Patent value information intelligent extraction method and device based on large language model

Abstract

The invention provides a patent value information intelligent extraction method and device based on a large language model, and the method comprises the steps: obtaining and preprocessing a target patent text, and recognizing and extracting key information in the target patent text based on a rule positioning engine; extracting a function triple from the key information based on a large language model and a Prompt template; extracting a feature vector of the function triad, and obtaining a value feature alignment vector through a retrieval enhancement generation technology; according to the feature vector of the function triple and the value feature alignment vector, mapping the function triple to a preset value information index system based on a pre-trained mapping model, and obtaining an overall value information vector of the target patent text on each index dimension; and in response to an analysis intention input by a user, driving a report generation engine to generate a structured evaluation report based on the total value information vector of the target patent text on each index dimension.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and intelligent patent analysis technology, specifically, to a method and apparatus for intelligent extraction of patent value information based on a large language model. Background Technology

[0002] The value analysis of patents, especially their functions and implementation methods, often relies on expert experience. Manually extracting multi-dimensional patent value information is time-consuming and struggles to handle massive amounts of patent data. Traditional manual reading, indexing, and evaluation methods can no longer meet the current business needs of industrial innovation and development for high-efficiency, high-precision, and high-quality processing of massive patent documents in areas such as proof of concept, technology transfer, patent operation, technology investment and financing, and infringement risk screening.

[0003] The definition and evaluation of the functional value of patents, especially the evaluation of scientific value, technological value, economic value, social value, cultural value, and potential for transformation and promotion that are consistent with standards such as the "Standard for Evaluation of Scientific and Technological Achievements" (GB / T 44731-2024), mainly rely on the experience of experts and the data and materials provided by the rights holders for qualitative evaluation. There is a lack of intelligent analysis and evaluation based on the patent text itself and the main characteristics of high-value patents, as well as a systematic evaluation framework for the potential social and cultural impacts of the technology application and whether it conforms to the ethical orientation of "science for good". It is impossible to fully utilize artificial intelligence technology to improve the efficiency and accuracy of patent functional value evaluation, thereby promoting the rapid transformation and extended application of patent technology achievements.

[0004] Traditional patent valuation techniques have primarily developed along two paths. Early methods based on keyword matching and regular expressions are essentially shallow text processing. This approach relies on predefined terminology dictionaries and fixed patterns, performing only superficial string matching and completely failing to grasp the inherent logic of the technical description and its potentially broad socio-cultural implications. For example, traditional methods can identify "delay unit" and "trigger" as keywords, but they ignore the functional significance implied by the "connection" between them, failing to understand the role of this connection within the overall technical solution, and unable to analyze the potential impact of the technology on industry efficiency, social security, or cultural life. Furthermore, this method has weak generalization ability; if the patent author changes the sentence structure or uses synonyms to express the same technical concept, the system may fail. Its fundamental limitation lies in its inability to construct a structured mapping between "technical function," "specific implementation form," and "multi-dimensional value characteristics." Its evaluation can only remain at the frequency statistics of explicit terms, lacking in-depth reasoning and value association analysis.

[0005] The subsequent emergence of deep learning models, such as CNN and LSTM, brought data-driven representation learning capabilities to patent text analysis. However, their essence remains limited by the structured assumptions of the models and the static nature of the training data. These models typically capture semantics through local context windows. While they can identify co-occurrence relationships between technical terms, they struggle to understand the complex technical logic chains spanning the entire patent document and the positioning of technical solutions within a broader socio-cultural context—for example, how a "delay unit" influences the temporal characteristics of a "trigger" through specific connections. This functional dependency is often obscured by the model's local attention. More importantly, the knowledge of deep learning models is embedded in the parameters trained on-site. They cannot actively acquire and understand the latest technological advancements, changes in legal precedents, or market dynamics during evaluation. This leads to a disconnect between their evaluation conclusions and the real-time evolution of industry realities and social expectations, resulting in a lack of timeliness, scenario adaptability, and comprehensive value.

[0006] Therefore, there is an urgent need for a patent multidimensional value assessment method and device that can deeply integrate the deep reasoning capabilities of large language models, real-time knowledge retrieval capabilities, and structured value modeling capabilities, in order to support high-quality, high-efficiency, and large-scale patent intelligence analysis and serve responsible technology innovation decision-making and achievement transformation. Summary of the Invention

[0007] Therefore, it is necessary to provide a method and apparatus for intelligent extraction of patent value information based on a large language model to address the above-mentioned technical problems.

[0008] To achieve the above objectives, the first aspect of the present invention provides an intelligent method for extracting patent value information based on a large language model, comprising the following steps: The target patent text is acquired and preprocessed. Based on the rule-based localization engine, key information in the target patent text is identified and extracted. The key information includes technical field, function, application scenario, pain point solved, beneficial effect, technical solution, working principle, functional carrier and its value characteristics. Based on the large language model and the Prompt template, functional triples are extracted from the key information. The functional triples are: function - function implementation form - value feature. Extract the feature vectors of functional triples, and obtain the value feature alignment vectors through retrieval enhancement generation techniques; Based on the feature vector and value feature alignment vector of the functional triple, the functional triple is mapped to the preset value information index system based on the pre-trained mapping model to obtain the overall value information vector of the target patent text in each index dimension. The mapping model is as follows: ; V iIndicates the first i The overall value information vector of each indicator dimension i ∈[1, N ], N This represents the total number of indicators in the value information indicator system. ΔY i Indicates the relationship with the first i The value feature alignment vector of each indicator dimension, π i Aligning the value feature vector with the first i The adjustment coefficient for each indicator dimension, with a value range of [0,1]; T j Indicates the first j The feature vector of a functional triple. n The number of functional triples; θ ij For the first j The feature vector of the t function triplet pairs the t function triplet. i The weight coefficients for each indicator dimension satisfy... ; In response to the user's analytical intent and based on the overall value information vector of the target patent text across various indicator dimensions, the report generation engine generates a structured evaluation report. The structured evaluation report includes the value mapping information of the target patent document across at least one indicator dimension, as well as the evaluation results across at least one indicator dimension.

[0009] This solution features systematic innovation at the technical level, proposing a hybrid extraction framework that combines a rule-based localization engine, a large language model, a structured Prompt template, and a mapping model. The framework first uses a rule engine to quickly and accurately locate high-value paragraphs in the patent text. Then, it invokes the large language model, combined with a Prompt template customized for the patent domain, to perform deep semantic analysis and information extraction on the high-value paragraphs, ultimately outputting a structured function-function implementation-value feature triplet. This design effectively integrates the localization accuracy of rule-based methods with the deep semantic understanding capabilities of the large language model, thus overcoming the problems of weak generalization in traditional methods and insufficient accuracy of general-purpose large language models in specialized domains. Furthermore, by extracting the feature vectors of the function triplets and using the retrieval enhancement generation technology of the large language model to obtain value feature alignment vectors, the value mapping and evaluation process is no longer limited to the patent text itself, but can integrate external academic achievements, market dynamics, and policy information, significantly improving the objectivity and persuasiveness of the evaluation results. Finally, through the preset mapping model, the system realizes the structured, end-to-end automatic conversion from technical detail descriptions to multi-dimensional value indicators, completing the efficient mapping of technical value to industrial and economic value.

[0010] To achieve the above objectives, a second aspect of the present invention provides an intelligent patent value information extraction device based on a large language model, comprising: The text preprocessing module is used to acquire and clean the original patent text and generate a structured paragraph sequence; The rule-based localization engine module is used to identify and extract key information from patent texts. The key information includes technical field, function, application scenario, pain point solved, beneficial effect, technical solution, working principle, functional carrier and its value characteristics. The functional triple generation module is used to extract information from the key information and generate functional triples using a pre-trained large language model and a Prompt template. The functional triples are function-function implementation form-value feature. The feature enhancement module is used to extract the feature vectors of functional triples and obtain the value feature alignment vectors through retrieval enhancement generation techniques. The value feature mapping module is used to map the functional triples to a preset value information index system based on the feature vectors of the functional triples and the value feature alignment vectors, using a pre-trained mapping model, to obtain the overall value information vector of the target patent text in each index dimension. The mapping model is as follows: ; V i Indicates the first i The overall value information vector of each indicator dimension i ∈[1, N ], N This represents the total number of indicators in the value information indicator system. ΔY i Indicates the relationship with the first i Alignment vector of value features across multiple indicator dimensions π i Aligning the value feature vector with the first i The adjustment coefficient for each indicator dimension, with a value range of [0,1]; T j Indicates the first j The feature vector of a functional triple. n The number of functional triples; θ ij For the first j The feature vector of the t function triplet pairs the t function triplet. i The weight coefficients for each indicator dimension satisfy... ; A multi-format output module is used to respond to the user's input analysis intent and drive the report generation engine to generate a structured evaluation report based on the overall value information vector of the target patent text in each indicator dimension. The structured evaluation report includes the value mapping information of the target patent document in at least one indicator dimension, as well as the evaluation results in at least one indicator dimension.

[0011] To achieve the above objectives, a third aspect of the present invention provides a computer device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory is used to store computer programs; and the processor, when executing the program stored in the memory, implements the intelligent extraction method for patent value information as described in the first aspect.

[0012] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the intelligent extraction method for patent value information as described in the first aspect.

[0013] To achieve the above objectives, the fifth aspect of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the intelligent extraction method for patent value information as described in the first aspect.

[0014] The beneficial effects of this invention are as follows: The core innovation of this invention lies in proposing a hybrid information extraction framework that combines a rule-based localization engine, a large language model, and a structured Prompt template. This framework first uses a rule engine to quickly locate high-value paragraphs in patent texts, such as sections on the technical field and invention content, as well as extended phrases like "characteristics" and "more specifically." Then, it invokes a large language model, combined with a structured Prompt template customized for the patent field, to perform deep semantic analysis on the identified content, ultimately automatically generating the function, its implementation form, and its value features. This design combines the precise localization capabilities of rule-based methods with the deep semantic understanding advantages of large language models, effectively solving the problems of weak generalization in traditional methods and insufficient specialization of general-purpose large language models.

[0015] To achieve a systematic mapping from technical descriptions to value indicators, this invention uses a functional triple—function-functional implementation form-value feature—as a standardized intermediate representation, and constructs a dynamic weight mapping model based on this. This model quantifies the correlation strength between text content and each indicator dimension through an attention mechanism, and introduces a residual learning strategy to dynamically correct evaluation biases, thereby improving the objectivity and accuracy of value assessment.

[0016] Furthermore, a retrieval enhancement generation technology based on a large language model is introduced to retrieve, in real time, the most relevant and novel comparative materials and background knowledge related to the current patent technology from multiple sources such as massive patent databases, scientific and technological literature, legal documents, and market reports. Then, based on this high-quality and highly relevant external knowledge, the large language model performs in-depth integration analysis and reasoning, so that the final value mapping and value assessment are not based on the patent text in isolation, but combine external academic, market, and policy evidence, making the assessment results more objective and persuasive.

[0017] Furthermore, this invention adopts the six-dimensional indicator system of the national standard "Evaluation Specification for Scientific and Technological Achievements" (GB / T 44731-2024), innovatively incorporating social and cultural values ​​into the automated evaluation system. It assesses the potential social and cultural impacts of patented technologies and whether they comply with ethical norms, breaking through the limitations of the technological and economic dimensions in traditional evaluations. Through standardized interpretation of mainstream cultural values ​​and matching and semantic association with policy documents, it achieves objective quantification of soft indicators, expanding the scope and decision support capabilities of patent valuation. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the intelligent extraction method for patent value information based on a large language model according to the present invention. Figure 2 This is a schematic diagram of the intelligent extraction device for patent value information based on a large language model according to the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be further described in detail below through specific embodiments.

[0020] Example 1 This embodiment provides an intelligent method for extracting patent value information based on a large language model, including the following steps: Step S1: Obtain the target patent text and perform preprocessing.

[0021] Manually upload or obtain the text file of the target patent automatically or manually through public databases, the official website of the Patent Office, or interfaces provided by partners to ensure the integrity and accuracy of the text.

[0022] It is understandable that after obtaining the text file of the target patent, an automated quality verification module can be embedded to ensure the quality of the input data from the source.

[0023] Specifically, the quality verification module first verifies the integrity of the file. After the integrity verification passes, it checks the readability of the text using a pre-trained model, identifies possible OCR errors and quality issues (such as missing or blurry pages), and prompts the user to re-upload.

[0024] Specifically, the quality verification indicators are based on the relevant provisions of the Patent Law of the People's Republic of China, the Implementing Regulations of the Patent Law of the People's Republic of China, and the Guidelines for Patent Examination.

[0025] Specifically, the training data for the pre-trained model consists of patent documents that won the China Patent Award from 2015 to 2024 (hereinafter the same; unless otherwise specified, all model training data referred to in this document are of this type). The range of this data will be dynamically updated as new patent awards are announced. For example, after the 2025 awards are announced, the data will be expanded to cover 2025.

[0026] After the quality check is passed, paragraph segmentation and language recognition and conversion are performed in sequence.

[0027] Paragraph segmentation: Using natural language processing technology, the patent text is automatically segmented according to paragraph markers such as periods and line breaks, generating an ordered sequence of natural paragraphs. Each paragraph is bound to the original text line number, which facilitates reverse location and information tracing in subsequent processing.

[0028] It's understandable that common, simple rule-based segmentation based on punctuation and line breaks is used: by recognizing sentence-ending markers such as periods, question marks, and exclamation marks, and combining them with formatting features like line breaks and indentation, text is divided into paragraphs. This method is effective when processing well-formatted documents, but it doesn't work well for patent texts with complex structures and lengthy sentences. Patent claims typically have one sentence per line break, but an entire claim is a complete semantic unit. Simply dividing by line breaks would disrupt its integrity, causing large language models to fail to understand the reference relationships in phrases like "according to claim 1...". Conversely, in the "detailed implementation" section, a single paragraph may be extremely long, containing multiple technical points; simple segmentation would lose fine-grained structure.

[0029] To address the aforementioned issues, this embodiment employs a hierarchical paragraph segmentation method that integrates the patented chapter structure with a pre-trained model: The first layer: legally mandated chapter division. Based on typical keywords of patent chapters, such as "technical field" and "claims," ​​the entire patent document is divided into legally mandated major chapters such as "Abstract," "Claims," ​​and "Specification."

[0030] The second layer: paragraph division within chapters. Specifically, this embodiment uses a large language model fine-tuned with the patented text. Chapter-level text is input into the model, and semantically coherent paragraph boundaries are identified through sequence labeling. The model predicts labels for each sentence boundary (B-PARA: paragraph start; I-PARA: paragraph middle; O: non-boundary), and segments the text based on semantic coherence changes between sentences and the patent-specific logical transitions, outputting semantically complete natural paragraphs.

[0031] Specifically, for multilingual mixed text, a multilingual pre-trained model is used to identify the main language and perform patent-optimized translation: Build and maintain a terminology database in the patent field, and make mandatory term replacements before and after translation (e.g., "embodiment" is fixed to be translated as "integration").

[0032] When calling the translation service, specific prompts are used to request a translation that conforms to the style and structure of patent legal texts and long sentences.

[0033] Check the position and syntactic rationality of key terms using rules or lightweight models, and mark suspected mistranslations.

[0034] Finally, text cleaning and standardization are performed to automatically identify and remove noise information in the patent text, including legal status, typesetting metadata, OCR garbled characters, duplicate entries, and hidden characters. At the same time, full-width to half-width characters and XML format standardization are performed to ensure that the input text is clean and formatted consistently.

[0035] Step S2: Based on the rule-based localization engine, identify and extract key information from the target patent text; the key information includes technical field, function, application scenario, pain point solved, beneficial effect, technical solution, working principle, functional carrier and its value characteristics.

[0036] Step S3: Extract functional triples from the key information based on the large language model and the Prompt template. The functional triples are: function - function implementation form - value feature.

[0037] Specifically, the following steps are included: Step S31, Paragraph Input and Preprocessing It is understandable that the patent text paragraphs containing key information extracted in step S2 are input into a pre-trained large language model. These paragraphs cover key information such as the patent's technical field, functional description, application scenarios, pain points to be solved, beneficial effects, technical solutions, working principles, functional carriers and their characteristics.

[0038] Before input, paragraphs undergo necessary preprocessing, including but not limited to language standardization, format standardization, and noise removal, to ensure the quality and consistency of input data.

[0039] Step S32, Prompt Template Design and Application Prompt templates not only contain explicit instructions, but also incorporate professional terminology and structured output format requirements in the field of patent analysis, thereby achieving specific outputs that meet the requirements.

[0040] In this embodiment, the Prompt template for function extraction is as follows: Please extract the core function of the patent from the following paragraphs, and output it in the format: 'Function: XXX'.

[0041] Prompt template for extracting functional implementation forms and value characteristics: Please analyze the following paragraphs, extract the specific technical means to achieve this function and the value enhancement it brings, and output the following format: 'Function Implementation Form: XXX', 'Value Characteristics: XXX'.

[0042] The Prompt template effectively controls the output of large language models, ensuring that they meet the requirements of subsequent value feature mapping and multi-dimensional evaluation.

[0043] Step S33: Extraction of functions, functional implementation forms, and value characteristics. After receiving the input paragraph and Prompt template, the large language model uses its powerful natural language understanding capabilities to perform in-depth analysis of the paragraph.

[0044] Specifically, the large language model will identify and extract the core functions of a paragraph, the specific technical means to achieve those functions (i.e., the form of function implementation), and the value characteristics brought about by those technical means.

[0045] For example, in a patent paragraph describing a novel energy storage technology, the model might extract key information such as "Function: High-efficiency energy storage and rapid release," "Function Implementation: Adopting a combination of high-speed permanent magnet motor and matrix converter technology," and "Value Features: The system operates using a high-speed permanent magnet motor as an electric motor, achieving efficient conversion of electrical energy to mechanical energy, and combines this with a matrix converter to precisely control the charging process, ensuring the system has a large-capacity energy storage capacity. Simultaneously, the vacuum cavity formed by the metal casing effectively reduces the wind resistance loss of the system rotor, improving the overall energy conversion efficiency."

[0046] It is important to note that the target patent includes not only text information but also multimodal data such as images. Therefore, in some embodiments, it is necessary to follow a Prompt template when inputting multimodal information.

[0047] The large language model performs modal encoding on the data of each modality and uses a unified space mapping method to map the features of each modality to a unified space. The modal encoding adopts a general technique and uses machine learning-related models such as neural networks to uniformly represent multimodal information, providing a unified input for subsequent processing by the large language model.

[0048] Attention mechanisms are used for cross-modal interaction to generate cross-modal fusion features, and these features are then processed by a multilayer perceptron to generate task-related outputs. y = MLP ( hfused ) y The sequence probability distribution of the functional triples; h fused = CrossAttention ( Q = h anchor , K ={ h1,..,h m}, V = { h 1 ,...h m}) represents cross-modal fusion features; Q / K / V For features from different modalities, m This represents the number of different modal features.

[0049] Step S34, Function—Function Implementation Form—Generation of Value Characteristics The large language model combines the extracted functions, function implementation forms, and value feature information into structured function-function implementation form-value feature function triplets.

[0050] These functional triples not only clearly express the core value of the patented technology, but also provide a convenient data foundation for subsequent value feature mapping and multi-dimensional evaluation.

[0051] For example, the aforementioned energy storage technology patents may generate the following functional triples and value mappings: Function Functional implementation form Value characteristics Energy storage The high-speed permanent magnet motor is an integrated charging and generating motor, with its input / output terminals connected to a matrix converter. When the flywheel energy storage system operates in charging mode, the high-speed permanent magnet motor functions as a motor with constant torque input. The system operates using a high-speed permanent magnet motor to achieve efficient conversion of electrical energy into mechanical energy. Combined with a matrix converter, the charging process is precisely controlled, ensuring the system has a large energy storage capacity. Simultaneously, the vacuum cavity formed by the metal casing effectively reduces the wind resistance loss of the system rotor, improving the overall energy conversion efficiency. Energy release The high-speed permanent magnet motor is an integrated charging and generating motor, with its input / output terminals connected to a matrix converter. In generator mode, the high-speed permanent magnet motor operates as a generator, providing constant power output. The system operates as a generator using a high-speed permanent magnet motor, which stably converts the mechanical energy stored in the flywheel into electrical energy output. The matrix converter ensures the quality of electrical energy and enables fast and stable charging and discharging switching. In summary, through the above-mentioned method for deep extraction of patent value information and generation of functional triples based on large language models and Prompt templates, this invention can achieve accurate identification and structured representation of high-value information in patent texts, laying the foundation for subsequent patent value assessment and multi-dimensional analysis.

[0052] Step S4: Extract the feature vectors of the functional triples and obtain the value feature alignment vectors through retrieval enhancement generation technology.

[0053] Specifically, common text vectorization techniques are used to extract feature vectors from functional triples, for example, encoding text into vectors using a neural network model. This will not be elaborated upon further here.

[0054] After obtaining the feature vector of the functional triple, the value feature alignment vector is further obtained through retrieval enhancement generation technology. This ensures that the subsequent value mapping and value assessment are not based on the patent text in isolation, but combine external academic, market and policy evidence, making the results more objective and persuasive.

[0055] It's understandable that retrieval-enhanced generative technology connects the powerful generation and reasoning capabilities of a large language model with a dynamically updated, specialized external knowledge base. In the context of patent valuation, when analyzing a patent, it can retrieve, in real time, the most relevant and novel comparative materials and background knowledge related to the current patent technology from multiple sources such as massive patent databases, scientific and technological literature, legal documents, and market reports. For example, when evaluating a patent on "silicon carbide power devices," the RAG framework can proactively retrieve core papers in the field from the past three years, competitors' patent portfolios, the evolution of relevant industry standards, and the latest market size data. Then, based on this high-quality, highly relevant external knowledge, the large language model performs in-depth integration analysis and reasoning to generate an evaluation report or value judgment.

[0056] Step S5: Based on the feature vector and value feature alignment vector of the functional triple, the functional triple is mapped to the preset value information index system based on the pre-trained mapping model to obtain the overall value information vector of the target patent text in each index dimension. The mapping model is as follows: ;

[0057] V i Indicates the first i The overall value information vector of each indicator dimension i ∈[1, N ], N This represents the total number of indicators in the value information indicator system. ΔY i Indicates the relationship with the first i Alignment vector of value features across multiple indicator dimensions π i Aligning the value feature vector with the first i The adjustment coefficient for each indicator dimension, with a value range of [0,1]; T j Indicates the first j The feature vector of a functional triple. n The number of functional triples; θ ij For the first j The feature vector of the t function triplet pairs the t function triplet. i The weight coefficients for each indicator dimension satisfy... .

[0058] Specifically, the weighting coefficients θ ij and adjustment coefficient π iThe patent value annotation corpus is obtained by fine-tuning a pre-trained large language model on the patent value annotation corpus; in particular, the patent text samples are annotated based on a preset value information index system to obtain the patent value annotation corpus.

[0059] The training process includes the following steps: Based on a pre-defined value information indicator system, a benchmark value feature vector is constructed for each indicator dimension as a benchmark for consistency comparison. based on The formula calculates the value information vector of the patent value annotation corpus in the i-th indicator dimension. V i ; The overall value information vector is quantified using the cosine similarity formula. V i The degree of consistency between the feature vector and the benchmark value of the corresponding indicator dimension: ; in, Representing the i Benchmark value feature vector for each indicator dimension; If the consistency level does not meet the preset threshold, the weighting coefficient will be adjusted. θ ij and adjustment coefficient π i It returns the value information vector of the patent value annotation corpus in the i-th indicator dimension. V i ; If the consistency level meets the preset threshold, then the current weight coefficients will be adjusted. θ ij and adjustment coefficient π i As the final weighting coefficient θ ij and adjustment coefficient π i .

[0060] Step S6: In response to the user's input analysis intent, and based on the overall value information vector of the target patent text in each indicator dimension, drive the report generation engine to generate a structured evaluation report. The structured evaluation report includes the value mapping information of the target patent document in at least one indicator dimension, and the evaluation results in at least one indicator dimension.

[0061] Specifically, responding to the user's analytical intent and based on the overall value information vector of the target patent text across various indicator dimensions, the report generation engine generates a structured evaluation report, including: Based on the user's input analysis intent, the overall value information vector of the corresponding indicator dimension is sent to the vector decoder. After the semantic de-decoding operation of the vector decoder, textual value mapping information is output. Based on the overall value information vector of the corresponding indicator dimensions, a weighted summation is applied to generate the final multi-dimensional value assessment value: ;

[0062] in, ω l The dynamic weights for each indicator dimension are determined based on domain expert experience or data-driven methods, reflecting the relative importance of each indicator dimension in the overall evaluation. , L ∈[1, N ].

[0063] Based on the value mapping information and multi-dimensional value assessment values, the structured evaluation report is synthesized.

[0064] For example, if users prioritize the technological value of a patent, then that technological value should be given a high weight. Alternatively, the weight of other technological values ​​that users don't care about can be set to 0.

[0065] Specifically, the vector decoder can be a fine-tuned lightweight text generation model that decodes descriptive text based on vectors; alternatively, it can be a large language model that interprets vectors into text descriptions by constructing a specific prompt. Other forms are also possible, and the appropriate method can be chosen based on actual needs. Specific implementation methods are not provided here.

[0066] It is understandable that the above method allows users to perform customized evaluations of any subset of dimensions in the indicator system based on specific analytical intentions. The specific implementation steps are as follows: Dimension Intent Resolution: The system first parses the evaluation instructions entered by the user and identifies the set of target dimensions specified by the user. This set can be a single dimension, such as only "innovation", a combination of some dimensions, such as "innovation" and "business potential", or all dimensions.

[0067] Vector Subset Extraction and Weighting: Based on the parsed target dimension set, the corresponding dimension scores are extracted from the complete overall value information vector of the target patent, forming a focused value sub-vector. Subsequently, the system applies preset or user-defined weight coefficients to perform weighted summation on this sub-vector to generate a comprehensive evaluation value.

[0068] Dynamic computation logic: This process is implemented through a configurable weighted computation engine. Its computation logic can be formally represented as: Overall evaluation value = Σ( ω l • V l ),in l This belongs to the user-specified set of dimensions. If the user does not specify weights, the system's default equal weights or domain-standard weights will be used.

[0069] Output and Interpretation: The system will ultimately output the comprehensive evaluation value and clearly explain the dimensions it covers, the weights used, and the calculation process in the evaluation report to ensure the traceability and interpretability of the results.

[0070] It is understandable that the output of the aforementioned patent valuation report goes far beyond a single value score; rather, it is a multi-dimensional set of value mapping information. This mapping information transforms abstract value concepts into concrete, interpretable, and actionable feature vectors, thereby providing structured insights for in-depth analysis and decision-making.

[0071] Example 2 The difference between this embodiment and Embodiment 1 is that it provides the steps for obtaining the value feature alignment vector.

[0072] Specifically, the steps for obtaining the value feature alignment vector include: Core keywords and concepts are extracted from the functional triples of the target patent text. Based on these core keywords and concepts, external data is queried using retrieval enhancement generation technology to obtain a set of external key information. X ; external key information set X eigenvectors of functional triples T Fusion, to obtain X ⊕ T ,in T ={ T 1,... T j , ... T n}; Targeting external key information sets X Through manual annotation or large language model processing, the indicator dimensions are divided and corresponding initial value feature vectors are generated. The initial value feature vectors of each indicator dimension are aligned using a pre-trained aligner model to obtain the value feature alignment vector.

[0073] Specifically, in the intelligent system of patent valuation, to ensure that the evaluation results generated by the large language model are deeply aligned with the professional standards, legal requirements, and multi-dimensional analytical needs of the patent field, this embodiment introduces a domain aligner based on residual learning. This aligner is fine-tuned on a high-quality evaluation dataset annotated by patent experts, and its core objective is to learn the professionally corrected residuals between the excellent evaluation conclusions judged by experts and the original output of the model. These residuals essentially capture the systematic gap between professional human cognition and general model understanding across multiple dimensions.

[0074] Specifically, the training steps of the aligner model include the following steps: A set of patent text samples is obtained. Each patent text sample is preprocessed, identified, and key information is extracted. Based on the large language model and the Prompt template, functional triples are extracted from the key information, and feature vectors of the functional triples are extracted. Extract the core keywords and concepts of the functional triples, query external data based on the core keywords and concepts, and obtain a set of external key information. X ; Targeting external key information sets X Through manual annotation or large language model processing, the indicator dimensions are divided and corresponding initial value feature vectors are generated. Yo (i) ; Initial value feature vectors for each indicator dimension by domain experts or advanced models Yo (i) Corrections are performed to form a modified value feature vector. Yc (i) ; Build a corrected dataset; Using the residual learning framework of the Aligner model, an aligner model is constructed to learn and correct the value feature vector. Yc (i) and initial value feature vector Yo (i) The residuals between them are used to define the objective function as:

[0075] in, φ These are the parameters of the aligner model; The objective function is minimized through an iterative optimization algorithm, which drives the model parameter update. The training of the aligner model is completed when the model converges or reaches a preset standard.

[0076] After training, this domain aligner works in a nested manner with the preceding general large language model.

[0077] When the model performs a preliminary analysis of a patent, the aligner intervenes and calibrates the generated mid-level feature vectors or preliminary conclusions in real time. Specifically, it uses a pre-trained aligner model to align the initial value feature vectors for each indicator dimension, including the following steps: Obtain external key information sets X Then, using a large language model, the external key information set... X The data is divided into dimensions, and initial value feature vectors for each indicator dimension are generated. Yo (i) ; Combined data X ⊕ T The pre-trained aligner model is used to initialize the feature vectors of each indicator dimension. Yo (i) Alignment is performed to obtain the value feature alignment vector: .

[0078] When a large language model performs a preliminary analysis of a target patent text, the aligner intervenes and calibrates the generated initial value feature vector in real time. It doesn't completely rewrite the model's output, but rather, by adding learned residual correction terms, it finely adjusts the model's initial, potentially superficial or cognitively biased, answer distribution towards a more technically accurate, legally rigorous, and value-relevant direction. For example, it corrects a vague description of a technological advantage into a rigorous statement with specific performance indicator comparisons and differentiation from existing technologies; or it adjusts a general market potential judgment into a quantitative analysis relating to specific application scenarios and the position within the industry chain.

[0079] Through this domain-specific alignment process, the generalized text feature vectors generated by common models are transformed into domain-specific feature vectors that are deeply tailored to the professional needs of patent valuation, emphasizing both structure and semantics. This optimized feature vector not only contains the original information of the patent text but also embeds the professional logic and evaluation criteria for patent valuation, thus providing a solid, reliable, and directly applicable semantic representation foundation for subsequent higher-level tasks such as valuation.

[0080] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0081] Example 3 The difference between this embodiment and Embodiment 1 or Embodiment 2 is that it provides specific steps for identifying and extracting key information from the target patent text based on a rule-based localization engine.

[0082] The steps include: (1) Chapter title recognition: Regular expressions are used to match fixed chapter titles in the patent text, such as “Technical Field”, “Invention Content”, “Specific Implementation”, etc., to quickly locate the core chapters in the patent text and provide scope guidance for subsequent information extraction.

[0083] This embodiment does not rely on a few fixed regular expressions, but instead constructs a multilingual, multi-variant patent chapter title knowledge base. Specifically: This knowledge base includes common and rare variations of expressions in major patent languages ​​such as Chinese, English, Japanese, and German, as well as for each statutory section. For example, for the "Summary of the Invention" section, the dictionary may contain: ["Summary of the Invention", "Summary of the Invention", "Summary of the Invention", "Summary of the Invention", "Summary of the Invention",...].

[0084] During recognition, for each line of the target patent text, string matching is used to match it with the patent chapter title in the patent chapter title knowledge base. Fuzzy string matching algorithms (such as Levenshtein distance and cosine similarity) are also used to compare it with the title variants in the patent chapter title knowledge base, and a matching confidence score is calculated to effectively combat OCR errors or subtle format differences.

[0085] Finally, prior knowledge of the chapter order in patent documents is introduced. For example, after identifying the "technical field," the next chapter is expected to be "background art" or "invention content," rather than "claims." Using this chapter order relationship as a constraint, identification results with low confidence can be verified or corrected.

[0086] (2) After fixing the chapter title, the extended relation identification model based on multi-dimensional language features and dependency parsing is used to locate the extended sentences in the technical field, so as to extract the specific features and extended scope of the technical field.

[0087] It is understood that the extended relation recognition model based on multi-dimensional language features and dependency parsing in this embodiment goes beyond simple keyword matching. It comprehensively judges whether an extended relation exists by analyzing the multi-dimensional language features of the sentence.

[0088] Specifically, the extended sentence localization technology based on multi-dimensional language features and dependency parsing extended relation recognition models includes: A comprehensive and categorized lexicon is constructed. Based on the lexicon, the patent content following the chapter titles is matched to extract the original sentences containing the lexicon, forming a preliminary candidate set.

[0089] The guiding vocabulary is divided into the following categories: Examples: "for example", "like", "as shown"; Detailed descriptions: "specifically," "more specifically," "that is"; Preferred categories: "preferred," "more preferably," "best of all"; Explanatory terms: "that is to say", "in other words".

[0090] The syntactic structure of sentences is analyzed using syntactic analysis tools to examine the grammatical relationship between introductory words and preceding and following components. Based on the analysis results, grammatical constraint rules are established to clean the initial candidate set.

[0091] For example, using syntactic analysis tools to determine whether “specifically” introduces an inserted explanatory element or actually introduces a purpose clause or adverbial clause can help eliminate ambiguity.

[0092] Finally, the cleaned results are subjected to in-depth semantic verification and localization based on semantic roles and referential resolution.

[0093] By using semantic roles and referential resolution to determine whether the content of an extended sentence truly refers to the core concept expounded in the preceding chapter title or topic sentence, sentences that appear to be extensions but actually refer to other content can be filtered out.

[0094] Construct a deep semantic template based on semantic role labeling and a patent value element dictionary to extract the core functions, application scenarios and technical problems solved by the patent.

[0095] Specifically, semantic role labeling analysis involves labeling sentences with semantic roles to identify semantic components such as predicate (core action), agent, patient, purpose, and result.

[0096] Patent Value Elements Dictionary: Construct a keyword dictionary of patent value elements, containing core words expressing "function", "scenario", and "problem" (such as "improve efficiency", "solve...problem", "in...scenario").

[0097] Semantic template definition: The criterion for determining key sentences is no longer the string itself, but rather the combination pattern of semantic components. For example, a sentence is determined to describe a "function" if its semantic structure satisfies: [Patent subject] + [functional verb / predicate, such as "improve," "achieve," "avoid"] + [technical effect object], and it contains keywords from the value element dictionary. This method can capture sentences with different structures but the same semantic meaning, such as "This invention improves heat dissipation efficiency through structure A" and "This design aims to achieve lossless signal transmission."

[0098] (3) Based on the semantic role labeling and the deep semantic template of the patent value element dictionary, identify key sentences describing key information to extract the core functions, application scenarios and technical problems solved by the patent.

[0099] The purpose of information extraction and standardization is to standardize the extracted key information, including deleting modifiers, retaining core verbs and nouns, and merging synonyms. This ensures that the extracted information is concise and accurate, facilitating subsequent large language model analysis and value assessment.

[0100] Traditional basic text preprocessing techniques have the following problems: Domain-specificity deficiencies lead to information distortion: For example, the adverb "significantly" can be removed in general text, but in a patent, "significantly improves efficiency" is an important technical effect statement, and removing "significantly" results in the loss of crucial information. General thesaurus cannot understand whether "device" and "equipment" are truly interchangeable in a specific patent context.

[0101] Disruption of syntactic structure leads to semantic loss: Simply "retaining the core verb and noun" is a bag-of-words model approach that destroys the semantic relationships between words. For example, when refining "motor drives gear," if only "motor," "drive," and "gear" are retained, the core verb-object relationship of "who drives whom" is lost, and the semantics become ambiguous.

[0102] A lack of semantic understanding leads to superficial standardization: true standardization requires understanding the precise meaning of a concept in a specific technological context. For example, in a patent, "processor" could refer to a central processing unit (CPU) or a graphics processing unit (GPU).

[0103] This embodiment avoids the above problems by constructing an information compression and standardized prompt word project based on a large language model, and also reflects the deep integration with the large language model: Specialized prompt words are designed, and the large language model is used to standardize and transform the key sentences in the input.

[0104] Example prompt: Please condense the technical functional information contained in the following patent sentence into a concise (subject, function, object) format. Please use standard technical terminology for the subject and object. Sentence: [Enter sentence] Few-shot learning: Provides several standardized examples to allow large language models to learn the format and requirements of standardized output through few-shot learning.

[0105] Output structured data: The large language model is required to directly output structured data such as JSON, which is convenient for subsequent processing by the program.

[0106] (4) Use a large language model to extract end-to-end relational data, design prompt words, and generate structured relational data from technical solutions and working principles to identify position and connection relationships.

[0107] In the technical solution and working principle section, specific signal words (such as "composed of," "connected to," "arranged in," etc.) are used to identify the composition, connection relationships, and spatial layout of components. The technical implementation details of the patent are extracted to provide a basis for subsequent value presentation, such as technology maturity analysis and core performance indicator evaluation.

[0108] Common approaches include: pre-defining a list of trigger words (such as "includes", "connects", "installed in", "set in", etc.). When these words appear in the text, it is assumed that there may be a component relationship between them, and attempts are made to identify the noun phrases before and after the trigger word as two entities in the relationship. In addition, simple syntactic rules can be combined, such as matching patterns like "[component A] + connection + [component B]", which is a slight improvement over simple trigger words.

[0109] However, the following main problems exist: ① Diverse expressions lead to low recall: The ways in which relationships are described in patents are extremely flexible and varied. In addition to "connected to," countless other methods can be used, such as "fixed to," "coupled to," "connected to," "the end of which is provided with," and "one end of which is inserted into." The list of fixed trigger words is inexhaustible, resulting in a large number of relationships that cannot be identified.

[0110] ② Syntactic complexity leads to extraction errors: Patent sentences are often long and complex, containing multiple clauses and modifiers. Simple patterns cannot handle long-distance dependencies. For example, in the sentence "The first component, which is connected to the second component located inside the shell by a rotatable axis," the subject "first component" and the object "second component" of the relation are far from the trigger word "connected," and there is a complex structure in between, making it easy to extract errors or failures.

[0111] ③ Ambiguous relational semantics lead to low accuracy: A trigger word may correspond to multiple relation types. For example, "set" can refer to spatial location ("set on top") or installation method ("set by bolts"). Relying solely on trigger words cannot distinguish these subtle differences, resulting in inaccurate relation extraction. In addition, non-technical relations may also be extracted (such as "this invention sets a target").

[0112] This embodiment uses a large language model to extract end-to-end relational data. By designing prompt words, it directly generates structured relational data from the text.

[0113] Example prompt: Please extract all information describing the composition, connection, or positional relationships between components from the following patent text fragment. Please output in JSON list format, with each element containing three fields: subject, relation, and object. Text: [Input text].

[0114] (5) Design prompt words and intelligent purification based on big language model semantic understanding to desubjectify and condense the extracted sentences.

[0115] During information extraction, non-technical modifiers (such as "advanced" or "new type"), repetitive adjectives, and adverbs of degree (such as "significant" or "extremely large") should be excluded. This ensures that the extracted information is objective and accurate, avoiding the influence of subjective modifications on subsequent analysis.

[0116] Conventional methods have the following problems: Simply removing all degree adverbs and adjectives leads to information loss. In patents, certain degree adverbs (such as "significantly improves" and "greatly simplifies") and adjectives (such as "high precision" and "low cost") are themselves important objective components of technical effects or performance indicators. Removing them results in the loss of key quantitative or qualitative information, making the information inaccurate. For example, after filtering out "significant," the value difference between "improves efficiency" and "significantly improves efficiency" disappears.

[0117] Lacking domain adaptability, general filtering rules may mistakenly delete valid information. For example, "novel" is a subjective promotional term in general contexts, but in a patent for "a novel connection structure," it may be an objective description of a technical state.

[0118] This embodiment utilizes the intelligent purification of semantic understanding by a large language model, designing prompt words to allow the large language model to "de-subjectify" and "condense" the extracted sentences.

[0119] Specifically, the prompt example is: Please separate the objective technical facts from the subjective promotional descriptions in the following patent sentence. Output only the objective technical facts portion and rewrite it as a concise declarative sentence.

[0120] Sentence: This invention provides an unprecedented advanced device that can greatly improve production efficiency and has outstanding economic benefits.

[0121] Expected output: This device will improve production efficiency.

[0122] Few-shot learning: Provide several examples before and after purification to allow the large language model to learn the purification criteria.

[0123] The above-mentioned rule-based positioning engine can quickly locate high-value paragraphs (such as "Technical Field", "Summary of Invention", "Detailed Implementation" and other chapters, as well as extended sentences such as "more specifically" and "characterized by".

[0124] Example 4 The difference between this embodiment and embodiments 1, 2, or 3 is that: The preset value information indicator system is the national standard "Specification for Evaluation of Scientific and Technological Achievements" (GB / T 44731-2024), which includes six dimensions of indicators: scientific value, technological value, economic value, social value, cultural value, and potential for transformation and promotion.

[0125] The following provides examples of external key information sets obtained through manual annotation or large language models. X The literature information is used to make a preliminary assessment of its value across six dimensions, generating initial features of the patent text in each indicator dimension. Yo One possible implementation.

[0126] Scientific Value of the Target Patent: Based on the structural relationships between the core elements of the target patent's function (e.g., "using a high-speed permanent magnet motor + matrix converter to improve system cycle efficiency" or "constructing a vacuum chamber + water-cooled jacket structure to reduce standby power loss"), academic paper databases in relevant disciplines are searched, targeting the key scientific issues and technological innovations involved in the patent technology. Specifically, searches are conducted using precise keyword combinations (e.g., patent technology name, core invention point, technical problem solved, etc.) to locate high-level academic conference papers and articles published in core journals that directly discuss or apply the key scientific issues and technological innovations in the patent technology. The core functions, functional implementation forms, technical effects, or scientific principles in the patent technology are analyzed or verified. The impact factor data of journals publishing papers directly related to the core elements of the patent technology are statistically analyzed, especially focusing on the impact factor trends over the past five years.

[0127] Based on this analysis, the scientific value of the target patent is: Theoretical Innovation: Evaluate the theoretical innovation of the patented technology by combining the scientific questions, theoretical models, experimental designs, or data analysis methods presented in the paper. For example, if the patented technology solves a long-standing scientific problem or proposes a completely new theoretical framework, its scientific value is significant.

[0128] Interdisciplinary nature: This analyzes whether the patented technology promotes the cross-disciplinary integration of different disciplines, such as collaborative innovation in materials science, electronic engineering, and energy technology. Patented technologies with strong interdisciplinary nature often drive the expansion of scientific boundaries and have high scientific value.

[0129] Academic Influence: When assessing the academic influence of a patented technology, multiple factors need to be considered, particularly the impact factor of the journals in which directly related papers are published, the citation frequency of these papers, and the presentations of the patented technology at academic conferences. Specifically, attention should be paid to papers that deeply explore or verify the core functions, functional implementation forms, technical effects, or scientific principles of the patented technology. Their publication records in high-impact journals and the frequency of citations are important indicators for measuring the academic influence of a patented technology.

[0130] Scientific value assessment and its connection with policy documents: Citing relevant policy provisions in policy documents concerning scientific and technological innovation, discipline development, and evaluation of scientific research achievements in related fields, such as national support policies for basic research, cutting-edge technology exploration, and major issues, to further demonstrate the scientific value of patented technologies and their contribution to the national science and technology development strategy.

[0131] Taking scientific value as the core starting point, the following quantitative formula is introduced to expand the assessment of comprehensive scientific value: SV =α 1 Innovation+α 2 Interdisciplinarity+α 3 Impact+α 4 Policy Alignment in, SV Represents the initial scientific value feature vector Innovation Represents theoretical innovation, Interdisciplinarity Representing interdisciplinary nature, Impact Represents academic influence. Policy Alignment Represents the relevance of policy documents; α 1. α 2. α 3. α 4 represents the weighting coefficient for the corresponding indicator dimension, determined based on the experience of domain experts or data-driven methods, to reflect the relative importance of different scientific value elements in the overall evaluation.

[0132] The technological value of the target patent: 1) Search for industry analysis reports, white papers on new products from leading companies, patents, third-party testing reports, and other technical route materials with equivalent functional implementation forms to the patent, and analyze the innovativeness of the patented technology: Analyze existing industry technologies from dimensions such as performance parameters, process / structure / materials, distribution of technical routes, standards and regulations (current national / industry standards / IEC / JEDEC numbers and versions), literature and patents (average annual number of patents from the same IPC / CPC group in the past 5 years; the 3 most cited non-patent documents; core patent publication numbers and independent claims of leading enterprises).

[0133] Analyze changes in performance, processes / structures / materials (whether new materials, structures, or nodes are introduced for the first time; the impact of adding or removing process steps on yield), technological leaps (whether new routes are opened or old routes are phased out), standard updates (whether any indicators or clauses consistent with this patent have been added or tightened in the past 5 years or in drafts for comments), and literature evidence (search results showing that the same "functional implementation form" does not appear in domestic and foreign published patents, papers, and reports in the past 5 years; information on the value of changes in existing technologies, such as the first mention of equivalent technologies and parameter comparisons in white papers by leading enterprises in the past 5 years).

[0134] 2) Search for papers, patents, test reports, and other materials on technical routes with equivalent functional implementations to this patent, and analyze the maturity of this patented technology: The analysis includes verification environment information from various dimensions such as environment category (laboratory / simulated field / real scenario), environmental boundary conditions (temperature, humidity, air pressure, power supply, load, operating condition curve), device / prototype version, and data acceptance level (third-party independent testing / commissioned testing / self-testing).

[0135] Analyze the verification conclusions from dimensions such as core performance indicators, failure modes, repeatability / reproducibility, safety / environmental friendliness, and comparison with standards / specifications.

[0136] 3) Search for standards, papers, patents, test reports, industry analysis reports, and other materials that provide equivalent technical solutions to the patented functionality, and analyze the comparison between the patented technology and mainstream technologies: Analyze the technical specifications in international / national standards (such as ISO, IEC, GB, ASTM) that are functionally equivalent to the patents, focusing on comparing performance testing methods, index definitions, and limit requirements; analyze highly cited papers from the past 5 years to extract performance data of equivalent technologies (such as energy density, efficiency, and loss); compare the scope of claims of patent families and competitor patents; analyze collected third-party testing reports on equivalent technology products and extract key performance indicators and comparisons of mainstream technologies, etc., for valuable information.

[0137] Taking technological value as the core starting point, the following quantitative formula is introduced to further expand the evaluation of its comprehensive technological value: TV = γ 1 Innovation+γ 2 Maturity + γ 3 Benchmark in, TV As the initial technical value feature vector, Innovation Represents technological innovation, Maturity Represents maturity. Benchmark Represents core performance metrics and their weights compared to mainstream technologies; γ 1. γ 2. γ 3 represents the weighting coefficient for the corresponding indicator dimension, determined based on the experience of domain experts or data-driven methods, to reflect the relative importance of different technical value elements in the overall evaluation.

[0138] Economic value of the target patent: Search and analyze the core products of domestic and foreign patent technologies with equivalent technical routes to the patent's functional implementation form, and compare their degree of innovation, technical coverage, patent quality (such as the number of times the patent is cited and the number of claims, which may be higher, reflecting the maturity and influence of the technology), identification of key core technologies, and degree of technological monopoly (such as identifying the fields related to "bottleneck" technologies by analyzing patent data), and analyze the potential economic value contained in the patent.

[0139] Search for the market application prospects of the patented technology's leading products or services, the development stage of the relevant leading products, and the apparent and potential market size. For example, search industry analysis reports, key performance / process parameters of the listed or prototype products of the top 10 domestic and foreign leading companies in the past 3 years, as well as market value information such as market share, sales revenue, and net profit margin of their products. Output economic value information such as the expected market size, expected sales revenue, and expected net profit margin after the patent is put into the market.

[0140] Search for the expandable fields of a patented technology, and analyze its versatility, adaptability, and market demand, such as the diversity of application scenarios. The wider the application scope of the patented technology, the better its market prospects. Taking 5G communication technology as an example, it can be applied to multiple fields such as the Internet of Things, autonomous driving, and telemedicine, with huge market potential. Expandable fields include new materials, medical care, next-generation information technology, green energy, biotechnology, and manufacturing.

[0141] Taking economic value as the core starting point, the following quantitative formula is introduced to further expand the assessment of its comprehensive economic value: EV = β 1 Intrinsic+β 2 Marketable + β 3 Scalable This formula connects the technological coverage and practical elements in scientific value analysis, forming a complete evaluation chain from theoretical breakthroughs to economic transformation: whereby... EV This represents the initial economic value feature vector; Intrinsic Reflecting the core innovation and quality of technology, Marketable To measure a product's market potential and economic benefits, Scalable Assess the value of cross-domain applications; β 1. β 2. β 3 represents the weighting coefficient for the corresponding indicator dimension, reflecting the contribution ratio of different economic values ​​in the overall assessment, and is determined based on the experience of domain experts or data-driven methods.

[0142] Social value of the target patent Based on the specific functions claimed by the target patent (such as "energy storage") and its explicit implementation form (such as "the high-speed permanent magnet motor is an integrated charging and generating motor, with the input / output terminals connected to a matrix converter. When the flywheel energy storage system is running in charging mode, the high-speed permanent magnet motor operates as a motor with constant torque input"), analyze the overall social value of the patented technology in promoting green transformation, energy conservation and carbon reduction, and industrial upgrading: 1) Policy document system retrieval and comparison Relevant national and local policy documents were retrieved and compared, with a focus on documents such as the "14th Five-Year Plan for National Energy Conservation and Emission Reduction," the "14th Five-Year Plan for the Development of New Energy Storage," and the "Action Plan for Achieving Carbon Peak Before 2030." These policy documents provided the macro-level background and specific indicators for the analysis. Keywords closely related to the patented technology, such as "green and low-carbon," "energy conservation and consumption reduction," "system efficiency improvement," and "promotion of new energy storage technologies," were extracted from these policy documents as policy benchmarks for analyzing the social value of the patents.

[0143] 2) Establish a definite correlation between the form of functional implementation and policy orientation. Green and low-carbon dimension: Analyzing how patented technologies promote green and low-carbon development through specific functional implementations. For example, the patent significantly reduces wind power loss during equipment operation by "constructing a vacuum chamber + water-cooled jacket structure," a technological implementation that directly responds to the policy document's requirement to "promote green and low-carbon technological innovation."

[0144] Energy conservation and consumption reduction dimension: Specifically quantify the energy conservation and consumption reduction effects of patented technologies. For example, a patented technology that reduces system standby losses by "adopting a high-speed permanent magnet motor + matrix converter" directly aligns with the policy document's goal of "promoting the application of highly efficient energy-saving technologies and equipment." Relevant materials include energy efficiency test reports, specific energy conservation and consumption reduction indicators required by policy documents, and evidence that the patented technology meets or exceeds these indicators.

[0145] System efficiency improvement dimension: Evaluate the role of patented technology in improving the overall system efficiency. For example, if patented technology improves the system's cycle efficiency, this performance aligns with the goal of "improving the efficiency of energy storage systems and promoting green and low-carbon transformation" in the "Implementation Plan for the Development of New Energy Storage." Relevant materials may include system efficiency test reports, specific requirements for system efficiency improvement in policy documents, and detailed data on the efficiency improvement achieved by the patented technology.

[0146] Promotion of New Energy Storage Technologies: Analyzing the contribution of patented technologies to the promotion of new energy storage technologies. If a patent involves a new energy storage technology, its innovative functional implementation (such as a highly efficient thermal management system) will promote the application and promotion of this technology in the market, aligning with the policy documents' encouragement measures regarding "accelerating the commercial application of new energy storage technologies." Relevant materials may include reports on the market application of technologies with equivalent functional implementations, support policies for the promotion of new energy storage technologies in policy documents, and an analysis of the market acceptance and influence of the patented technology.

[0147] Taking social value as the core starting point, the following quantitative formula is introduced to further expand the assessment of its comprehensive social value: SoV=δ 1 Ecofriendly + δ 2 Energysaving + δ 3 Efficienc + δ 4 Adoption in, SoV This is the initial social value feature vector. Ecofriendly This represents the value contribution in green and low-carbon development. Energysaving This represents the value contribution in energy conservation and emission reduction. Efficiency This represents the value contribution to improving system efficiency. Adoption This represents the value contribution made in the promotion of new technologies; δ 1. δ 2. δ 3. δ 4 represents the weighting coefficients of each variable, determined based on the experience of domain experts or data-driven methods, to reflect the relative importance of different social value elements in the overall evaluation.

[0148] The cultural value of the target patent: Based on the value information regarding the function and implementation of the target patent, analyze its potential direct impact and value on the inheritance of outstanding Chinese culture and innovation in the cultural industry. Retrieve cultural value-related information concerning the target patent through various channels, including but not limited to: Policy documents and plans: Search and compare national and local government policy documents, special plans and action plans on the development of the cultural industry, such as the "Opinions on Promoting the High-Quality Development of the Digital Cultural Industry", analyze the degree of fit between the target patent technology and these policy orientations, and grasp the innovation direction under the guidance of policies.

[0149] Cultural and Tourism Industry Development Report: Referencing authoritative reports on the development of the cultural and tourism industry and market research reports, this report aims to understand the current development trends, hot areas, and potential demands of the cultural and tourism industry, assess the market potential and application prospects of the target patented technology in this field, and provide market basis for innovative practices.

[0150] Cultural tourism project case studies: Research successful cultural tourism project cases at home and abroad, especially those that utilize similar target patented technologies, and analyze their specific contributions to cultural inheritance, innovation and industrial upgrading.

[0151] Expert opinions and academic research: We searched for evaluations of the cultural value of patented technologies in the relevant functional implementation forms by experts in the cultural and technological fields; at the same time, we searched for relevant academic research papers to understand the academic community's evaluation of the impact of relevant technologies on cultural inheritance and development.

[0152] Based on this analysis, the cultural value of the target patent mainly includes: Contribution to cultural heritage preservation: Specifically, this refers to the contribution of technological innovation to cultural expression. It involves analyzing whether the technological innovations in patents provide new means or platforms for the creation, dissemination, or experience of cultural products. For example, enriching cultural tourism experiences through technologies such as virtual reality (VR) and augmented reality (AR) can foster an innovative cultural atmosphere that encourages exploration, thus perpetuating the cultural lineage in an innovative way.

[0153] The promotion of cultural innovation: Specifically, does it facilitate the inheritance and innovation of cultural content? This involves exploring whether patented technologies can promote the innovative expression or digital transmission of traditional cultural content, using innovative concepts to revitalize traditional culture. For example, using blockchain technology to protect the copyright and authenticity of intangible cultural heritage, or using new technologies to create works with unique cultural characteristics.

[0154] Impact on the cultural industry: Specifically, does it promote the upgrading and transformation of the cultural industry? Evaluate the role of patented technologies in improving the efficiency of the cultural industry, reducing costs, or opening up new markets, including whether they drive the development of related industrial chains and improve the quality and added value of cultural products.

[0155] Sociocultural benefits: Analyze the positive impacts of patented technologies on the socio-cultural environment, such as enhancing the public's innovative spirit, awareness, and scientific and cultural literacy, promoting cultural consumption, and strengthening cultural confidence, thereby fostering a socio-cultural atmosphere that respects creation and values ​​innovation.

[0156] Technology for Good: This analysis examines whether patented technologies adhere to the ethical principles of "people-centeredness and sustainable development," and whether their design, application, and potential impacts make positive contributions to social welfare, fairness and justice, and environmental protection. Specifically, it assesses whether the technological solutions incorporate value elements such as safety, reliability, universality, and inclusivity.

[0157] Taking cultural value as the core starting point, the following quantitative formula is introduced to further expand the evaluation of its comprehensive cultural value: CV = ε 1 Heritage + ε 2 Innovation + ε 3 Impact+ε 4 Benefit + ε 5 TechforGood in, CV This is the initial cultural value feature vector. Heritage Represents a contribution to cultural heritage. Innovation Represents the driving force of cultural innovation, Impact Representing the influence of the cultural industry, Benefit Represents social and cultural benefits. TechforGoodContributions or impacts representing the positive dimension of technology for good; ε 1. ε 2. ε 3. ε 4. ε 5 represents the weighting coefficients for each indicator dimension, determined based on the experience of domain experts or data-driven methods, to reflect the relative importance of different cultural value elements in the overall evaluation.

[0158] The potential for commercialization and promotion of the target patent Based on the value information regarding the functional implementation of the target patent (new value attributes of products in the application field), analyze the potential for transformation and promotion of the patent, including its ability to fill gaps in the field, its possibility of replacing existing technologies, and the potential market demand analysis for forming new products and services.

[0159] The specific target patent demonstrates new product value attributes in its functional implementation, including but not limited to how the patented technology improves the performance of existing products, reduces costs, enhances quality and efficiency, or creates entirely new market applications. Analyze the uniqueness of the patented technology and its potential to solve industry pain points and meet unmet market needs. Collect market, technological, and competitive intelligence related to the target patent by searching various scientific and technological literature, industry standards, market consulting reports, special research reports, and enterprise demand information. Specifically, this includes: The ability of technology to fill gaps in a field: By comparing the functions, performance parameters, and application scope of patented technology with existing technology, we can analyze whether the patented technology has filled a gap in a key field in the existing technology system, solved a long-standing technical problem, or created a completely new technological field.

[0160] Potential to Replace Existing Technologies: This assessment comprehensively considers factors such as the cost-effectiveness, compatibility, user acceptance, and market penetration of the patented technology to evaluate its potential and feasibility in replacing existing related product technologies. Through market research and competitive analysis, it aims to understand the dissatisfaction with existing technologies and the demand for improvement in the target market, thereby determining the market potential of the patented technology.

[0161] Barriers to the Development of New Products and Services: Identify and analyze existing and potential barriers that may hinder the transformation of patented technologies into new products or services. These include, but are not limited to, technological maturity, funding requirements, intellectual property protection, market acceptance, policy and regulatory restrictions, and cultural and customary limitations. Through case studies and policy analysis, assess the impact of these barriers on the commercialization and promotion of patented technologies.

[0162] Taking conversion and promotion potential as the core starting point, the following quantitative formula is introduced to further expand the evaluation of its comprehensive conversion and promotion potential: TP = ζ 1 Gapfilling + ζ 2 Substitutablity + ζ 3 Barrier in, TP This is the initial feature vector representing the potential for conversion and promotion. Gapfilling This represents the ability of technology to fill gaps in the field. Substitutablity This represents the possibility of replacing existing technologies. Barrier These represent obstacles to the formation of new products and services; ζ 1. ζ 2. ζ 3 represents the weighting coefficients for each indicator dimension, determined based on the experience of domain experts or data-driven methods, to reflect the relative importance of different conversion and promotion potential elements in the overall evaluation.

[0163] It is understandable that traditional patent evaluation focuses on the technical and economic dimensions. This invention is the first to incorporate soft indicators such as social value and cultural value into the automated evaluation system and achieves objective quantification through policy document matching.

[0164] Taking Table 1 as an example, after the aforementioned mapping steps, the six-dimensional patent value mapping information is shown in Table 2: ; ;

[0165] Example 5 Based on the same inventive concept, this application also provides an intelligent patent value mapping information extraction device for implementing the intelligent patent value information extraction method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the intelligent patent value mapping information extraction device provided below can be found in the limitations of the intelligent patent value information extraction method described above, and will not be repeated here.

[0166] The intelligent patent value information extraction device based on a large language model includes: The text preprocessing module is used to acquire and clean the original patent text and generate a structured paragraph sequence; The rule-based localization engine module is used to identify and extract key information from patent texts. The key information includes technical field, function, application scenario, pain point solved, beneficial effect, technical solution, working principle, functional carrier and its value characteristics. The functional triple generation module is used to extract information from the key information and generate functional triples using a pre-trained large language model and a Prompt template. The functional triples are function-function implementation form-value feature. The feature enhancement module is used to extract the feature vectors of functional triples and obtain the value feature alignment vectors through retrieval enhancement generation techniques. The value feature mapping module is used to map the functional triples to a preset value information index system based on the feature vectors of the functional triples and the value feature alignment vectors, using a pre-trained mapping model, to obtain the overall value information vector of the target patent text in each index dimension. The mapping model is as follows: ; V i Indicates the first i The overall value information vector of each indicator dimension i ∈[1, N ], N This represents the total number of indicators in the value information indicator system. ΔY i With the i Alignment vector of value features across multiple indicator dimensions π i Aligning the value feature vector with the first i The adjustment coefficient for each indicator dimension, with a value range of [0,1]; T j Indicates the first j The feature vector of a functional triple. n The number of functional triples; θ ij For the first j The feature vector of the t function triplet pairs the t function triplet. i The weight coefficients for each indicator dimension satisfy... ; A multi-format output module is used to respond to the user's input analysis intent and drive the report generation engine to generate a structured evaluation report based on the overall value information vector of the target patent text in each indicator dimension. The structured evaluation report includes the value mapping information of the target patent document in at least one indicator dimension, as well as the evaluation results in at least one indicator dimension.

[0167] It is understood that the large language model is DeepSeek, GPT, or other variants.

[0168] Example 6 This embodiment provides a computer device, which can be a terminal. The computer device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, communication interface, and memory communicate with each other via the communication bus. The memory is used to store computer programs. When the processor executes the program stored in the memory, it implements the intelligent extraction method of patent value information as described in any one of Examples 1-4.

[0169] Example 7 Based on the above embodiments, this embodiment provides a computer-readable storage medium storing a computer program thereon, characterized in that the computer program, when executed by a processor, implements the intelligent extraction method for patent value information as described in any one of embodiments 1-4.

[0170] Example 8 Based on the above embodiments, this embodiment provides a computer program product, including a computer program that, when executed by a processor, implements the intelligent extraction method for patent value information as described in any one of embodiments 1-4.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for intelligent extraction of patent value information based on a large language model, characterized in that, Includes the following steps: The target patent text is acquired and preprocessed. Based on the rule-based localization engine, key information in the target patent text is identified and extracted. The key information includes technical field, function, application scenario, pain point solved, beneficial effect, technical solution, working principle, functional carrier and its value characteristics. Based on the large language model and the Prompt template, functional triples are extracted from the key information. The functional triples are: function - function implementation form - value feature. Extract the feature vectors of functional triples, and obtain the value feature alignment vectors through retrieval enhancement generation techniques; Based on the feature vector and value feature alignment vector of the functional triple, the functional triple is mapped to the preset value information index system based on the pre-trained mapping model to obtain the overall value information vector of the target patent text in each index dimension. The mapping model is as follows: ; V i Indicates the first i The overall value information vector of each indicator dimension i ∈ [1,N ] ，N This represents the total number of indicators in the value information indicator system. ΔY i Indicates the relationship with the first i Alignment vector of value features across multiple indicator dimensions π i Aligning the value feature vector with the first i The adjustment coefficient for each indicator dimension, with a value range of [0,1]; T j Indicates the first j The feature vector of a functional triple. n The number of functional triples; θ ij For the first j The feature vector of the t function triplet pairs the t function triplet. i The weight coefficients for each indicator dimension satisfy... ; In response to the user's analytical intent and based on the overall value information vector of the target patent text across various indicator dimensions, the report generation engine generates a structured evaluation report. The structured evaluation report includes the value mapping information of the target patent document across at least one indicator dimension, as well as the evaluation results across at least one indicator dimension.

2. The method for intelligent extraction of patent value information based on a large language model according to claim 1, characterized in that, The steps for obtaining the value feature alignment vector include: Core keywords and concepts are extracted from the functional triples of the target patent text. Based on these core keywords and concepts, external data is queried using retrieval enhancement generation technology to obtain a set of external key information. X ; external key information set X eigenvectors of functional triples T Fusion, to obtain X ⊕ T ,in T ={ T 1, ... T j ,... T n }; Targeting external key information sets X Through manual annotation or large language model processing, the indicator dimensions are divided and corresponding initial value feature vectors are generated. The initial value feature vectors of each indicator dimension are aligned using a pre-trained aligner model to obtain the value feature alignment vector.

3. The method for intelligent extraction of patent value information based on a large language model according to claim 2, characterized in that, The training steps for the aligner model include: A set of patent text samples is obtained. Each patent text sample is preprocessed, identified, and key information is extracted. Based on the large language model and the Prompt template, functional triples are extracted from the key information, and feature vectors of the functional triples are extracted. Extract the core keywords and concepts of the functional triples, query external data based on the core keywords and concepts, and obtain a set of external key information. X ; Targeting external key information sets X Through manual annotation or large language model processing, the indicator dimensions are divided and corresponding initial value feature vectors are generated. Yo (i) ; Initial value feature vectors for each indicator dimension by domain experts or advanced models Yo (i) Corrections are performed to form a modified value feature vector. Yc (i) ; Build the corrected dataset ; An aligner model is constructed using the residual learning framework of the Aligner model. To learn and correct value feature vectors Yc (i) and initial value feature vector Yo (i) The residuals between them are used to define the objective function as: ； in, φ These are the parameters of the aligner model; The objective function is minimized through an iterative optimization algorithm, which drives the model parameter update. The training of the aligner model is completed when the model converges or reaches a preset standard.

4. A method for intelligent extraction of patent value information based on a large language model according to claim 2 or 3, characterized in that, Aligning the initial value feature vectors of each indicator dimension using a pre-trained aligner model includes the following steps: Obtain external key information sets X Then, through manual annotation or large language models, the external key information set is analyzed. X The data is divided into dimensions, and initial value feature vectors for each indicator dimension are generated. Yo (i) ; Combined data X ⊕ T The pre-trained aligner model is used to initialize the feature vectors of each indicator dimension. Yo (i) Alignment is performed to obtain the value feature alignment vector: 。 5. The method for intelligent extraction of patent value information based on a large language model according to claim 4, characterized in that, The preset value information indicator system is the national standard "Evaluation Specification for Scientific and Technological Achievements" (GB / T 44731-2024), which includes six dimensions of indicators: scientific value, technological value, economic value, social value, cultural value, and potential for transformation and promotion. The initial value feature vector of scientific value is: SV = α 1 Innovation+α 2 Interdisciplinarity+α 3 Impact+α 4 Policy Alignment in, SV Represents the initial scientific value feature vector Innovation Represents theoretical innovation, Interdisciplinarity Representing interdisciplinary nature, Impact Represents academic influence. Policy Alignment Represents the relevance of policy documents; α 1. α 2. α 3. α 4 represents the weighting coefficient for the corresponding indicator dimension, determined based on the experience of domain experts or data-driven methods, to reflect the relative importance of different scientific value elements in the overall evaluation; The initial value feature vector of technological value is: TV = γ 1 Innovation+γ 2 Maturity + γ 3 Benchmark in, TV As the initial technical value feature vector, Innovation Represents technological innovation, Maturity Represents maturity. Benchmark Represents core performance metrics and their weights compared to mainstream technologies; γ 1. γ 2. γ 3 represents the weighting coefficient for the corresponding indicator dimension, which is determined based on the experience of domain experts or data-driven methods to reflect the relative importance of different technical value elements in the overall evaluation. The initial value feature vector of economic value is: EV = β 1 Intrinsic+β 2 Marketable + β 3 Scalable EV This represents the initial economic value feature vector; Intrinsic Reflecting the core innovation and quality of technology, Marketable To measure a product's market potential and economic benefits, Scalable Assess the value of cross-domain applications; β 1. β 2. β 3 represents the weighting coefficient for the corresponding indicator dimension, reflecting the contribution ratio of different economic values ​​in the overall assessment, and is determined based on the experience of domain experts or data-driven methods. The initial value feature vector of social value is: SoV=δ 1 Ecofriendly + δ 2 Energysaving + δ 3 Efficienc + δ 4 Adoption in, SoV This is the initial social value feature vector. Ecofriendly This represents the value contribution in green and low-carbon development. Energysaving This represents the value contribution in energy conservation and emission reduction. Efficiency This represents the value contribution to improving system efficiency. Adoption This represents the value contribution made in the promotion of new technologies; δ 1. δ 2. δ 3. δ 4 represents the weighting coefficients of each variable, determined based on the experience of domain experts or data-driven methods, to reflect the relative importance of different social value elements in the overall evaluation; The initial value feature vector of cultural value is: CV = ε 1 Heritage + ε 2 Innovation + ε 3 Impact+ε 4 Benefit + ε 5 TechforGood in, CV This is the initial cultural value feature vector. Heritage Represents a contribution to cultural heritage. Innovation Represents the driving force of cultural innovation, Impact Representing the influence of the cultural industry, Benefit Represents social and cultural benefits. TechforGood Contributions or impacts representing the positive dimension of technology for good; ε 1. ε 2. ε 3. ε 4. ε 5 represents the weighting coefficients for each indicator dimension, determined based on the experience of domain experts or data-driven methods, to reflect the relative importance of different cultural value elements in the overall evaluation; The initial value feature vector of conversion and promotion potential is: TP = ζ 1 Gapfilling + ζ 2 Substitutablity + ζ 3 Barrier in, TP This is the initial feature vector representing the potential for conversion and promotion. Gapfilling This represents the ability of technology to fill gaps in the field. Substitutablity This represents the possibility of replacing existing technologies. Barrier These represent obstacles to the formation of new products and services; ζ 1. ζ 2. ζ 3 represents the weighting coefficients for each indicator dimension, determined based on the experience of domain experts or data-driven methods, to reflect the relative importance of different conversion and promotion potential elements in the overall evaluation.

6. A method for intelligent extraction of patent value information based on a large language model according to claim 1, 2, 3, or 5, characterized in that, The weighting coefficient θ ij and adjustment coefficient π i The patent value annotation corpus is obtained by fine-tuning a pre-trained large language model on a patent value annotation corpus; specifically, patent text samples are annotated based on a pre-set value information index system to obtain the patent value annotation corpus. The training process includes the following steps: Based on a pre-defined value information indicator system, a benchmark value feature vector is constructed for each indicator dimension as a benchmark for consistency comparison. based on The formula calculates the patent value annotation corpus in the first... i Value information vector of each indicator dimension V i ; The overall value information vector is quantified using the cosine similarity formula. V i The degree of consistency between the feature vector and the benchmark value of the corresponding indicator dimension: ； in, Representing the i Benchmark value feature vector for each indicator dimension; If the consistency level does not meet the preset threshold, the weighting coefficient will be adjusted. θ ij and adjustment coefficient π i And return the corpus of annotations used to calculate the patent value in the first instance. i Value information vector of each indicator dimension V i ; If the consistency level meets the preset threshold, then the current weight coefficients will be adjusted. θ ij and adjustment coefficient π i As the final weighting coefficient θ ij and adjustment coefficient π i .

7. A method for intelligent extraction of patent value information based on a large language model according to claim 1, 2, 3, or 5, characterized in that, When the target patent includes multimodal data, the key information in the target patent text is identified and extracted, and the multimodal data is fed into the large language model through the corresponding Prompt template. The large language model performs modal encoding on the data of each modality and uses a unified space mapping method to map the features of each modality to a unified space; Attention mechanisms are used for cross-modal interaction to generate cross-modal fusion features, and these features are then processed by a multilayer perceptron to generate task-related outputs. y = MLP ( h fused ) y The sequence probability distribution of the functional triples; h fused = CrossAttention ( Q = h anchor , K ={ h1,..,h m }, V = { h 1 ,...h m }) represents cross-modal fusion features; Q / K / V For features from different modalities, m This represents the number of different modal features.

8. The method for intelligent extraction of patent value information based on a large language model according to claim 7, characterized in that, Responding to the user's analytical intent and based on the overall value information vector of the target patent text across various indicator dimensions, the report generation engine generates a structured evaluation report, including: Based on the user's input analysis intent, the overall value information vector of the corresponding indicator dimension is sent to the vector decoder. After the semantic de-decoding operation of the vector decoder, textual value mapping information is output. Based on the overall value information vector of the corresponding indicator dimensions, a weighted summation is applied to generate the final multi-dimensional value assessment value: ； in, ω l The dynamic weights for each indicator dimension are determined based on domain expert experience or data-driven methods, reflecting the relative importance of each indicator dimension in the overall evaluation. ,L∈[1,N]; Based on the value mapping information and multi-dimensional value assessment values, the structured evaluation report is synthesized.

9. A method for intelligent extraction of patent value information based on a large language model according to claim 1, 2, 3, 5, or 8, characterized in that, The rule-based localization engine identifies and extracts key information from the target patent text, including: Match fixed chapter titles in the target patent text using regular expressions; After fixing the chapter titles, the extended relation identification model based on multi-dimensional language features and dependency parsing is used to locate the extended sentences in the technical field, so as to extract the specific features and scope of the technical field. Based on semantic role labeling and a patent value element dictionary, a deep semantic template is used to identify key sentences that describe key information in order to extract the core functions, application scenarios and technical problems solved by the patent. Based on the large language model, information compression and standardized prompt word engineering are used to standardize the identified key sentences; The design prompts drive the large language model to parse and generate structured relational data from the technical solution and working principle in an end-to-end manner. The relational data is used to identify a specific location or connection relationship, including three elements: subject, relation type, and object. In addition, design prompts and intelligent purification based on semantic understanding of large language models are used to desubjectify and condense the extracted sentences.

10. The method for intelligent extraction of patent value information based on a large language model according to claim 9, characterized in that, Matching fixed chapter titles in patent text using regular expressions, including: Construct a multilingual, multi-variant patent chapter title knowledge base; For each line of the target patent text, string matching is used to match it with the patent chapter title in the patent chapter title knowledge base, and fuzzy string matching algorithm is used to compare it with the title variant in the patent chapter title knowledge base. A matching confidence score is calculated, and the matching result is retained when the matching confidence score is greater than a preset threshold. By introducing prior knowledge of the chapter order in patent documents and using this prior knowledge as a constraint, identification results with matching confidence levels below a preset threshold are verified or corrected.

11. The method for intelligent extraction of patent value information based on a large language model according to claim 9, characterized in that, The extended relation identification model based on multi-dimensional language features and dependency parsing locates extended sentences in the technical field, including: Construct a comprehensive and categorized lexicon, match the patent text following the chapter titles based on the lexicon, extract the original sentences containing the lexicon, and form a preliminary candidate set; Dependency parsing is performed on the original sentences in the preliminary candidate set to examine the grammatical relationship between the introductory words and the preceding and following components. Based on the analysis results, grammatical constraint rules are established to clean the preliminary candidate set. Deep semantic verification and localization are performed on the cleaned results based on semantic roles and referential resolution.

12. A device for intelligent extraction of patent value information based on a large language model, characterized in that, include: The text preprocessing module is used to acquire and clean the original patent text and generate a structured paragraph sequence; The rule-based localization engine module is used to identify and extract key information from patent texts. The key information includes technical field, function, application scenario, pain point solved, beneficial effect, technical solution, working principle, functional carrier and its value characteristics. The functional triple generation module is used to extract information from the key information and generate functional triples using a pre-trained large language model and a Prompt template. The functional triples are function-function implementation form-value feature. The feature enhancement module is used to extract the feature vectors of functional triples and obtain the value feature alignment vectors through retrieval enhancement generation techniques. The value feature mapping module is used to map the functional triples to a preset value information index system based on the feature vectors of the functional triples and the value feature alignment vectors, using a pre-trained mapping model, to obtain the overall value information vector of the target patent text in each index dimension. The mapping model is as follows: ; V i Indicates the first i The overall value information vector of each indicator dimension i ∈[1,N], where N is the total number of indicators in the value information indicator system; ΔY i Indicates the relationship with the first i The value feature alignment vector of each indicator dimension, π i Aligning the value feature vector with the first i The adjustment coefficient for each indicator dimension, with a value range of [0,1]; T j Indicates the first j The feature vector of a functional triple. n The number of functional triples; θ ij For the first j The feature vector of the t function triplet pairs the t function triplet. i The weight coefficients for each indicator dimension satisfy... ; A multi-format output module is used to respond to the user's input analysis intent and drive the report generation engine to generate a structured evaluation report based on the overall value information vector of the target patent text in each indicator dimension. The structured evaluation report includes the value mapping information of the target patent document in at least one indicator dimension, as well as the evaluation results in at least one indicator dimension.

13. A computer device, characterized in that: It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the intelligent extraction method for patent value information as described in any one of claims 1 to 11.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the intelligent extraction method of patent value information as described in any one of claims 1 to 11.

15. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the intelligent extraction method for patent value information as described in any one of claims 1 to 11.

Images