A system and method for intellectual property valuation

Abstract

The application relates to the technical field of intellectual property evaluation, and discloses an intellectual property value evaluation system and method, wherein patent feature dataset is acquired through distributed data collection; a pre-trained language model is combined with a multi-level feature fusion mechanism to obtain a patent representation vector; a spatiotemporal heterogeneous graph neural network is used for dynamic modeling to obtain a patent graph embedding representation; a multi-task meta-learning framework is used for collaborative evaluation to obtain legal value, technical value and economic value scores; driving factors of patent value are identified to obtain value attribution results with causal explanations; a time sequence neural network is used for dynamic prediction in combination with external technical evolution and market trend data to obtain value evolution trends at multiple time nodes; and a scene-adaptive report generation framework is used in combination with a user-specified application scene to obtain a structured patent value evaluation report and decision suggestion. The application can realize multi-dimensional, dynamic and interpretable intellectual property value evaluation.

Description

Technical Field

[0001] This invention relates to the field of intellectual property valuation technology, and more specifically, to an intellectual property valuation system and method. Background Technology

[0002] With the development of the knowledge economy, intellectual property rights are increasingly becoming a key element of a company's core competitiveness. Patents, as their primary form, hold significant value in commercial activities such as technology acquisitions, pledge financing, and litigation licensing. Accurately assessing patent value is crucial for corporate strategic decision-making and risk management; however, patent valuation involves multiple dimensions, including legal, technological, and market factors, making the process complex and highly dynamic. Existing evaluation methods mainly include traditional approaches such as the cost approach, market approach, and income approach, but each has its limitations: the cost approach ignores the market and future revenue, the market approach is constrained by data opacity and scarcity of cases, and the income approach relies on subjective predictions and parameter settings. None of these methods can comprehensively and accurately reflect the multidimensional value of a patent; In recent years, artificial intelligence technology has been introduced into patent valuation, but it still faces the following problems: the data source is singular and lacks integration of multi-source information; the feature extraction capability is insufficient and it is difficult to deeply explore the semantics of text and network structure; the evaluation dimension is singular and does not distinguish between different value aspects such as legal, technical and economic aspects; the model has poor interpretability and is mostly static evaluation, failing to reflect the dynamic evolution of value; in addition, there is a lack of adaptive adjustments for different application scenarios. Summary of the Invention

[0003] This invention provides an intellectual property value assessment system and method, which solves the technical problems of single dimension of patent value assessment, poor model interpretability, lack of dynamic evolution prediction and lack of scenario adaptability adjustment in related technologies.

[0004] This invention provides a method for assessing the value of intellectual property, comprising the following steps: S1. Based on the patent number and technical field identifier of the target patent, a patent feature dataset is obtained by distributed data collection. S2, obtain the patent feature dataset, and use a pre-trained language model combined with a multi-level feature fusion mechanism to obtain the patent representation vector; S3 receives the patent feature dataset and patent representation vector, and uses a spatiotemporal heterogeneous graph neural network for dynamic modeling to obtain the patent graph embedding representation. S4 receives patent representation vectors and patent graph embedding representations, and uses a multi-task meta-learning framework for collaborative evaluation to obtain scores for legal value, technical value, and economic value. S5 receives scores for legal value, technical value, and economic value, identifies the driving factors of patent value, and obtains value attribution results with causal explanation. S6: Obtain external technology evolution and market trend data, combine patent graph embedding representation with legal value, technical value and economic value scores, and use a time-series neural network to make dynamic predictions to obtain the value evolution trend at multiple time nodes; S7 receives value scores, value attribution results, and value evolution trends. Combined with the user-specified application scenarios, it uses a scenario-adaptive report generation framework to obtain a structured patent value assessment report and decision-making recommendations.

[0005] In a preferred embodiment, S1 includes: The basic attribute information of patents is extracted in batches from the patent database system using an API interface to obtain structured patent metadata; Web crawling technology and text parsing algorithms are used to extract the text and graphic content of patents from the full-text patent database, resulting in multi-level text data representation and graphic data representation. A citation relationship network for the patent is constructed using graph database technology to obtain the forward citation set and backward citation set of the target patent; A data quality assessment model is used to quantify and score the completeness, timeliness, and reliability of each data source and each data field, resulting in a data quality weight vector.

[0006] In a preferred embodiment, S2 includes: Based on the language model parameters pre-trained from a large-scale general patent corpus, model transfer technology is used as the initialization basis for domain coding to obtain a basic encoder with patent language understanding capabilities. Based on the unlabeled patent text set in the target technology field, a self-supervised task of mask language modeling is used to perform domain-adaptive fine-tuning on the basic encoder, resulting in a domain encoder that is sensitive to the technical terms and expression patterns of the target domain. Based on the technical classification labels of patents in the target domain, a contrastive learning strategy is used to discriminately optimize the domain encoder, resulting in a discriminative semantic encoder that can distinguish between different technical branches. Based on the segmented text of the patent specification, a discriminative encoder is used to perform hierarchical semantic encoding to obtain a multi-level semantic representation of the patent text. Based on the semantic vector of the specification, the semantic vector of the claims, and the embedding vector of the technology classification, a multi-head attention mechanism is used to fuse features to obtain the patent representation vector.

[0007] In a preferred embodiment, S3 includes: Based on patent citation data and applicant data, a patent association network is constructed using a heterogeneous graph modeling method, resulting in a heterogeneous graph structure containing multiple node types and edge types, including citation edges, technology similarity edges, competing edges, and family edges. Based on the heterogeneous graph and the temporal attributes of patents, an adaptive time window partitioning method is used to generate a dynamic graph snapshot sequence, thereby obtaining graph sequence data that reflects the temporal evolution of the patent network. Based on the graph snapshot of each time window, a heterogeneous graph attention network is used to encode the spatial topology of the graph, resulting in a node spatial embedding vector that integrates information from multiple types of neighbors. Based on the node spatial embedding vector sequence of each time window, a temporal convolutional network is used to model the temporal evolution pattern of patent nodes, resulting in a patent graph embedding representation that integrates temporal dynamics.

[0008] In a preferred embodiment, S4 includes: Based on data quality weight vector, semantic representation vector, graph embedding vector, and structured feature vector, a quality-aware feature fusion mechanism is adopted to obtain a comprehensive patent feature representation. Based on the fused feature vectors, a shared representation learning network for three evaluation sub-tasks—legal value, technological value, and economic value—is constructed using a multi-task learning architecture, resulting in deep feature representations shared by the tasks. Based on a multi-task network, an uncertainty weighting strategy is adopted to dynamically balance the weights of each subtask in joint training, resulting in a joint loss function for task balance. Based on historical patent evaluation data from multiple technology fields, a meta-learning algorithm is used to train a multi-task network to obtain model initialization parameters with rapid domain adaptation capabilities. A Bayesian neural network algorithm is used to quantify the uncertainty of the value assessment results, and the probability distribution and confidence interval of the value score are obtained.

[0009] In a preferred embodiment, S5 includes: Based on feature data and value annotation data of large-scale historical patents, a causal structure learning algorithm is used to infer the causal dependencies between feature variables and between features and values, and to obtain a causal directed acyclic graph. Based on the learned preliminary causal graph, domain expert knowledge and theoretical analysis are used to revise and verify the causal graph, resulting in a reliable causal relationship model. Based on the causal graph and the feature data of the target patent, the causal inference method is used to calculate the causal effect of each feature variable on the value of the target patent, and the causal contribution quantification results of the features are obtained. Based on the identified causal effects, a counterfactual inference method is used to conduct a value attribution analysis on the target patent, obtaining the specific contribution and attribution path of each key factor.

[0010] In a preferred embodiment, S6 includes: Based on large-scale historical patent data, a time window sliding method is used to construct a time-series evolution training dataset of patent value, and input-output sequence pairs are obtained for time-series model training. Based on technology evolution paths and market development data, time series analysis methods are used to extract the time series characteristics of external driving factors, thereby obtaining time series signals that reflect technology and market dynamics. Based on time-series training data and external time-series features, a sequence-to-series deep learning model is used to train a time-series evolution predictor of patent value, resulting in a time-series model that can predict future value trajectories. Based on the time-series prediction model, the scenario analysis method is used to generate value evolution predictions under different external environment assumptions, and the value prediction trajectories of three scenarios, namely optimistic, neutral and pessimistic, are obtained.

[0011] In a preferred embodiment, S7 includes: Based on the application scenario type identifiers input by users, the parameter configuration of the evaluation framework is determined by the scenario knowledge base matching method, so as to obtain the value weight and evaluation focus of scenario customization; Based on value scoring, attribution analysis, dynamic prediction, and scenario-customized scoring, a structured data integration method is used to obtain a comprehensive evaluation data structure containing multi-level evaluation information. Based on integrated structured assessment data, template matching and natural language generation technologies are used to obtain highly readable text assessment reports; based on the assessment data, data visualization technologies are used to generate various charts to provide an intuitive visualization of the assessment results.

[0012] In a preferred embodiment, the construction of the patent association network using a heterogeneous graph modeling method specifically includes: Create a set of patent nodes, with each patent corresponding to a node in the graph. Node attributes include patent semantic vector, patent application date, and legal status. Create citation edges based on citation relationships, establishing directed edges from cited patents to citing patents. Based on the technology classification, a technology similarity edge is created. For patent pairs with the same first few digits of the technology classification number, the cosine similarity of their semantic vectors is calculated. If the similarity exceeds the preset similarity threshold, an undirected technology similarity edge is established between the two patent nodes. Competitive edges are created based on applicant information. For patent pairs belonging to different applicants but with high overlap in technical fields, competitive edges are established. Family edges are created based on patent family relationships. For patents from different countries belonging to the same patent family, family-related edges are established.

[0013] In a preferred embodiment, an intellectual property valuation system is used to perform the steps of the above-described intellectual property valuation method, including: The data acquisition module is used to obtain a patent feature dataset by using distributed data acquisition based on the patent number and the technical field identifier of the target patent. The semantic representation module is used to obtain the patent feature dataset. It uses a pre-trained language model combined with a multi-level feature fusion mechanism to obtain the patent representation vector. The graph embedding module receives the patent feature dataset and the patent representation vector, and uses a spatiotemporal heterogeneous graph neural network for dynamic modeling to obtain the patent graph embedding representation. The value assessment module receives patent representation vectors and patent graph embedding representations, and uses a multi-task meta-learning framework to conduct collaborative assessments to obtain scores for legal value, technical value, and economic value. The attribution analysis module is used to receive legal value, technical value, and economic value scores, identify the driving factors of patent value, and obtain value attribution results with causal explanations. The trend prediction module is used to acquire external technology evolution and market trend data, and combine it with patent graph embedding representation and legal value, technical value and economic value scores. It uses a time sequence neural network to make dynamic predictions and obtain the value evolution trend at multiple time nodes. The report generation module receives value scores, value attribution results, and value evolution trends. Combined with the user-specified application scenario, it adopts a scenario-adaptive report generation framework to obtain a structured patent value assessment report and decision-making recommendations.

[0014] The beneficial effects of this invention are as follows: By integrating multi-source heterogeneous data such as patent texts, citation networks, legal events, transaction data, and technological evolution through a distributed data acquisition mechanism, and using a data quality assessment model to quantify and score the data's completeness, timeliness, and reliability, a data quality weight vector is obtained. A pre-trained language model combined with domain-adaptive fine-tuning and contrastive learning is used for multi-level semantic encoding, fusing semantic information from the specification, claims, and technical classification. A spatiotemporal heterogeneous graph neural network is employed to model multiple types of relationships in patents, such as citations, technical similarities, competition, and family relationships. A temporal convolutional network is used to capture the dynamic evolution patterns of the network, generating a patent graph embedding representation that integrates spatial topology and temporal dynamics, thereby improving the comprehensiveness and accuracy of feature representation. This study employs a multi-task meta-learning framework to collaboratively evaluate the legal, technological, and economic values. It achieves joint optimization of multi-dimensional value through shared representation learning and task-adaptive weight balancing. Causal structure learning and counterfactual inference methods are used to identify the driving factors of patent value, yielding causal attribution results. Sequence-to-sequence temporal neural networks, combined with external technological evolution and market trend data, predict future value evolution trajectories, and Bayesian neural networks quantify uncertainty. Finally, a scenario-adaptive report generation framework provides customized value weights, collaborative value assessments, and decision-making suggestions for different application scenarios, thus achieving multi-dimensional, dynamic, interpretable, and scenario-adaptive intellectual property value assessment. Attached Figure Description

[0015] Figure 1 This is a flowchart of an intellectual property valuation method according to the present invention; Figure 2 This is a module diagram of an intellectual property valuation system according to the present invention. Detailed Implementation

[0016] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0017] At least one embodiment of the present invention discloses a method for assessing the value of intellectual property, such as... Figure 1 As shown, it includes the following steps: S1. Based on the patent number and technical field identifier of the target patent, a patent feature dataset is obtained by distributed data collection. S11, Structured collection of basic patent data; Based on the patent number of the target patent, basic attribute information of the patent is extracted in batches from the patent database system using an API interface to obtain structured patent metadata. Specifically, by building data interfaces with multiple patent data sources, including the State Intellectual Property Office's patent search system, the European Patent Office's Espacenet system, and the US Patent and Trademark Office's full-text patent database, cross-jurisdictional patent data collection is achieved. For each patent, the extracted basic data includes fields such as patent name, patent number, application number, application date, publication date, grant date, applicant information, inventor information, agency information, current legal status, payment status, patent type, and technology classification system. For international patents, Patent Cooperation Treaty application information and Paris Convention priority information are collected simultaneously to construct patent family relationships. For Chinese patents, multi-level classification information of Chinese patent classification numbers and international patent classification numbers is extracted, and the main classification number and sub-classification number are parsed to construct a hierarchical identifier for the technical subject. For each data field collected, the data acquisition timestamp, data source identifier, and field integrity mark are recorded. The integrity mark includes three levels: completely missing, partially missing, and complete, providing a basis for subsequent data quality assessment.

[0018] S12, Multi-level extraction of patent text content; Based on the patent number obtained in step S11, web crawling technology and text parsing algorithms are used to extract the text and graphic content of the patent from the full-text patent database, resulting in multi-level text and graphic data representations. Specifically, the full text of the patent specification, the full text of the claims, the abstract, and the accompanying drawings are downloaded. For patent documents in different formats, including PDF, HTML, and XML, appropriate text extraction tools are used for content parsing. For PDF patents, optical character recognition technology is used to extract the text content while preserving structured information such as paragraph structure, chapter titles, and figure annotations. The claims are analyzed in detail to identify independent and dependent claims, extract the citation relationships of each claim, and construct a tree-like dependency structure of the claims. The specification text is automatically segmented according to standard chapters such as technical field, background technology, invention content, description of drawings, and specific embodiments, and chapter boundaries are identified using a combination of rule-based and machine learning methods. For the specific embodiments section, key technical content such as method step descriptions, system module descriptions, and algorithm flow descriptions are further extracted, and structured information such as technical terms, parameter names, and numerical ranges are extracted using entity recognition technology. For the patent drawings, image processing technology is used to extract the technical information in the drawings, such as the flowchart structure, module connection relationship, and data flow, and to convert the graphic information into a structured graph data representation.

[0019] S13, Bidirectional construction of the patent citation network; Based on the patent data collected in step S11, a citation relationship network for the patents is constructed using graph database technology, resulting in the forward citation set and backward citation set for the target patent. Specifically, for the target patent, all prior patents cited in the background section of the specification and during the examination process are extracted to form the backward citation set. These patents represent the technological foundation and prior art background of the target patent. For each cited patent, information such as its patent number, citation location, and citation type are recorded. The citation type includes applicant citations and examiner citations; examiner citations typically indicate a stronger technological relevance between the cited patent and the target patent. Simultaneously, all subsequent patents that cited the target patent are retrieved to form the forward citation set. These patents represent the technological influence and technology dissemination path of the target patent. By analyzing indicators such as the number of forward citation patents, the citation time distribution, and the quality of the cited patents, the continuity of the target patent's technological value can be assessed. Furthermore, the citation network is expanded using multi-hop methods, considering not only direct citation relationships but also second-degree and third-degree citation relationships to construct a broader technological evolution network. For each citation edge, the citation strength weight is calculated, taking into account factors such as the number of citations, the importance of the cited patent, and the relevance of the cited text, to provide weighted edge information for subsequent graph neural network modeling.

[0020] S14, Timeline capture of patent legal events; Based on the patent number of the target patent, a multi-source data fusion strategy is employed to extract the legal event sequence of the patent from the patent legal status database, litigation case database, and patent licensing and transfer registration system, obtaining time-series legal status evolution data. Specifically, records of changes in patent rights are extracted from the patent register copy, including legal events such as patent transfer, licensing registration, pledge registration, patent termination, and patent restoration, recording information such as the occurrence time, event type, and involved parties for each event. All litigation cases involving the target patent are retrieved from the patent litigation database, including infringement lawsuits, invalidation requests, and patent ownership disputes, extracting detailed information such as the filing time, litigation type, parties involved, litigation outcome, compensation amount, and court of trial. For invalidation proceedings, process data such as the filing time of the invalidation request, invalidation grounds, examination decision, and re-examination decision are extracted to analyze the legal stability of the patent right. For patents with litigation records, statistical characteristics such as litigation success rate and average compensation amount are further analyzed; these characteristics reflect the enforcement intensity and market value of patent rights. All legal events are arranged chronologically to construct a patent legal event timeline, providing legal-dimensional evolutionary data for subsequent time-series modeling. In cases of missing data, where no litigation records for a patent can be found in certain jurisdictions, the data is marked as missing, but no false data is filled in.

[0021] S15, Similar case matching of patent transaction data; Based on the target patent's technology classification number and applicant information, a similarity matching algorithm is used to retrieve comparable transaction cases from the patent transaction database, resulting in a historical transaction dataset for value reference. Specifically, patent transaction records are collected from channels such as publicly available patent licensing and transfer information, patent auction platforms, and technology trading markets, extracting basic information about the traded patents, transaction prices, transaction times, transaction types, and industry backgrounds of both parties. For the target patent, its similarity to historical transaction patents is calculated in dimensions such as technology classification, number of claims, citation relationships, and patent age. Weighted Euclidean distance or cosine similarity measurement methods are used to screen comparable transaction cases with similarity exceeding a preset threshold. For the selected comparable cases, their transaction prices are extracted and adjusted for time value. Based on the interval between the transaction time and the current valuation time, a technology depreciation model is used to convert the historical transaction prices to present value. Because patent transaction data is often scarce and information is not fully publicly available, for cases where a clear transaction price cannot be obtained, only the fact of the transaction is retained as a qualitative reference and marked as lacking price information. For patents with no comparable transaction cases at all, the data source is marked as entirely missing and processed in subsequent modeling using a multi-view learning mechanism.

[0022] S16, Literature mining of technological evolution paths; Based on the technical theme keywords and technical classification numbers of the target patent, literature retrieval and text mining techniques are used to extract data on technological evolution and market development from academic paper databases, technical report databases, and industry analysis reports to obtain macro-level trend characteristics of the technical field. Specifically, using the claims text and technical solution section of the target patent, keyword extraction algorithms are employed to identify core technical terms and construct a technical theme vocabulary. Using technical theme vocabulary as search keywords, research papers in related technical fields are retrieved from academic databases, extracting information such as publication time, citation count, research topic, and technical route to analyze the changing trends in research popularity of the technical theme. By performing text clustering on paper abstracts and keywords, sub-themes and technical branches within the technical field are identified, constructing an evolutionary map of the technical theme. Macro-level indicators such as market size, growth rate, technological maturity, and competitive landscape of the industry to which the target patent belongs are extracted from industry research reports and market analysis reports; these indicators reflect the market application prospects of the technology. Using time series analysis methods, trend analysis is performed on indicators such as the number of published papers and patent applications for the technical theme to determine the stage of the technology's life cycle, including the nascent stage, growth stage, maturity stage, and decline stage. By associating the technological evolution characteristics obtained from literature mining with the target patents, a macro-feature vector reflecting the technological era is generated.

[0023] S17, Data integrity assessment and quality weight calculation; Based on the multi-source data collected in steps S11 to S16, a data quality assessment model is used to quantitatively score the completeness, timeliness, and reliability of each data source and each data field, resulting in a data quality weight vector. Specifically, for each data source, its data completeness rate is calculated, which is the ratio of the number of non-missing fields to the total number of fields. For data sources with partial missing data, the importance of the missing fields is further analyzed, with higher quality penalties imposed for missing core fields. The timeliness of the data is assessed by calculating the time difference between the data acquisition time and the data generation time. For time-sensitive data, such as legal status and market data, a large time difference will lower the data quality score. The reliability of the data source is assessed by classifying it according to the authority of the data source. Data from official databases is given the highest reliability score, data from third-party data platforms is scored based on the platform's reputation, and unstructured data crawled from the web is given a lower reliability score. Combining the three dimensions of completeness, timeliness, and reliability, a weighted summation method is used to calculate the comprehensive quality score of each data source, normalized to between 0 and 1, forming the quality weight vector. This quality weight vector is used in the subsequent feature fusion stage to weight features from different sources, making the model rely more on high-quality data and reducing the interference of low-quality data on the evaluation results.

[0024] S18, intelligent completion of missing data based on collaborative filtering; Based on the data missing information identified in step S17, collaborative filtering and matrix completion algorithms are used to intelligently fill in the missing data that can be completed, resulting in a patent feature matrix with enhanced integrity. Specifically, the feature data of all patents are organized into a patent feature matrix, where rows represent individual patents, columns represent feature dimensions, and missing values ​​in the matrix correspond to features that were not obtained during the data collection phase. A collaborative filtering method based on similar patents is used. For a missing feature of a patent, several patents most similar to it in terms of known features are found. The values ​​of these similar patents in the missing feature dimension are weighted and averaged to estimate the missing feature value of the target patent. Similarity calculation uses Pearson correlation coefficient or cosine similarity, and is performed only based on non-missing common feature dimensions. For structured numerical features, such as patent age, number of claims, and number of citations, a weighted average method is used for completion. For categorical features, such as technical field and applicant type, a majority voting method is used to select the most frequently occurring category as the completion value. For textual features, such as missing technical descriptions, no automatic completion is performed, and the missing state is maintained. For each completed feature value, a data source marker is added, indicating it as an inferred value rather than a true observation, and its weight in the quality weight vector is reduced. Through this completion process, the integrity of the feature matrix is ​​improved while maintaining data accuracy, providing richer input for subsequent machine learning models.

[0025] S2, obtain the patent feature dataset, and use a pre-trained language model combined with a multi-level feature fusion mechanism to obtain the patent representation vector; S21, Transfer loading of the general patent language model; Based on language model parameters pre-trained on a large-scale general patent corpus, model transfer programming is used as the initialization basis for domain encoding to obtain a basic encoder with patent language understanding capabilities. Specifically, a Transformer architecture language model pre-trained on a corpus of millions of patent texts is selected. This model has already learned the general language patterns, syntactic structures, and basic technical terms of patent texts. The parameters of the word embedding layer, multi-layer self-attention encoding layer, and feedforward neural network layer of the pre-trained model are loaded to retain the model's basic understanding of patent language. A domain adaptation layer, including domain-specific pooling and projection layers, is added to the top layer of the model for subsequent domain fine-tuning. The vocabulary of the pre-trained model is examined, and for professional terms that frequently appear in the target technology field but are not in the original vocabulary, the vocabulary is expanded and the embedding vectors of the newly added words are randomly initialized. A fine-tuning strategy for the model parameters is set: a smaller learning rate is used for the pre-trained layers to retain general knowledge, and a larger learning rate is used for the newly added adaptation layer to quickly learn domain features.

[0026] S22, self-supervised fine-tuning of unlabeled corpora in the target domain; Based on an unlabeled patent text set within the target technology field, a domain-adaptive fine-tuning of the basic encoder in step S21 is performed using a masked language modeling self-supervised task, resulting in a domain encoder sensitive to the technical terms and expression patterns of the target domain. Specifically, all patents belonging to the same international patent classification category as the target patent are retrieved from the patent database, and their specification texts are extracted to construct a domain patent corpus. The text in the corpus undergoes sentence and word segmentation preprocessing, and the text is divided into sub-word units using a byte-pair encoding word segmentation method. A masked language modeling training task is constructed, randomly selecting a certain proportion of words in the text for masking, and training the model to predict the original content of the masked words based on the context. To enhance the model's understanding of technical terms, an entity-aware masking strategy is adopted to increase the probability of technical terms being masked, making the model more focused on the semantic learning of professional vocabulary. During the fine-tuning process, an adaptive learning rate scheduling strategy is adopted, using a larger learning rate in the initial stage to quickly adapt to the domain distribution, and then reducing the learning rate in the later stage to fine-tune the parameters. Through self-supervised fine-tuning on target domain corpora, the model learns the co-occurrence patterns of technical terms specific to that domain, the relationships between technical concepts, and domain-specific expression habits, thereby improving its semantic understanding of patent texts in the target domain.

[0027] S23, Discriminative optimization of technology similarity comparison learning; Based on the technical classification labels of patents in the target domain, a contrastive learning strategy is used to discriminately optimize the domain encoder in step S22, resulting in a discriminative semantic encoder capable of distinguishing different technical branches. Specifically, utilizing the hierarchical structure of the International Patent Classification (IPC), patents belonging to the same subclass are considered positive sample pairs with similar technology, while patents belonging to different major classes are considered negative sample pairs with dissimilar technology. For each patent, several patents are randomly sampled from its same technical subclass as positive samples, and several patents are randomly sampled from different technical major classes as negative samples, constructing training triples for contrastive learning. A contrastive loss function is used, with the optimization objective being to make the semantic vectors of positive sample pairs close in the embedding space and the semantic vectors of negative sample pairs far apart in the embedding space. Cosine similarity is used as the distance metric; for positive sample pairs, the cosine similarity of their vectors is maximized, and for negative sample pairs, their cosine similarity is minimized. A temperature coefficient is introduced to adjust the difficulty of contrastive learning; a smaller temperature coefficient makes the model more sensitive to difficult negative samples. Through comparative learning training, the model not only learns the semantic content of patent texts, but also learns how to distinguish patents with different technical directions in the embedding space, which enhances the discriminative ability of semantic representation, so that patents with similar technologies are clustered in the vector space, while patents with large technical differences are separated in the vector space.

[0028] S24, Layered coding of the entire patent specification; Based on the segmented patent specification text extracted in step S1, a hierarchical semantic encoding is performed using the discriminative encoder optimized in step S23 to obtain a multi-level semantic representation of the patent text. Specifically, the patent specification is divided into multiple paragraphs according to chapters, such as technical field, background technology, invention content, and specific implementation methods, with each paragraph containing several sentences. At the sentence level, each sentence is input into a domain encoder, and the words in the sentence are encoded using a multi-layer self-attention mechanism of the Transformer to extract the encoding state sequence of the last layer of the sentence. An attention pooling mechanism is used to generate sentence-level semantic vectors for the lexical encoding sequence of the sentence. The attention weights are automatically learned based on the contribution of words to the semantics of the sentence, with key technical terms receiving higher weights. At the paragraph level, all sentence vectors within a paragraph are combined into a sequence and input into another layer of the Transformer encoder to model the logical connections and semantic coherence between sentences. The encoding state sequence of the last layer of the paragraph is extracted, and paragraph-level semantic vectors are generated through attention pooling. At the document level, the semantic vectors of all paragraphs are combined into a sequence and input into a document-level Transformer encoder to model the structured relationships between different chapters and extract the global semantic representation of the document. Through a three-layer hierarchical encoding, semantic information is aggregated layer by layer from words to sentences to paragraphs to documents. This not only preserves the semantics of local technical details but also captures the overall technical theme of the document, generating a comprehensive semantic representation of the patent specification.

[0029] S25, Structural analysis of the technical features of the claims; Based on the claim text extracted in step S1, dependency parsing and technical feature extraction algorithms are used to perform structured parsing of the claims, resulting in a set of technical features and a logical structure representation of the claims. Specifically, syntactic analysis is performed on each claim to identify the claim's subject name, list of technical features, and logical relationships between features. Claims typically use a standard format of "...characterized by...", and the boundaries between the subject and feature parts are identified through rule matching. The feature parts are further broken down, and each technical feature item is identified based on punctuation and logical connectors. For each technical feature item, named entity recognition technology is used to extract technical entities, including structured elements such as component names, action verbs, attribute adjectives, and parameter values. A dependency tree of technical features is constructed to represent the modification and combination relationships between features. For dependent claims, the independent claim numbers they reference are parsed to construct a reference dependency graph between claims. The text of each technical feature item is input into the domain encoder in step S23 to generate a semantic vector of the technical features. All technical feature vectors of a claim are aggregated, and a weighted average or max pooling method is used to generate a semantic vector at the claim level. For independent claims and dependent claims, their feature vectors are calculated separately, and they are hierarchically organized according to the reference relationships to generate a structured semantic representation of the claims.

[0030] S26, Hierarchical embedding of technical classification numbers; Based on the International Patent Classification (IPC) numbers extracted in step S1, a hierarchical embedding method is used to encode the classification numbers into vector representations, resulting in a structured representation of the patent's technical subject. Specifically, the IPC adopts a five-level hierarchical structure: department, major category, minor category, main group, and subgroup, with each level representing a different granularity of technical classification. An independent embedding dictionary is established for each level of the classification number, mapping the classification symbols to embedding vectors. For the main classification number of a patent, its five-level classification symbols are extracted, and embedding vectors at each level are found and concatenated in order from coarse to fine to form the hierarchical embedding vector of the classification number. This vector encodes the complete classification path of the patent from macro-technical field to micro-technical direction. For multiple sub-classification numbers of a patent, the same method is used to generate their respective hierarchical embedding vectors. Then, an attention mechanism is used to weighted aggregate the multiple classification vectors to generate a comprehensive technical classification representation of the patent. Through hierarchical embedding, not only is the precise technical orientation of the classification number preserved, but the hierarchical structure also implicitly expresses the similarity relationship between different technical categories. Patents belonging to the same major category but different minor categories are relatively close in the embedding space, while patents belonging to completely different major categories are far apart in the embedding space.

[0031] S27, Attention fusion of multi-level semantic features; Based on the specification semantic vector generated in step S24, the claim semantic vector generated in step S25, and the technical classification embedding vector generated in step S26, a multi-head attention mechanism is used for feature fusion to obtain the patent representation vector. Specifically, the three types of semantic vectors are used as input to the attention mechanism. A query-key-value attention calculation framework is adopted, using the claim vector as the query vector and the specification vector and classification vector as the key and value vectors, respectively. The attention weights of the claims on different information sources are calculated to generate a comprehensive representation that integrates specification details and classification themes. The multi-head attention mechanism sets up multiple parallel attention heads, each focusing on a different semantic subspace to capture diverse semantic association patterns. The output vectors of multiple attention heads are concatenated and passed through a linear transformation layer to generate a fusion vector with a unified dimension. During the fusion process, residual connections and layer normalization operations are introduced to maintain training stability and accelerate convergence. Through the attention fusion mechanism, the model can automatically learn the contribution of different semantic sources to the final representation. For patents with detailed technical solutions, specification semantics receives higher weight; for patents with concise claims, claim semantics receives higher weight; and for patents with clear technical themes, classification information receives higher weight. The final generated patent representation vector integrates multi-level semantic information of the patent text, including both macro-level technical theme positioning and micro-level technical solution details, providing rich semantic features for subsequent value assessment.

[0032] S3 receives the patent feature dataset and patent representation vector, and uses a spatiotemporal heterogeneous graph neural network for dynamic modeling to obtain the patent graph embedding representation. S31, Heterogeneous graph construction with multiple types of relationships; Based on the patent citation data and applicant data extracted in step S1, a heterogeneous graph modeling method is used to construct a patent association network, resulting in a heterogeneous graph structure containing various node and edge types. Specifically, a set of patent nodes is created, with each patent corresponding to a node in the graph. Node attributes include the patent representation vector generated in step S2, patent application date, legal status, etc. Citation edges are created based on citation relationships. For each pair of patents with a citation relationship, a directed edge is established from the cited patent to the citing patent. Edge attributes include citation type, citation time, etc. Technical similarity edges are created based on technology classification. For patent pairs with the same first few digits of their technology classification numbers, the cosine similarity of their semantic vectors is calculated. If the similarity exceeds a preset threshold, an undirected technical similarity edge is established between the two patent nodes, with the edge weight being the similarity value. Competitive edges are created based on applicant information. For patent pairs belonging to different applicants but with high overlap in technical fields, competitive relationship edges are established, representing market and technological competition. Family-related edges are created based on patent family relationships. For patents belonging to the same patent family but from different countries, family-related edges are established, representing the distribution of the same invention in different jurisdictions. Through the above steps, a heterogeneous graph was constructed, with node types including patents and edge types including citation edges, technology similarity edges, competition edges, and family edges. Each edge type has a different impact mechanism on patent value: citation edges reflect technological influence, technology similarity edges reflect technological positioning, competition edges reflect market competition, and family edges reflect internationalization. These multidimensional relationships are modeled uniformly through the heterogeneous graph.

[0033] S32, Adaptive time-granular dynamic graph snapshot generation; Based on the heterogeneous graph constructed in step S31 and the temporal attributes of the patents, an adaptive time window partitioning method is used to generate a dynamic graph snapshot sequence, resulting in graph sequence data reflecting the temporal evolution of the patent network. Specifically, the granularity of the time window is adaptively determined according to the technological update speed of the technical field to which the target patent belongs. For rapidly iterating technical fields, such as artificial intelligence and mobile communications, a quarterly time window is used; for slowly evolving technical fields, such as traditional machinery and chemical materials, an annual time window is used. The technological update speed is quantified by analyzing the application time density and citation time interval statistical characteristics of historical patents in that field. The entire time span is divided into several time windows according to the determined time granularity. For each time window, a graph snapshot for that time period is generated. The generation rule for the graph snapshot is that the node set contains all patents with application dates before the time window, and the edge set contains newly added citation edges within the time window and other types of edges that are still valid within the time window. By sliding the time window, a series of graph snapshots are generated, each snapshot reflecting the patent network state for a specific time period. The structural change rate between adjacent time window graph snapshots is calculated, including metrics such as the number of newly added nodes, the number of newly added edges, and changes in edge weights. These change rate metrics reflect the activity level of technological evolution. The graph snapshot sequence and change rate metrics are used as a temporal representation of the dynamic graph, providing input for subsequent spatiotemporal graph neural networks.

[0034] S33, Spatial feature extraction using heterogeneous graph attention networks; Based on the graph snapshots generated in step S32 for each time window, a heterogeneous graph attention network is used to encode the spatial topology of the graph, obtaining a node spatial embedding vector that integrates information from multiple types of neighbors. Specifically, for each patent node in the graph, different types of neighbor node information are aggregated according to the edge type. For citation-type edges, all citing neighbors and cited neighbors of the target node are extracted, forming a forward neighbor set and a backward neighbor set, respectively. For each neighbor node, its attention weight to the target node is calculated, based on the semantic vector similarity between the target node and neighbor nodes, as well as the edge weight. A multi-head attention mechanism is adopted, setting multiple attention heads to learn different neighbor importance patterns. The feature vector of each neighbor node is multiplied by its attention weight and then summed to obtain the aggregated neighbor features of that type of edge. For other types of edges, such as technology-similar edges, competing edges, and family edges, the same attention aggregation method is used to generate aggregated features for each type of edge. The aggregated features of different types of edges are concatenated with the semantic vector of the target node itself, and a nonlinear transformation layer is used to generate the spatial embedding vector of the target node. By stacking multiple heterogeneous graph attention networks, the node embedding vectors can aggregate information from multi-hop neighbors, capturing a wider range of network structural features. Ultimately, each patent node obtains a spatial embedding vector that integrates information from multiple types of neighbors and the multi-hop network structure. This vector not only contains the semantic features of the patent itself, but also its positional and relational features within the technology network.

[0035] S34, Dynamic evolution modeling of temporal graph convolutional networks; Based on the node spatial embedding vector sequence for each time window generated in step S33, a temporal convolutional network is used to model the temporal evolution pattern of patent nodes, resulting in a patent graph embedding representation. Specifically, for each patent node, its spatial embedding vectors in each time window are extracted and arranged in chronological order to form the temporal feature sequence of the node. For a new patent that appears only in a certain time window, its embedding vectors before that time window are set as zero vectors or estimated using interpolation methods. A one-dimensional temporal convolutional network is used, with the convolution kernel sliding along the time dimension to extract local patterns and trend features of the time series. The temporal convolutional network adopts a causal convolutional structure, ensuring that the output at the current moment depends only on the input at the current and previous moments, conforming to temporal causality. By stacking multiple layers of temporal convolutions, the receptive field is expanded, enabling the model to capture long-term temporal dependencies. At the top layer of the temporal convolutional network, a temporal attention mechanism is used to assign different weights to the features of different time windows. Recent time windows usually receive higher weights because they are more relevant to the current value assessment. By weighted aggregation of the convolutional features of each time window, the temporal aggregated features of the node are generated. The temporal aggregation features are concatenated with the spatial embedding vectors of nodes in the latest time window, and a patent graph embedding representation is generated through nonlinear transformation. This patent graph embedding representation contains both the spatial location information of nodes in the current network and the dynamic trend information of nodes evolving over time, providing rich network features for evaluating the current value and future potential of patents.

[0036] S35, Adaptive learning of importance weights for graph structures; Based on the patent graph embedding representation generated in step S34, a graph-level attention pooling mechanism is used to learn the importance weights of different nodes and edge types to the value of the target patent, resulting in the importance distribution of the global graph structure. Specifically, for the target patent node, a subgraph centered on it is constructed, containing the target node and its multi-hop neighbor nodes. The importance score of each node in the subgraph for the value assessment of the target node is calculated. Using the graph attention pooling method, the patent graph embedding representation of each node is mapped to an importance score through a learnable weight matrix. The scores are normalized to obtain the importance weight of each node. The importance weight reflects the degree of influence of different neighboring patents on the value of the target patent; high-value cited patents or those cited by high-value patents will receive higher importance weights. For different types of edges, the contribution of each type of edge to the value assessment is calculated separately, and the attention calculation is performed through learnable edge type embedding vectors. The importance weights of different types of edges, such as cited edges, technology-similar edges, and competing edges, are adaptively learned during the training process. The model automatically adjusts the weights of each type of edge based on the feedback of value labels in the training data. The learned node importance weights and edge type weights are applied to the neighbor aggregation process in step S33 to achieve importance-aware graph feature extraction. Through an adaptive learning mechanism, the model can not only capture graph structure information but also identify which structural patterns are most critical for patent value assessment, thus improving the relevance and discriminativeness of graph representation.

[0037] S4 receives patent representation vectors and patent graph embedding representations, and uses a multi-task meta-learning framework for collaborative evaluation to obtain scores for legal value, technical value, and economic value. S41, Quality-perceived fusion of multi-source features; Based on the data quality weight vector output in step S1, the patent representation vector output in step S2, the patent diagram embedding representation output in step S3, and the structured feature vector, a quality-aware feature fusion mechanism is employed to obtain a comprehensive patent feature representation. Specifically, the structured features are first encoded. These structured features include numerical features such as patent age, number of claims, number of independent claims, number of pages in the specification, number of figures, number of cited patents, number of cited patents, number of patent families, and total number of patents held by the applicant, as well as categorical features such as applicant type, legal status, and whether there is a litigation record. The numerical features are normalized to eliminate the dimensional differences between different features. The categorical features are embedded and encoded, mapping the categories to low-dimensional continuous vectors. The encoded structured feature vector, the patent representation vector from step S2, and the patent diagram embedding representation from step S3 are concatenated to form a preliminary comprehensive feature vector. According to the quality weights of each data source output in step S1, different parts of the feature vector are assigned quality weights. A weighted transformation method is used, multiplying the feature components corresponding to high-quality data sources by a larger weight coefficient and multiplying the feature components corresponding to low-quality data sources by a smaller weight coefficient. A multilayer perceptron is used to perform a nonlinear transformation on the weighted feature vectors, learning the interaction relationships between features to generate a fused feature representation. During the fusion process, a feature importance learning module is introduced, employing a feature selection attention mechanism to automatically learn the contribution of different feature dimensions to value assessment, assigning greater weight to important features and less weight to redundant features. The final fused feature vector integrates information from text semantics, network topology, structured attributes, and other aspects, and is adaptively weighted according to data quality, providing reliable feature input for subsequent value assessment.

[0038] S42, Construction of Multi-Task Shared Representation Learning Network; Based on the fused feature vectors generated in step S41, a multi-task learning architecture is used to construct shared representation learning networks for three assessment sub-tasks: legal value, technological value, and economic value, resulting in shared deep feature representations for each task. Specifically, a multi-task neural network is designed, comprising a shared low-level feature extraction module and a task-specific top-level prediction module. The shared module employs a multi-layer fully connected neural network to abstract and learn representations of the fused features layer by layer, extracting common features useful across all value dimensions. The parameters of the shared module are updated simultaneously during the training of the three sub-tasks, learning more robust and generalized feature representations through joint optimization across multiple tasks. At the top layer of the shared module, task-specific modules are constructed for each of the three sub-tasks: legal value assessment, technological value assessment, and economic value assessment. Each task-specific module contains several layers of fully connected networks, specifically learning the discriminative features for that task. The input to the task-specific module is the output feature of the shared module, and the output is the value score prediction for that task. Through this architecture design of sharing the low-level and separating the high-level, the model can learn task-specific knowledge while sharing common knowledge between tasks, leveraging the correlation between tasks to improve the learning performance of each task. For example, patents with high legal value often also have high technical value. Multi-task learning can take advantage of this positive correlation. When there is little labeled data for a certain task, data from other tasks can be used to assist in learning.

[0039] S43, a dynamic balancing mechanism for task-adaptive weights; Based on the multi-task network constructed in step S42, an uncertainty weighting strategy is used to dynamically balance the weights of each sub-task in joint training, resulting in a task-balanced joint loss function. Specifically, in multi-task learning, the training difficulty and data volume of different sub-tasks may vary. Simply summing the loss functions of each task with equal weights can easily lead to some tasks dominating the training process while others are under-learned. To address the task imbalance problem, a learnable task uncertainty weight parameter is introduced for each sub-task. This parameter represents the inherent uncertainty of the task; tasks with high uncertainty should receive smaller weights, and tasks with low uncertainty should receive larger weights. The joint loss function is designed as a weighted sum of the losses of each task, with the task weights being a function of the task uncertainty parameter. During training, the task uncertainty parameter and network parameters are optimized simultaneously, and the model automatically learns appropriate task weight configurations. When the prediction error of a task remains large, its uncertainty parameter increases, causing the weight of that task in the joint loss to decrease, preventing the gradient of that task from excessively affecting the shared parameters. When the prediction error of a task is small and stable, its uncertainty parameter decreases, and the weight of that task in the joint loss increases, fully utilizing the supervision signal of that task. Through a dynamic balancing mechanism, multi-task learning can adaptively adjust the training emphasis of each task, thereby coordinating the learning progress of each task and improving the joint optimization effect.

[0040] S44, rapid cross-domain adaptation based on meta-learning; Based on historical patent evaluation data from multiple technical fields, a multi-task network in step S42 is trained using a meta-learning algorithm to obtain model initialization parameters with rapid domain adaptation capabilities. Specifically, the goal of meta-learning is to learn a good model initialization, such that these initialization parameters can quickly adapt to new technical fields with only minor adjustments using a small amount of labeled data. Data from multiple technical fields are organized into a meta-learning task set, with each technical field corresponding to a meta-task. During the meta-training phase, several domains are sampled from the task set. For each sampled domain, the data is divided into a support set and a query set. The support set is used for model adaptation, and the query set is used to evaluate the performance after adaptation. A model-independent meta-learning algorithm is used. For each sampled domain, the model parameters are first updated using several gradient steps with the support set data to obtain parameters adapted to that domain. Then, the loss is calculated on the query set, and this loss serves as the meta-objective function. The goal of meta-optimization is to find a set of initialization parameters that minimizes the loss on the query set after a small number of gradient steps. Through meta-training in multiple domains, the model learns cross-domain general feature extraction capabilities and a rapid adaptation strategy. In practical applications, when it is necessary to evaluate patents in a new technology field, the initial parameters obtained by meta-learning can be used as a starting point. Only a small number of labeled samples in the field are needed for fine-tuning, which can quickly build a value evaluation model for the field, avoiding the data requirements and time costs of training from scratch.

[0041] S45, Uncertainty quantification of Bayesian neural networks; Based on the multi-task model trained in step S44, a Bayesian neural network method is used to quantify the uncertainty of the value assessment results, obtaining the probability distribution and confidence interval of the value score. Specifically, in the neural network constructed in step S42, the deterministic network weight parameters are replaced with probability distributions. Each weight parameter is no longer a fixed value, but a probability distribution, typically modeled using a Gaussian distribution. During forward propagation, specific weight values ​​are sampled from the weight distribution, and the network is used to calculate the predicted output. Through multiple sampling and forward propagation, multiple predicted output samples are obtained, which constitute the probability distribution of the predicted results. The mean of the predicted distribution is calculated as a point estimate of the value score, and the standard deviation of the predicted distribution is calculated as a measure of uncertainty. A variational inference method is used to train the Bayesian neural network, optimizing the parameters of the weight distribution by minimizing the variational lower bound. In actual assessment, multiple forward samplings are performed on the target patent to obtain score sample sets for legal value, technical value, and economic value, and the mean and confidence interval of each value dimension are calculated respectively. Confidence intervals are typically calculated using several percentiles of the predicted distribution; for example, the 2.5th percentile and the 97.5th percentile could form a 95% confidence interval. The uncertainty quantified using Bayesian methods reflects the model's confidence in the evaluation results. When the input data is of high quality and the model is well-trained, the uncertainty is low and the confidence interval is narrow. Conversely, when there are many missing input data points or a significant difference between the target patent and the training data distribution, the uncertainty is high and the confidence interval is wide, providing decision-makers with a basis for risk assessment.

[0042] S46, Output and interpretation of multidimensional value scoring; Based on the mean value score and confidence interval obtained in step S45, a multidimensional value assessment result for the patent is generated using a structured output format, resulting in a value report containing a score, confidence level, and grade. Specifically, for the three dimensions of legal value, technical value, and economic value, the mean score, lower bound of the confidence interval, upper bound of the confidence interval, and standard deviation of uncertainty are output respectively. Continuous score values ​​are mapped to discrete value grades, for example, divided into five grades: high value, medium-high value, medium value, medium-low value, and low value. The threshold for grade division is determined based on the distribution statistics of the training data. The assessment confidence level is calculated based on the width of the confidence interval; a narrower confidence interval indicates a higher assessment confidence level, and a wider confidence interval indicates a lower assessment confidence level. The confidence level is quantified as a percentage value. A structured representation of the value assessment result is generated, including fields such as value type, mean score, confidence interval, value grade, and confidence level. To enhance the interpretability of the results, additional explanations of the scoring are provided. For example, the legal value score reflects the legal stability and enforcement of the patent right; the technological value score reflects the technological innovation and influence of the patent; and the economic value score reflects the market application potential and economic profitability of the patent. This structured and interpretable output makes the evaluation results easy to understand and apply, supporting subsequent decision analysis and report generation.

[0043] S5 receives scores for legal value, technical value, and economic value, identifies the driving factors of patent value, and obtains value attribution results with causal explanation. S51, Learning the causal structure of historical data; Based on large-scale historical patent feature data and value-annotated data, a causal structure learning algorithm is used to infer the causal dependencies between feature variables and between features and values, resulting in a causal directed acyclic graph (DAG). Specifically, a historical patent dataset is collected, containing multi-dimensional feature vectors and known value scores for each patent. Feature variables include patent text features, network topology features, and structured attribute features. A constraint-based causal discovery algorithm is employed to identify causal relationships between variables through conditional independence tests. For any two variables, their independence is tested under given subsets of other variables. If they are not independent under all conditional sets, a causal edge may exist between the two variables. By performing independence tests on all variable pairs, a preliminary causal graph structure is constructed. Since purely data-driven causal discovery may produce edges that do not conform to domain common sense, domain prior knowledge is introduced to constrain the causal graph. For example, patent application time cannot be causally affected by other features, and value scores, as the final result variable, cannot causally affect other features. Based on prior knowledge, unreasonable causal edges are removed, and the structure of the causal graph is corrected. For causal edges whose direction cannot be determined through independence tests, a score-based structure learning method is employed. The goodness of fit of different causal graph structures is evaluated using scoring functions such as the Bayesian information criterion, and the structure with the highest score is selected. This results in a causal directed acyclic graph (DAG), where nodes represent feature variables and value variables, directed edges represent causal relationships, and the direction of the edges indicates the direction of causal influence. This causal graph reveals the mechanism of patent value generation, identifying which features directly affect value and which features indirectly affect value through mediating variables, providing a structural foundation for subsequent causal inference.

[0044] S52, Correction and verification of causal graphs of domain knowledge; Based on the preliminary causal graph learned in step S51, the causal graph is revised and verified using domain expert knowledge and theoretical analysis to obtain a reliable causal relationship model. Specifically, experts in the field of intellectual property valuation are invited to review the learned causal graph. Experts judge the rationality of causal edges based on valuation theory and practical experience. Edges that experts believe cannot have a causal relationship are deleted from the causal graph, while causal edges that experts believe should exist but were not learned in the data are added to the causal graph. For example, expert knowledge indicates that the number of patent citations causally affects technological value, and technological value causally affects economic value; these causal paths should be reflected in the graph. For edges with uncertain causal directions, judgment is made based on chronological order; variables that occur earlier causally affect variables that occur later. For example, if the patent application date precedes the grant date, the application date may affect the grant date, but the converse is not true. The revised causal graph is verified through counterfactual testing. Intervention predictions are made using the causal graph, and the prediction results are compared with actual observation data or expert judgments. If the prediction is reasonable, the causal graph is reliable; if the prediction is unreasonable, further revision is performed. After multiple rounds of revision and verification, a causal graph was obtained that conforms to both statistical data patterns and domain knowledge. This causal graph can reliably represent the causal generation mechanism of patent value and provide a credible causal model for value attribution.

[0045] S53, Identification of causal effects of the target patent; Based on the causal graph obtained in step S52 and the feature data of the target patent, a causal inference method is used to calculate the causal effect of each feature variable on the value of the target patent, resulting in a quantified causal contribution of the features. Specifically, for each feature variable of the target patent, the average causal effect of that feature on the value is calculated through intervention using the do operator. The do operator intervention involves artificially setting the value of a feature variable, cutting off all causal input edges of that variable, and retaining only the causal output edges, simulating the impact on the value when the feature is assumed to take a specific value. When calculating the causal effect, the expected differences in value scores under different intervention values ​​are compared. For example, for the feature of citation count, the difference between the expected value when the citation count is assumed to be high and the expected value when the citation count is assumed to be low is calculated; this difference is the average causal effect of citation count on the value. Since the do operator intervention requires a causal model, a causal model is trained using the causal graph and historical data from step S52, and the causal graph is parameterized using a structural equation model or a causal Bayesian network. For the target patent, based on the causal model and the observed characteristic values ​​of the target patent, the expected value under different interventions is calculated, and the numerical estimate of the causal effect is obtained through Monte Carlo sampling or analytical calculation. The causal effect of all characteristic variables is calculated to obtain the ranking of the causal contribution of each characteristic to the value, and the key factors with the greatest impact on the value are identified.

[0046] S54, Value Attribution Analysis of Counterfactual Inferences; Based on the causal effects identified in step S53, a counterfactual inference method is used to conduct a value attribution analysis of the target patent, obtaining the specific contribution and attribution path of each key factor. Specifically, counterfactual inference answers the question: how would the value of the target patent change if a certain feature of the target patent were to take different values? For the identified key factors, counterfactual scenarios are constructed, such as assuming the number of citations of the target patent increases by a certain number, or assuming the target patent has no litigation record, and the predicted value of the patent under these counterfactual scenarios is calculated. Counterfactual prediction utilizes a causal model. According to the structure of the causal graph, when a feature is counterfactually intervened, the influence propagates along the causal path, updating all affected downstream variables, and finally obtaining the counterfactual predicted value of the value variable. The difference between the counterfactual predicted value and the actual predicted value is calculated; this difference is the attribution contribution of the factor to the value. Counterfactual analysis is performed on multiple key factors separately to quantify the contribution of each factor. Further analysis of the interactions between factors is conducted. For factor pairs with causal paths, it is analyzed how a factor affects the value through mediating factors, identifying the complete causal attribution chain. For example, high technological innovation in a patent leads to a large number of citations, which in turn leads to high technological value—a complete attribution chain. Analyzing this attribution chain not only reveals which factors are important but also how these factors function, providing actionable suggestions for value enhancement.

[0047] S55, Reliability assessment of attribution results; Based on the attribution analysis results obtained in step S54 and the assessment uncertainty output in step S4, sensitivity analysis is used to evaluate the reliability of the attribution results and obtain an attribution confidence index. Specifically, the causal inference results depend on the correctness of the causal graph structure and causal model parameters. If there are errors in the causal graph or the parameter estimates are inaccurate, the attribution results may be unreliable. Sensitivity analysis is performed on the key edges of the causal graph to test the degree of change in the attribution results when a causal edge is deleted or added. If the attribution results are not sensitive to changes in the causal graph structure, the attribution conclusion is robust; if the attribution results change significantly, the attribution conclusion depends on the causal graph assumption and has low reliability. Perturbation analysis is performed on the parameters of the causal model. Random sampling is performed within the confidence interval of the parameters, and the attribution results are recalculated. The distribution of the attribution results is observed. If the distribution is concentrated, the attribution is reliable; if the distribution is dispersed, the attribution uncertainty is high. Combined with the assessment uncertainty output in step S4, when the uncertainty of the value assessment itself is large, the reliability of the attribution analysis also decreases accordingly, because the attribution is based on the value assessment results. By comprehensively analyzing sensitivity and assessing uncertainty, a reliability indicator is assigned to each attribution factor. High confidence factors indicate reliable attribution conclusions, while low confidence factors indicate uncertainties requiring careful consideration. The attribution report presents both the attribution results and confidence levels, allowing users to understand the reliability of the attribution conclusions and avoid making incorrect decisions based on unreliable attributions.

[0048] S6: Obtain external technology evolution and market trend data, combine patent graph embedding representation with legal value, technical value and economic value scores, and use a time-series neural network to make dynamic predictions to obtain the value evolution trend at multiple time nodes; S61, Construction of historical data on the patent lifecycle; Based on large-scale historical patent data, a time-series evolution training dataset for patent value is constructed using a sliding time window method, resulting in input-output sequence pairs for training the time-series model. Specifically, for each patent in the historical database, feature snapshots and value annotations are extracted from multiple time points within the time span from the application date to the present. The interval between time points is determined according to the update cycle of the technical field; quarterly or semi-annual intervals are used for rapidly iterating fields, and annual intervals are used for slowly evolving fields. For each time point, observable features of the patent at that time are extracted, including the cumulative number of citations, the legal status at that time, and the network location features at that time. For value annotation, if actual value data exists for that time point, such as transaction price or pledge amount, the actual data is used as the annotation; if no actual data exists, the model from step S4 is used to back-evaluate the features of that time point to obtain the value annotation. Time-series training samples are constructed, with each sample containing the feature sequence and value sequence within a historical time window of a patent as input, and the value sequences of several time points after that window as the prediction target. For example, the input is the feature and value sequences of the patent for the first 3 years, and the prediction target is the value of the patent in the 4th and 5th years. By using a sliding time window, multiple training samples are extracted from the lifecycle of each patent to construct a large-scale time series prediction training dataset.

[0049] S62, Temporal feature extraction of external driving factors; Based on the technology evolution path and market development data collected in step S1, time series analysis is used to extract the temporal characteristics of external driving factors, obtaining time series signals reflecting technology and market dynamics. Specifically, for the technical field to which the target patent belongs, indicators such as the number of published papers and patent applications in that field over the years are extracted from the technical literature database to construct a time series of technology activity. A trend decomposition method is used to decompose the time series into trend, cyclical, and random components. The trend component reflects the long-term development direction of the technology, while the cyclical component reflects the periodic fluctuations in technology popularity. Technology maturity indicators are calculated, and by analyzing changes in the growth rate of patent applications, the life cycle stage of the technology is identified. Rapid growth in applications indicates that the technology is in the growth stage, slowing growth indicates that the technology has entered the maturity stage, and a decline in applications indicates that the technology has entered the decline stage. Historical time series of indicators such as market size, market growth rate, and market concentration in that technical field are extracted from market research reports to reflect dynamic changes in market demand. Time series data such as patent application trends and market share changes of major competitors are extracted from competitive intelligence data to reflect the evolution of the competitive landscape. By aligning these external time-series data with the internal time-series features of the patents, a multivariate time series containing internal and external driving factors is constructed, providing rich contextual information for time-series prediction models.

[0050] S63, Sequence-to-Sequence Value Evolution Model Training; Based on the temporal training data constructed in step S61 and the external temporal features extracted in step S62, a sequence-to-sequence deep learning model is used to train a temporal evolution predictor of patent value, resulting in a temporal model capable of predicting future value trajectories. Specifically, an encoder-decoder architecture is adopted. The encoder receives the patent feature sequence and value sequence within a historical time window, while the decoder generates the value prediction sequence for future time nodes. The encoder employs a long short-term memory network or gated recurrent units, capable of capturing long-term temporal dependencies and processing temporal patterns in the historical sequence. The encoder's input at each time step contains a concatenated vector of the patent's internal features and external driving factors at that time point. Through iterative computation, the encoder encodes the historical sequence into a fixed-dimensional context vector, which summarizes the historical information. The decoder also employs a long short-term memory network, with its initial state initialized by the encoder's context vector. At each decoding time step, the decoder receives the predicted value from the previous time step and the external driving factors for the current time step, outputting the value prediction for the current time step. An attention mechanism is introduced, enabling the decoder to focus on the most relevant time steps in the historical sequence when predicting each future time step, dynamically weighting historical information. The training objective is to minimize the mean squared error between the predicted value and the actual future value. During training, a teacher-forced strategy is employed, where the decoder input uses real historical values ​​instead of predicted values, accelerating model convergence. After training, the model is able to predict the future value evolution trajectory of a patent based on its historical evolution patterns and changes in the external environment.

[0051] S64, Scenario analysis for multi-scenario prediction; Based on the time-series prediction model trained in step S63, a scenario analysis method is used to generate value evolution predictions under different external environment assumptions, resulting in value prediction trajectories for optimistic, neutral, and pessimistic scenarios. Specifically, future technological evolution and market development are uncertain, and a single prediction trajectory cannot fully reflect possible future states; therefore, multiple scenarios need to be constructed for prediction. Three typical scenarios are defined: the optimistic scenario assumes continued rapid development in the technological field, strong market demand growth, and a favorable competitive environment; the neutral scenario assumes stable development of technology and market according to current trends; and the pessimistic scenario assumes technological bottlenecks, shrinking market demand, and intensified competition. For each scenario, corresponding future values ​​of external driving factors are set; for example, the technological activity index and market growth rate index are set to high values ​​in the optimistic scenario and low values ​​in the pessimistic scenario. The historical feature sequence and current state of the target patent are input into the time-series prediction model in step S63. In the decoding stage, the external driving factors of the three scenarios are input respectively, generating three different value evolution prediction curves. The forecast curve for the optimistic scenario represents the upper limit trend of patent value under the most favorable conditions, the forecast curve for the pessimistic scenario represents the lower limit trend of patent value under unfavorable conditions, and the forecast curve for the neutral scenario represents the most likely value evolution path. Through multi-scenario forecasting, decision-makers are provided with the range of value evolution, helping to assess investment risks and formulate response strategies.

[0052] S65, Time propagation modeling of prediction uncertainty; Based on the value prediction trajectory generated in step S64 and the current assessment uncertainty output in step S4, the uncertainty propagation method is used to quantify the uncertainty of future predictions, obtaining the confidence intervals for value predictions at each time point. Specifically, the uncertainty of value prediction stems from two aspects: the uncertainty of the current value assessment itself and the prediction error of the time series prediction model. The uncertainty of the current value is quantified by the Bayesian neural network in step S4, manifested as the probability distribution of the current value score. The uncertainty of the time series prediction model is quantified through ensemble learning or Monte Carlo Dropout methods. Multiple time series prediction models are trained, or multiple random Dropout samplings are performed during prediction to obtain multiple prediction trajectory samples. The distribution of these samples reflects the prediction uncertainty. The uncertainty of the current value is used as the initial conditional uncertainty for time series prediction. Through dynamic propagation of the time series model, the uncertainty at each future time point is calculated. Uncertainty increases with the increase of the prediction time span due to the accumulation of more prediction errors and uncertainties caused by environmental changes. For each future time point, Monte Carlo simulation is used to sample values ​​from the current value distribution as initial values ​​and from the distribution of external driving factors as inputs. The sequential prediction model is run to obtain a value prediction sample for that time point. This sampling is repeated multiple times to obtain a sample distribution of value predictions. The mean of the sample distribution is calculated as the expected value prediction for that time point, and the percentiles of the sample distribution are calculated to obtain confidence intervals; for example, the 5th percentile and the 95th percentile constitute a 90% confidence interval. Through uncertainty propagation modeling, the prediction results not only include a point estimate of value but also an uncertainty interval that varies over time, allowing decision-makers to understand the reliability and risk range of the prediction.

[0053] S66, Visual output of value evolution trends; Based on the multi-scenario prediction trajectory in step S64 and the confidence interval in step S65, time series visualization technology is used to generate a value evolution trend chart, providing an intuitive display of dynamic value prediction. Specifically, a two-dimensional coordinate system is constructed with time as the horizontal axis and value score as the vertical axis to plot the historical value trajectory and future value prediction trajectory of the patent. The historical trajectory is represented by solid lines, drawn based on historical actual data or retrospective evaluation data. The future prediction trajectory is represented by dashed lines, distinguishing between optimistic, neutral, and pessimistic scenarios, each using different colors or line types. Confidence interval bands are drawn around the prediction trajectory to represent the range of prediction uncertainty; the width of the confidence interval band increases with time, intuitively reflecting the growth of prediction uncertainty. Key time nodes, such as the expected patent grant date, the expiration date of the protection period, and technology upgrade nodes, are marked to help understand the driving factors of value changes. Trend curves of external driving factors, such as technology maturity curves and market size curves, are added to the chart and compared with the value curve to show the impact of the external environment on value. An interactive visualization interface is generated, allowing users to adjust scenario parameters, prediction time span, confidence levels, etc., and dynamically view the prediction results under different settings. Through visualization output, complex time-series forecast results are transformed into intuitive and easy-to-understand charts, helping decision-makers quickly understand the dynamic trends and risk characteristics of patent value and supporting forward-looking strategic decisions.

[0054] S7 receives value scores, value attribution results, and value evolution trends. Combined with the application scenarios specified by the user, it adopts a scenario-adaptive report generation framework to obtain a structured patent value assessment report and decision-making suggestions. Specifically, a scenario-adaptive report generation framework is adopted, which obtains a complete evaluation output for a specific application scenario through parameter configuration, result integration, report generation, chart display, and decision reasoning.

[0055] S71, parameter configuration and weight customization for application scenarios; Based on the application scenario type identifier input by the user, a scenario knowledge base matching method is used to determine the parameter configuration of the evaluation framework, obtaining the customized value weights and evaluation focuses for each scenario. Specifically, an application scenario knowledge base is constructed, including various typical scenarios such as pledge financing, technology acquisition, infringement litigation, licensing negotiations, and strategic layout. Each scenario is associated with a set of parameter configuration schemes. The configuration schemes include weight coefficients for three dimensions: legal value, technological value, and economic value, as well as specific evaluation indicators for each scenario. For example, the configuration for the pledge financing scenario is: legal value weight 0.4, economic value weight 0.4, technological value weight 0.2, with indicators including patent stability, expected returns, and monetization difficulty. The configuration for the technology acquisition scenario is: technological value weight 0.5, economic value weight 0.3, legal value weight 0.2, with indicators including technological innovation, technological complementarity, and market expansion potential. After the user selects an application scenario, the system extracts the corresponding weight configuration from the knowledge base. Users can fine-tune the default configuration, adjusting the weight coefficients according to specific needs to achieve personalized customization. Customized weights are applied to the three-dimensional value score output in step S4, and a scenario-customized comprehensive value score is calculated by weighted summation. This score reflects the comprehensive value level of the patent in a specific application scenario.

[0056] S72, Collaborative Value Assessment in Technology Acquisition Scenarios; For technology acquisition scenarios, based on the correlation analysis between the target patent and the acquirer's existing patent portfolio, a synergy effect quantification method is used to calculate the incremental value after the acquisition, resulting in a patent portfolio synergy score. Specifically, data on the acquirer's existing patent portfolio is obtained, extracting features such as technology classification, technology themes, and market coverage. The technological complementarity between the target patent and the acquirer's patent portfolio is calculated by comparing the technology classification of the target patent with the technology classification distribution of the acquirer's patent portfolio to identify whether the target patent fills a technological gap for the acquirer or strengthens the acquirer's technological advantages. High technological complementarity indicates high synergy value. Market coverage enhancement is calculated by analyzing the geographical distribution and application areas of the target patent overlap and expansion with the acquirer's patent portfolio. If the target patent covers important markets not yet covered by the acquirer, the market synergy value is high. Patent barrier strengthening is calculated by assessing whether the introduction of the target patent can strengthen the acquirer's patent fence and enhance its technological barriers against competitors. The barrier strengthening effect is quantified by analyzing the competitive relationship between the target patent and competitor patents, as well as the synergistic defensive capabilities with the acquirer's patents. The scores from three dimensions—technological complementarity, increased market coverage, and strengthened patent barriers—are combined to generate a total patent portfolio synergy score. This synergy score reflects the incremental contribution of the target patents to the acquirer's patent portfolio and is an important component of M&A valuation. In the final comprehensive value score, the independent value and synergistic value of the target patents are added together to obtain the total value assessment in the M&A scenario.

[0057] S73, Structured integration of multidimensional evaluation results; Based on the value scoring in step S4, the attribution analysis in step S5, the dynamic prediction in step S6, the scenario-customized scoring in step S71, and the collaborative value scoring in step S72, a structured data integration method is used to obtain a comprehensive evaluation data structure containing multi-level evaluation information. Specifically, a hierarchical data model of the evaluation results is constructed. The top layer is the comprehensive value score and value level; the second layer is the sub-scores and confidence intervals for legal value, technical value, and economic value; the third layer is the key influencing factors and causal contribution of each value dimension; the fourth layer is the predicted trajectory and confidence interval for the future evolution of value; and the fifth layer is the scenario-specific collaborative value and risk warning. The evaluation results output from each step are organized according to the hierarchical structure, and the relationships between data fields are established to ensure data consistency and integrity. Numerical evaluation results are formatted to unify precision and units, and textual evaluation results are structured and labeled with entities and relationships. Metadata for the evaluation results is constructed, including evaluation time, evaluation model version, data source, confidence level, etc., to ensure the traceability of the evaluation results. A structured evaluation data package is generated to support subsequent report generation, data analysis, and system integration.

[0058] S74, Intelligent generation of natural language assessment reports; Based on the structured evaluation data integrated in step S73, template matching and natural language generation technologies are used to obtain a highly readable text evaluation report. Specifically, a standard template for the evaluation report is designed, including modules such as executive summary, basic patent information, value evaluation conclusion, legal value analysis, technical value analysis, economic value analysis, key influencing factors, value evolution prediction, and application suggestions. For each module, rules and templates for text generation are defined. The template contains a fixed narrative framework and variable placeholders, which will be filled with specific evaluation data. For example, the template for the value evaluation conclusion module is: The comprehensive value evaluation result of this patent is [Comprehensive Score], at the [Value Level] level. Among them, the legal value score is [Legal Value Score], the technical value score is [Technical Value Score], and the economic value score is [Economic Value Score]. The evaluation confidence level is [Confidence Level]. The evaluation data from step S73 is filled into the placeholders in the template to generate specific narrative text. For the key influencing factors module, based on the attribution analysis results of step S5, a list of factors and a contribution description are generated according to their importance. The template for the key factors affecting the patent value includes: [Factor 1] contribution [Contribution 1], [Factor 2] contribution [Contribution 2]. For the value evolution prediction module, based on the prediction results of step S6, a trend description is generated. The template for the future [predicted time span] is that the patent value is expected to show an [rising / falling / stable] trend, with the predicted value range being from the [lower bound] to the [upper bound]. Natural language generation technology is used to smooth the text generated from the template, eliminating mechanical feel and enhancing readability, for example, by using synonym replacement and sentence transformation techniques. Finally, a complete text evaluation report is generated, with professional and standardized language, clear and coherent logic, meeting the quality requirements of a professional evaluation report.

[0059] S75, multi-dimensional visualization of charts; Based on the evaluation data from step S73, data visualization technology is used to generate various charts, providing an intuitive visual representation of the evaluation results. Specifically, a pie chart of value composition is generated, showing the proportion of legal value, technological value, and economic value in the overall value, with the area of ​​each sector visually reflecting its contribution. A bar chart of key influencing factors is generated, with the horizontal axis representing the name of the influencing factor and the vertical axis representing the causal contribution. The importance of different factors is compared by the height of the bars, and color coding is used to distinguish between positive and negative impacts. A line chart of value evolution trend is generated, with the horizontal axis representing time and the vertical axis representing the value score, plotting the historical value trajectory and future multi-scenario prediction trajectory, with shaded areas representing confidence intervals. A patent network location map is generated, using graph visualization technology to show the position of the target patent in the citation network. The size of the node indicates the importance of the patent, and the thickness of the edge indicates the citation strength, highlighting the target patent and its key neighbor nodes. A radar chart is generated to show the performance of the patent on multiple evaluation indicators. Each axis of the radar chart represents a different evaluation dimension, such as innovativeness, stability, influence, and market potential, with the shape of the radar chart visually reflecting the advantages and disadvantages of the patent. For M&A scenarios, generate a synergy value comparison chart to compare the independent value of the target patent with the synergy value after the M&A, showcasing the value increment brought by the acquisition. Integrate all visualizations into the evaluation report, complementing the text descriptions to achieve a visually appealing presentation, enhancing the report's professionalism and persuasiveness.

[0060] S76, Scenario-customized decision suggestion generation; Based on the scenario configuration in step S71, the evaluation data in step S73, and the domain knowledge base, a combination of rule-based reasoning and case-based reasoning is used to generate targeted decision suggestions, resulting in actionable application guidance. Specifically, a decision rule base is constructed, with pre-defined decision suggestion rules for different application scenarios and combinations of evaluation results. For example, in a pledge financing scenario, if the legal value score is higher than the first threshold and the value evolution prediction is stable, it is suggested that the patent is suitable as a pledge asset with low legal risk and stable value, and the credit limit can be referenced to the economic value score. If the legal value score is lower than the second threshold or there is a risk of invalidity, it is suggested that the patent has potential legal instability issues, and further legal due diligence should be conducted to prudently determine the pledge value. In a technology acquisition scenario, if the technology complementarity score is high and the market synergy is high, it is suggested that the patent is highly synergistic with the acquirer's technology layout, and acquisition should be given priority, as it can create significant synergistic value. If the technology overlaps significantly but lacks innovation, it is suggested that the patent overlaps with the acquirer's existing technology with limited marginal value, and alternative solutions should be evaluated. Based on the evaluation results, corresponding decision rules are matched to generate preliminary suggestion text. Furthermore, a case-based reasoning approach is employed to retrieve historical cases with similar features to the current patent from a historical evaluation case database. The decision-making experience and actual results of these historical cases are extracted as a reference for decision-making recommendations. By combining the results of rule-based reasoning and case-based reasoning, a complete decision-making recommendation is generated, including elements such as decision direction, risk warnings, operational suggestions, and reference bases. This decision-making recommendation is integrated into the application recommendation module of the evaluation report, providing users with closed-loop support from evaluation to decision-making.

[0061] In one embodiment of the present invention, an intellectual property valuation system is provided for performing the steps in the above-described intellectual property valuation method, such as... Figure 2 As shown, it includes: The data acquisition module is used to obtain a patent feature dataset by using distributed data acquisition based on the patent number and the technical field identifier of the target patent. The semantic representation module is used to obtain the patent feature dataset. It uses a pre-trained language model combined with a multi-level feature fusion mechanism to obtain the patent representation vector. The graph embedding module receives the patent feature dataset and the patent representation vector, and uses a spatiotemporal heterogeneous graph neural network for dynamic modeling to obtain the patent graph embedding representation. The value assessment module receives patent representation vectors and patent graph embedding representations, and uses a multi-task meta-learning framework to conduct collaborative assessments to obtain scores for legal value, technical value, and economic value. The attribution analysis module is used to receive legal value, technical value, and economic value scores, identify the driving factors of patent value, and obtain value attribution results with causal explanations. The trend prediction module is used to acquire external technology evolution and market trend data, and combine it with patent graph embedding representation and legal value, technical value and economic value scores. It uses a time sequence neural network to make dynamic predictions and obtain the value evolution trend at multiple time nodes. The report generation module receives value scores, value attribution results, and value evolution trends. Combined with the user-specified application scenario, it adopts a scenario-adaptive report generation framework to obtain a structured patent value assessment report and decision-making recommendations.

[0062] This invention focuses on the application of intellectual property due diligence in cross-border technology mergers and acquisitions. The following example, a multinational corporation's acquisition of an artificial intelligence startup, demonstrates the practical application effect of the method of this invention.

[0063] A major European automaker plans to acquire a Chinese startup focused on autonomous driving perception algorithms. This startup holds 12 core patents in China, the US, and Europe. The acquiring company needs to conduct a comprehensive evaluation of these patents, assessing not only the value of individual patents but also the synergies between the patent portfolio and the acquiring company's existing autonomous driving patent portfolio, and predicting the post-merger technology integration value. The acquiring company uses the intellectual property valuation system of this invention to intelligently evaluate the target company's patent portfolio.

[0064] Examples of data collection are shown in Table 1; The system collected complete data on 12 patents belonging to the target company from the China National Intellectual Property Administration, the United States Patent and Trademark Office, and the European Patent Office. Taking one of the core patents, CN202110XXXXXX, as an example, the data collected by the system includes: Table 1, Example of Data Collection Examples of evaluation results are shown in Table 2; After a systematic, multi-dimensional evaluation, the evaluation results for patent CN202110XXXXXX are as follows: Table 2, Examples of Evaluation Results The system predicts the value evolution trend of this patent over the next three years. In a neutral scenario, the patent value score is expected to rise from the current 85 points to 89 points in 2027, peaking at 91 points in 2028, and then gradually declining with technological iterations. In an optimistic scenario, if autonomous driving technology accelerates its commercialization, the patent value could reach a peak of 96 points. In a pessimistic scenario, if there is a significant shift in the technological roadmap, the patent value may drop to 75 points. The prediction confidence interval widens over time, from 8 points in 2026 to 15 points in 2028, reflecting the increased uncertainty in long-term predictions.

[0065] Based on the assessment results, the system generates the following decision recommendations for the acquirer: The overall value of the patent portfolio is high, with strong technological innovation and high synergy with the acquirer's technological layout. It is recommended to prioritize the acquisition. Pay particular attention to the five core patents, including CN202110XXXXXX, which are technologically leading in the field of perception algorithms and can significantly enhance the acquirer's technological competitiveness after the acquisition. It is recommended to emphasize synergistic value during the acquisition negotiations and reasonably determine the valuation range. After the acquisition, it is recommended to strengthen the global patent portfolio, supplementing patent applications in key markets such as Japan and South Korea to further enhance the strategic value of the patent portfolio. At the same time, pay attention to the risks of technological evolution. Autonomous driving perception technology iterates rapidly; it is recommended to continuously upgrade the technology and update patents after the acquisition to maintain technological leadership.

[0066] Through the method of this invention, the acquiring party obtains comprehensive, accurate, and dynamic patent valuation information in the M&A decision-making process, which effectively supports M&A negotiations and strategic decisions, ultimately leading to the successful completion of the acquisition transaction and the realization of the expected technological synergy and market value enhancement after the acquisition.

[0067] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for assessing the value of intellectual property, characterized in that, Includes the following steps: S1. Based on the patent number and technical field identifier of the target patent, a patent feature dataset is obtained by distributed data collection. S2, obtain the patent feature dataset, and use a pre-trained language model combined with a multi-level feature fusion mechanism to obtain the patent representation vector; S3 receives the patent feature dataset and patent representation vector, and uses a spatiotemporal heterogeneous graph neural network for dynamic modeling to obtain the patent graph embedding representation. S4 receives patent representation vectors and patent graph embedding representations, and uses a multi-task meta-learning framework for collaborative evaluation to obtain scores for legal value, technical value, and economic value. S5 receives scores for legal value, technical value, and economic value, identifies the driving factors of patent value, and obtains value attribution results with causal explanation. S6: Obtain external technology evolution and market trend data, combine patent graph embedding representation with legal value, technical value and economic value scores, and use a time-series neural network to make dynamic predictions to obtain the value evolution trend at multiple time nodes; S7 receives value scores, value attribution results, and value evolution trends. Combined with the user-specified application scenarios, it uses a scenario-adaptive report generation framework to obtain a structured patent value assessment report and decision-making recommendations.

2. The intellectual property valuation method according to claim 1, characterized in that, S1 includes: The basic attribute information of patents is extracted in batches from the patent database system using an API interface to obtain structured patent metadata; Web crawling technology and text parsing algorithms are used to extract the text and graphic content of patents from the full-text patent database, resulting in multi-level text data representation and graphic data representation. A citation relationship network for the patent is constructed using graph database technology to obtain the forward citation set and backward citation set of the target patent; A data quality assessment model is used to quantify and score the completeness, timeliness, and reliability of each data source and each data field, resulting in a data quality weight vector.

3. The intellectual property valuation method according to claim 1, characterized in that, S2 includes: Based on the language model parameters pre-trained from a large-scale general patent corpus, model transfer technology is used as the initialization basis for domain coding to obtain a basic encoder with patent language understanding capabilities. Based on the unlabeled patent text set in the target technology field, a self-supervised task of mask language modeling is used to perform domain-adaptive fine-tuning on the basic encoder, resulting in a domain encoder that is sensitive to the technical terms and expression patterns of the target domain. Based on the technical classification labels of patents in the target domain, a contrastive learning strategy is used to discriminately optimize the domain encoder, resulting in a discriminative semantic encoder that can distinguish between different technical branches. Based on the segmented text of the patent specification, a discriminative encoder is used to perform hierarchical semantic encoding to obtain a multi-level semantic representation of the patent text. Based on the semantic vector of the specification, the semantic vector of the claims, and the embedding vector of the technology classification, a multi-head attention mechanism is used to fuse features to obtain the patent representation vector.

4. The intellectual property valuation method according to claim 1, characterized in that, S3 includes: Based on patent citation data and applicant data, a patent association network is constructed using a heterogeneous graph modeling method, resulting in a heterogeneous graph structure containing multiple node types and edge types, including citation edges, technology similarity edges, competing edges, and family edges. Based on the heterogeneous graph and the temporal attributes of patents, an adaptive time window partitioning method is used to generate a dynamic graph snapshot sequence, thereby obtaining graph sequence data that reflects the temporal evolution of the patent network. Based on the graph snapshot of each time window, a heterogeneous graph attention network is used to encode the spatial topology of the graph, resulting in a node spatial embedding vector that integrates information from multiple types of neighbors. Based on the node spatial embedding vector sequence of each time window, a temporal convolutional network is used to model the temporal evolution pattern of patent nodes, resulting in a patent graph embedding representation that integrates temporal dynamics.

5. The intellectual property valuation method according to claim 1, characterized in that, S4 includes: Based on data quality weight vector, semantic representation vector, graph embedding vector, and structured feature vector, a quality-aware feature fusion mechanism is adopted to obtain a comprehensive patent feature representation. Based on the fused feature vectors, a shared representation learning network for three evaluation sub-tasks—legal value, technological value, and economic value—is constructed using a multi-task learning architecture, resulting in deep feature representations shared by the tasks. Based on a multi-task network, an uncertainty weighting strategy is adopted to dynamically balance the weights of each subtask in joint training, resulting in a joint loss function for task balancing. Based on historical patent evaluation data from multiple technology fields, a meta-learning algorithm is used to train a multi-task network to obtain model initialization parameters with rapid domain adaptation capabilities. A Bayesian neural network algorithm is used to quantify the uncertainty of the value assessment results, and the probability distribution and confidence interval of the value score are obtained.

6. The intellectual property valuation method according to claim 1, characterized in that, S5 includes: Based on feature data and value annotation data of large-scale historical patents, a causal structure learning algorithm is used to infer the causal dependencies between feature variables and between features and values, and to obtain a causal directed acyclic graph. Based on the learned preliminary causal graph, domain expert knowledge and theoretical analysis are used to revise and verify the causal graph, resulting in a reliable causal relationship model. Based on the causal graph and the feature data of the target patent, the causal inference method is used to calculate the causal effect of each feature variable on the value of the target patent, and the causal contribution quantification results of the features are obtained. Based on the identified causal effects, a counterfactual inference method is used to conduct a value attribution analysis on the target patent, obtaining the specific contribution and attribution path of each key factor.

7. The intellectual property valuation method according to claim 1, characterized in that, S6 includes: Based on large-scale historical patent data, a time window sliding method is used to construct a time-series evolution training dataset of patent value, and input-output sequence pairs are obtained for time-series model training. Based on technology evolution paths and market development data, time series analysis methods are used to extract the time series characteristics of external driving factors, thereby obtaining time series signals that reflect technology and market dynamics. Based on time-series training data and external time-series features, a sequence-to-series deep learning model is used to train a time-series evolution predictor of patent value, resulting in a time-series model that can predict future value trajectories. Based on the time-series prediction model, the scenario analysis method is used to generate value evolution predictions under different external environment assumptions, and the value prediction trajectories of three scenarios are obtained: optimistic, neutral and pessimistic.

8. The intellectual property valuation method according to claim 1, characterized in that, S7 includes: Based on the application scenario type identifiers input by users, the parameter configuration of the evaluation framework is determined by the scenario knowledge base matching method, so as to obtain the value weight and evaluation focus of scenario customization; Based on value scoring, attribution analysis, dynamic prediction, and scenario-customized scoring, a structured data integration method is used to obtain a comprehensive evaluation data structure containing multi-level evaluation information. Based on integrated structured assessment data, template matching and natural language generation technologies are used to obtain highly readable text assessment reports; based on the assessment data, data visualization technologies are used to generate various charts to provide an intuitive visualization of the assessment results.

9. The intellectual property valuation method according to claim 4, characterized in that, The method of constructing a patent association network using heterogeneous graph modeling specifically includes: Create a set of patent nodes, with each patent corresponding to a node in the graph. Node attributes include patent semantic vector, patent application date, and legal status. Create citation edges based on citation relationships, establishing directed edges from cited patents to citing patents. Based on the technology classification, a technology similarity edge is created. For patent pairs with the same first few digits of the technology classification number, the cosine similarity of their semantic vectors is calculated. If the similarity exceeds the preset similarity threshold, an undirected technology similarity edge is established between the two patent nodes. Competitive edges are created based on applicant information. For patent pairs belonging to different applicants but with high overlap in technical fields, competitive edges are established. Family edges are created based on patent family relationships. For patents from different countries belonging to the same patent family, family-related edges are established.

10. An intellectual property valuation system, used to perform the steps of an intellectual property valuation method as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to obtain a patent feature dataset by using distributed data acquisition based on the patent number and the technical field identifier of the target patent. The semantic representation module is used to obtain the patent feature dataset. It uses a pre-trained language model combined with a multi-level feature fusion mechanism to obtain the patent representation vector. The graph embedding module receives the patent feature dataset and the patent representation vector, and uses a spatiotemporal heterogeneous graph neural network for dynamic modeling to obtain the patent graph embedding representation. The value assessment module receives patent representation vectors and patent graph embedding representations, and uses a multi-task meta-learning framework to conduct collaborative assessments to obtain scores for legal value, technical value, and economic value. The attribution analysis module is used to receive legal value, technical value, and economic value scores, identify the driving factors of patent value, and obtain value attribution results with causal explanations. The trend prediction module is used to acquire external technology evolution and market trend data, and combine it with patent graph embedding representation and legal value, technical value and economic value scores. It uses a time sequence neural network to make dynamic predictions and obtain the value evolution trend at multiple time nodes. The report generation module receives value scores, value attribution results, and value evolution trends. Combined with the user-specified application scenario, it adopts a scenario-adaptive report generation framework to obtain a structured patent value assessment report and decision-making recommendations.

Images