's large language models. For the target training set, The first feature-efficacy matrix and the first patent summary are output by the i-th teacher's large language model. argmax represents the confidence score output by the i-th teacher's large language model, and argmax represents the majority vote operation.

[0101] In one embodiment, the step of obtaining the first student large language model by performing full-parameter pre-training, supervised fine-tuning, reinforcement learning, or knowledge editing training on the base large language model using the first labeled training set includes:

[0102] (7)

[0103] in, Let L be the knowledge distillation loss function; α', β', and γ' are hyperparameters balancing the three losses, satisfying α'+β'+γ'=1; L_KL is the KL divergence between the teacher model and the student model on the first feature-power matrix dimension; L_reg is the graph regularization loss based on the first feature-power matrix. For full-parameter pre-training loss function, supervised fine-tuning loss function, reinforcement learning loss function, or knowledge editing training loss function;

[0104] (8)

[0105] in, For the full-parameter pre-training loss function, Let x be a token sequence pair in the unsupervised patent corpus, where x represents the context, y represents the next token, and D represents the first labeled training set. The conditional probabilities given for the model;

[0106] (9)

[0107] Where L_SFT is the supervised fine-tuning loss function; is a training sample in the first labeled training set; D_SFT is the first labeled training set.

[0108] (10)

[0109] Where L_RL is the reinforcement learning loss function, x is the cue word, D_RL is the cue set used for RLHF; y~π_θ is the student's large language model's response sampled on cue x; r(x,y) is the human feedback reward score; β_RL is the KL penalty coefficient in RL (different from distillation β); KL(π_θ||π_ref) is the KL distance between the current policy and the frozen reference policy π_ref; Clip is the pruning operation to ensure that the policy update magnitude is controlled.

[0110] (11)

[0111] Where L_KE is the loss function for editing training, (x_k, y_k) is a patent text to be edited, a target feature power matrix, and a summary pair; f_θ(x_k) is the output vector of the student model on the input x_k; ‖·‖ 2 The squared L2 norm measures the difference between the prediction and the target.

[0112] (12)

[0113] Wherein, L_KL is the KL divergence between the teacher model and the student model on the first feature-power matrix dimension, which is used to convey the structured knowledge of the teacher model; ∈[0,1] is the prediction probability of the teacher's large language model for the i-th patent text in the first labeled training set in terms of feature information f and efficacy information e; ∈[0,1] is the predicted probability of the student large language model for the same position in the same patent text in the first labeled training set. ln represents the natural logarithm; N is the number of patent texts in the first labeled training set; C is the complete set of feature information of the patent texts in the first labeled training set; E is the complete set of efficacy information of the patent texts in the first labeled training set;

[0114] (13)

[0115] Where L_reg is the graph regularization loss based on the first feature-power matrix, used to constrain the consistency of feature predictions for students at the same power nodes. It is the edge set of an undirected graph composed of functional nodes; (m,m')∈ Indicates adjacent functional nodes; | | represents the total number of edges in the graph; and Let m and m' represent the prediction probabilities of the student's large language model for the same feature information f at adjacent power nodes m and m', respectively; C is the complete set of feature information.

[0116] student_model=F(Transformer_θ(X;θ_s+θ_task[q]),L) (14)

[0117] Where student_model is the first student large language model; Transformer_θ(·) represents the complete Transformer module with parameter θ; θ_s represents the parameters of the base large language model; θ_task[q] represents the task embedding vector indexed by patent quality level information q; X is the embedding matrix of the target training set; F is the optimization function; and L is the knowledge distillation loss function.

[0118] In one embodiment, the teacher large language model can be the FLAN-T5 model with an Encoder-Decoder architecture or the Qwen model with a Decoder-Only architecture, and the number of parameters can be 7B or 13B; the base large language model can be the FLAN-T5 model with an Encoder-Decoder architecture or the Qwen model with a Decoder-Only architecture, and the number of parameters can be any value smaller than that of the teacher large language model; wherein, the number of parameters "B" refers to "hundred million", such as 7B is a large language model with 700 million parameters.

[0119] In one embodiment, the construction of the feature-efficacy vector retrieval library and the feature-efficacy numerical retrieval library includes:

[0120] S701: Regarding the third feature-efficacy matrix<S,O,PV,PMU,PR> The PMU items are uniformly mapped to basic units or industry-standard units to obtain the fourth feature-efficacy matrix, and then the updated feature-efficacy vector <Patent-Patent Abstract-Third Feature Efficacy Matrix> is obtained.

[0121] S702: The updated feature-efficacy vector <patent-patent abstract-fourth feature efficacy matrix> is vectorized and compressed using the base large language model to obtain a compressed feature-efficacy vector; a feature-efficacy vector retrieval library corresponding to the compressed feature-efficacy vector is constructed based on the K-means method; wherein, Euclidean distance is used in the iteration of K-means;

[0122] S703: Using the fourth feature-efficacy matrix as a sample, based on patent S and unit of measurement. The database is partitioned and constructed using an SQL database to retrieve feature-efficacy numerical values.

[0123] This application leverages the text generalization and downstream task potential of large language models to construct feature-efficacy matrices from a set of patent texts to be processed, eliminating the need for manual annotation and thus reducing costs. Specifically, the analysis and construction steps of the large language model yield at least one feature-efficacy matrix and a patent summary contained within the patent texts to be processed. Based on the feature-efficacy matrix, this application constructs a hybrid vector and efficacy interval retrieval method to accurately match complex patent query needs. Simultaneously, a multi-dimensional patent quality evaluation method based on the feature-efficacy matrix, comprehensively evaluating from dimensions such as novelty, efficacy, and reputation, is applied to patent retrieval, further improving the analysis and retrieval effectiveness of relevant patents.

[0124] Figure 3 An internal structural diagram of a computer device in one embodiment is shown. This computer device can specifically be a terminal or a server. Figure 3 As shown, the computer device includes a processor, memory, and network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a patent search method based on a feature-efficacy matrix and a large language model. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement the patent search method based on a feature-efficacy matrix and a large language model. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0125] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0126] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0127] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A patent retrieval method based on feature-efficacy matrix and large language model, characterized in that, The method includes: S1: Obtain the first set of patent texts to be processed, and divide the first set of patent texts to be processed into a target training set and a target verification set according to different levels of patent quality labels; S2: Extract features from the target training set to obtain training text feature information, input the training text feature information into the teacher large language model, the teacher large language model outputs the first feature-efficacy matrix and the first patent summary for each patent text in the target training set; use the first feature-efficacy matrix and the first patent summary as the final label and fuse them with the corresponding target training set to obtain the first labeled training set; S3: The first student large language model is obtained by performing full-parameter pre-training, supervised fine-tuning, reinforcement learning or knowledge editing training on the base large language model through the first labeled training set; S4: Periodically extract a second set of patent texts to be processed from newly published patents. After being annotated by the teacher's large language model in step S2, a second annotated training set is formed. When the distribution drift index of the second annotated training set exceeds a preset threshold δ, the first student large language model is updated by training the second annotated training set through step S3 to obtain the second student large language model. In step S3, an elastic weight consolidation or memory replay mechanism is introduced to prevent the second student large language model from experiencing performance regression on the first annotated training set during the training of the second annotated training set. S5: Extract features from the target verification set to obtain verification text feature information, input the verification text feature information into the second student big language model, and the second student big language model outputs the second feature-efficacy matrix and the second patent summary for each patent text in the target verification set; S6: Compare each item of the different second feature-efficacy matrices of the same patent text in the target validation set; if one or more items are not equal, call the teacher large language model, and each teacher large language model outputs a triplet and the confidence of each item: sort the triplets in descending order of their corresponding confidence, select the triplet with the highest confidence as the final authority value of the target validation set, and fuse the final authority value with the corresponding patent text to obtain the authoritative patent text, and write the authoritative patent text back to the target training set and the target validation set; The step of dividing the first set of patent texts to be processed into a target training set and a target validation set according to different levels of patent quality labels includes: discretizing multiple levels of patent quality labels into several level label intervals according to manually defined threshold intervals; each level label interval corresponds to a set of patent text samples in the first set of patent texts to be processed, and the set of patent text samples with the smallest proportion of the number of patent texts in the first set of patent texts to be processed is a rare level sample, and an amplification factor is set. ri Extract from rare-level samples ri* a% Each patent text is placed into the first training set; where, if ri*α%>1 Then take ri*α% Set the value to 1; for the patent text sample set corresponding to each level label interval except for rare level samples, set a reduction factor. ti Extract from each patent text sample set ti*a% The first set of patent texts constitutes a second training set, and the first and second training sets together constitute a target training set; from the first set of patent texts to be processed... 1-ri*a%-ti*a% Randomly selected from patent texts b% Each patent text constitutes a target validation set, which is not scaled up or down synchronously with the target training set. S7: Determine the third feature-efficacy matrix of the second set of patent texts to be processed through steps S1 and S2. Multiple third feature-efficacy matrices constitute a retrieval library, which includes a feature-efficacy vector retrieval library and a feature-efficacy numerical retrieval library. S8: Receive search keywords in natural language input, call the second student large language model to calculate the keyword feature-efficacy matrix of the search keywords, perform semantic matching between the keyword feature-efficacy matrix and the third feature-efficacy matrix in the feature-efficacy vector retrieval library, and output multiple third feature-efficacy matrices with a similarity greater than a first threshold to the keyword feature-efficacy matrix. Multiple patents corresponding to multiple third feature-efficacy matrices constitute a first candidate patent list. Receive the specified feature name, numerical range, and unit of measurement, convert the numerical range to a numerical range under a standard unit system according to unit conversion relationships, filter the range in the execution interval of the feature-efficacy numerical retrieval library corresponding to the third feature-efficacy matrix, and output a second candidate patent list that meets the conditions.

2. The patent retrieval method based on feature-efficacy matrix and large language model according to claim 1, characterized in that, The process for obtaining the graded patent quality label is as follows: in, The citation score is assigned to each patent text in the first set of patent texts to be processed. c_i The number of patent citations for each patent text; and These are the average and standard deviation of patent citations for the first set of patent texts to be processed, respectively. recency_score A timeliness score is given for each patent text. time_now The evaluation benchmark date for each patent text; time_grant The date of authorization announcement for each patent text; authorize_score Grant rate score for each patent text; g_i The number of patents granted for each patent text; g_total The number of patents granted in the same patent family within the first set of patent texts to be processed; A graded patent quality label for each patent text; , , All are fusion coefficients.

3. The patent retrieval method based on feature-efficacy matrix and large language model according to claim 2, characterized in that, The step of dividing the first set of patent texts to be processed into a target training set and a target verification set according to different levels of patent quality labels includes: S101: Multiple graded patent quality labels constitute a graded patent quality label set; S102: Discretize multiple levels of patent quality labels into several level label ranges according to manually defined threshold ranges; S103: Calculate the proportion of each grade label range in the patent quality label set of the stated grade; S104: For each level label interval, determine the first number of patent texts corresponding to it in the first set of patent texts to be processed, and randomly select from the first number of patent texts. a% The target verification set consists of 1 patent texts, starting from the first set of patent texts to be processed. -a% Randomly selected from patent texts b% The target training set consists of several patent texts, wherein... a , b It is a conventional constant.

4. The patent retrieval method based on feature-efficacy matrix and large language model according to claim 1, characterized in that, The final labels are obtained as follows: the first feature-efficacy matrix with the highest number of votes and the first patent summary with the highest number of votes are selected as the final labels; if there is a tie, random and uniform sampling is performed. in, The maximum number of votes for the first feature-efficacy matrix or the maximum number of votes for the first patent summary. i This refers to the sequence number of the teacher's large language model. For the first i A large language model for teachers, L For a large set of language models for teachers, For the target training set, For the first i The first feature-efficacy matrix or first patent summary output by the teacher's large language model argmax This is an operation to obtain a majority vote.

5. The patent retrieval method based on feature-efficacy matrix and large language model according to claim 1, characterized in that, The final label is obtained by selecting the first feature-efficacy matrix with the highest weighted votes and the first patent summary with the highest weighted votes as the final label; in, The maximum weighted vote in the first feature-efficacy matrix or the maximum weighted vote in the first patent summary. i This is the serial number of the teacher's large language model. For the first i A large language model for teachers, L For a large set of language models for teachers, For the target training set, For the first i The first feature-efficacy matrix and first patent summary of the output of the teacher's large language model For the first i The confidence level of the output of a large language model for each teacher. argmax This is an operation to obtain a majority vote.

6. The patent retrieval method based on feature-efficacy matrix and large language model according to claim 1, characterized in that, The process of obtaining the first student large language model by performing full-parameter pre-training, supervised fine-tuning, reinforcement learning, or knowledge editing training on the base large language model using the first labeled training set includes: in, The knowledge distillation loss function; α', β', γ' It is a hyperparameter that balances the three losses and satisfies α'+β'+γ'= 1; L_KL For the teacher model and student model on the first feature-power matrix dimension KL Divergence; L_reg The graph regularization loss is based on the first feature-power matrix; For full-parameter pre-training loss function, supervised fine-tuning loss function, reinforcement learning loss function, or knowledge editing training loss function; in, For the full-parameter pre-training loss function, One of the unsupervised patent corpora token Sequence pairs, x For context, y For the next token ; D This is the first labeled training set; The conditional probabilities given for the model; in, L_SFT For supervised fine-tuning of the loss function; This is a training sample in the first labeled training set; D_SFT This is the first labeled training set; in L_RL To reinforce the learning loss function, x As a prompt word, D_RL For the cue set used in RLHF; y~π_θ For students' large language model in prompting x Upsampled answers; r(x,y) Provide reward scores to humans; β_RL for RL middle KL Penalty coefficient KL (π_θ||π_ref) For the current strategy and freeze reference strategy π_ref of KL distance; Clip To ensure the scope of strategy updates is controlled during the pruning operation; in, L_KE To identify the training loss function, (x_k,y_k) For a patent text that needs editing, there is a target feature efficacy matrix and an abstract pair. f_θ(x_k) For the student model in input x_k The output vector on; ‖·‖2 for L2 Norm squared measures the difference between the prediction and the target. in, L_KL For the teacher model and student model on the first feature-power matrix dimension KL Divergence is used to convey structured knowledge from the teacher model; ∈[0,1] It is the teacher's large language model applied to the first labeled training set. i The patent text contains feature information f Efficacy information e The predicted probability; ∈[0,1] is the predicted probability of the student's large language model for the same position in the same patent text in the first labeled training set; ln Represents the natural logarithm; N The number of patent texts in the first labeled training set; C This is the complete set of feature information for the patent texts in the first labeled training set; E This is the complete set of efficacy information for the patent texts in the first labeled training set; in, L_reg The graph regularization loss is based on the first feature-power matrix and is used to constrain the consistency of feature predictions for students at nodes with the same power. It is the edge set of an undirected graph composed of functional nodes; (m,m') ∈ Indicates adjacent functional nodes; | | represents the total number of edges in the graph; and These represent the student's large language model at adjacent efficiency nodes. m , m' The same feature information f The predicted probability; C It is the complete set of feature information; student_model=F(Transformer_θ(X;θ_s+θ_task[q]),L) in, student_model For the first student's large language model; Transformer_θ(·) Represents the complete expression with parameter θ Transforme r module; θ_s The parameters represent the base language model; θ_task[q] This indicates information derived from patent quality levels. q The indexed task embedding vector; X The embedding matrix of the target training set; F To optimize the function, L Let be the knowledge distillation loss function.

7. The patent retrieval method based on feature-efficacy matrix and large language model according to claim 1, characterized in that, The teacher's large language model is Encoder-Decoder Architecture FLAN-T5 Model, Decoder-Only of Qwen The model has 7B or 13B parameters; The base large language model is Encoder-Decoder Architecture FLAN-T5 Model, also Decoder-Only of Qwen The model has a number of parameters that are any values ​​smaller than those in the teacher's large language model.

8. The patent retrieval method based on feature-efficacy matrix and large language model according to claim 1, characterized in that, The construction of the feature-efficacy vector retrieval library and the feature-efficacy numerical retrieval library includes: S701: Regarding the third feature-efficacy matrix... PMU The items are uniformly mapped to basic units or industry-standard units to obtain the fourth feature-efficacy matrix, and then the updated feature-efficacy vector is obtained. S702: The updated feature-efficacy vector <patent-patent abstract-fourth feature-efficacy matrix> is vectorized and compressed using the aforementioned base large language model to obtain a compressed feature-efficacy vector; based on... K-means The method constructs a feature-efficacy vector retrieval library corresponding to the compressed feature-efficacy vector; wherein, in K-means The Euclidean distance is used in the iteration; S703: Using the fourth feature-efficacy matrix as a sample, based on patent S and unit of measurement. The database is partitioned and constructed using an SQL database to retrieve feature-efficacy numerical values.

Images