Patent pre-assessment method and apparatus

Abstract

The patent pre-examination evaluation method and device of the present disclosure performs vectorization preprocessing on the to-be-processed file to obtain a structured sequence corresponding to the text content of the to-be-processed file; determines the patent object corresponding to the structured sequence based on a patent law clause rule library and a semantic judgment model; in response to the patent object determination passing, evaluates the implementation details and logical relationship of the technical solution to obtain a technical solution detail evaluation result; analyzes the to-be-processed file to obtain an evaluation result of the novelty and creativity of the to-be-processed file; in response to the passing of the evaluation result of the novelty and creativity, generates a structured patent pre-examination report containing a pre-examination conclusion based on the patent object, the technical solution detail evaluation result, and the evaluation result of the novelty and creativity; realizes full-process automatic processing of the patent pre-examination process, significantly improves the pre-examination efficiency and result consistency, reduces the labor cost, and provides a scientific basis for patent examination and application decision-making.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to an evaluation method and apparatus for patent pre-examination. Background Technology

[0002] Patent pre-examination, as a crucial preliminary step in intellectual property management, is widely used in patent application quality control and examination process optimization. With the development of artificial intelligence technology, patent pre-examination systems have gradually incorporated technologies such as Natural Language Processing (NLP), machine learning, and knowledge graphs, constructing an automated technology system from document parsing to examination decision-making.

[0003] However, common patent pre-examination methods typically employ static rule bases for patent subject matter determination, which suffers from discrepancies in the interpretation of ambiguous clauses. Furthermore, the three-dimensional analysis stage relies on manual comparison of patent databases, leading to lengthy pre-examination times for individual patents and a high rate of omission of innovative points. Therefore, while traditional pre-examination evaluation systems achieve automation in some aspects, they suffer from fragmented processing flows, insufficient multi-technology collaboration, and poor dynamic adaptability of examination standards, thus affecting the overall efficiency and consistency of patent pre-examination evaluation results. Summary of the Invention

[0004] This disclosure provides a method and apparatus for patent pre-examination evaluation. Its main purpose is to address the shortcomings of traditional pre-examination evaluation systems, which, although automated in some aspects, suffer from fragmented processing flows, insufficient coordination among multiple technologies, and poor dynamic adaptability of examination standards, thus affecting the overall efficiency and consistency of patent pre-examination evaluation results.

[0005] According to a first aspect of this disclosure, a method for evaluating patent pre-examination is provided, comprising: The file to be processed is preprocessed into vectorization to obtain a structured sequence corresponding to the text content of the file to be processed; Based on the patent law clause rule base and semantic determination model, the patent subject matter corresponding to the structured sequence is determined; In response to the successful determination of the patent subject matter, the implementation details and logical relationships of the technical solution are evaluated to obtain the results of the technical solution's comprehensiveness assessment. The document to be processed is analyzed to obtain the evaluation results of its novelty and inventiveness. In response to the successful evaluation of novelty and inventiveness, a structured patent pre-examination report containing the pre-examination conclusion is generated based on the patent subject matter, the detailed evaluation results of the technical solution, and the evaluation results of novelty and inventiveness.

[0006] Once the technical solution's level of detail assessment is passed, the technical keywords in the file to be processed are extracted, and the technical keywords are classified into IPCs based on a multi-label classification algorithm that combines patent semantic features and technical field features to obtain the target IPC classification. Based on the patent subject matter, the results of the detailed assessment of the technical solution, and the results of the novelty and inventiveness assessment, a structured patent pre-examination report containing a pre-examination conclusion is generated, including: Based on the patent subject matter, the results of the technical solution detail assessment, the target IPC classification, and the results of the novelty and inventiveness assessment, a structured patent pre-examination report containing the pre-examination conclusion is generated.

[0007] In some embodiments, the vectorization preprocessing of the file to be processed to obtain a structured sequence corresponding to the text content includes: Convert the file to be processed into plain text format data; The plain text data is segmented and standardized to obtain the segmentation results. The word segmentation results are preprocessed into vectors to obtain the structured sequence corresponding to the text content.

[0008] In some embodiments, the step of performing vectorization preprocessing on the word segmentation results to obtain the structured sequence corresponding to the text content includes: The word segmentation results are processed based on the whole word masking model to obtain an attention mask matrix of a preset sequence length. The attention mask matrix contains the positions of valid words and filler words that mark the word segmentation results. The attention mask matrix is ​​vectorized to obtain the structured sequence corresponding to the text content.

[0009] In some embodiments, determining the patent subject matter corresponding to the structured sequence based on the patent law rule base and semantic determination model includes: Based on the rule matching engine and the patent law clause rule library, the text content in the file to be processed is classified and filtered, and the filtering results are marked as to whether in-depth judgment is required. For a text sequence marked as depth-based, the text content is input into the semantic determination model for encoding processing to obtain the semantic determination result; The semantic determination result is then subjected to regulatory compliance verification to obtain the regulatory compliance verification result. Based on the semantic determination result and the regulatory compliance verification result, the determination result of whether the document to be processed belongs to patent subject matter is obtained.

[0010] In some embodiments, for a text sequence marked as depth-determined, inputting the text content into the semantic determination model for encoding processing to obtain a semantic determination result includes: The text content is input into the semantic determination model for encoding processing, and the hidden state vector corresponding to the target word of the text content is extracted; wherein, the hidden state vector contains the global semantic representation of the entire text sequence; The hidden state vector is classified based on a linear classification layer to obtain the classification result; The classification results are subjected to probability normalization to determine the semantic determination result that the document to be processed belongs to the patent subject matter.

[0011] In some embodiments, performing regulatory compliance verification on the semantic determination result to obtain a regulatory compliance verification result includes: The semantic judgment results are similar to those obtained by performing similarity calculations based on a pre-defined knowledge graph database. The knowledge graph database stores entities and their relationships among at least one of legal provisions, violation cases, and keywords in a node-relationship format. The compliance verification result of the regulations is obtained based on the similarity calculation result.

[0012] In some embodiments, the step of evaluating the implementation details and logical relationships of the technical solution in response to the successful determination of the patent subject matter, to obtain a result of the technical solution's comprehensiveness assessment, includes: The technical elements in the file to be processed are identified based on a preset technical element tagging system of the patent text; wherein, the technical elements constitute the technical solution; Based on the technical elements in the file to be processed, a technical element integrity score is determined; By constructing a sentence similarity graph and applying the PageRank algorithm to identify key sentences in the technical elements, and calculating the technical depth score based on the technical content and description depth of the key sentences to determine the sufficiency score of implementation details; The technical solution is converted into a directed graph, wherein the nodes of the directed graph represent technical steps, technical features or implementation elements, and the edges of the graph represent logical relationships between technical elements, including at least one of causal relationships, dependency relationships and temporal relationships. The directed graph is analyzed and processed. The relational representation between nodes is learned through graph convolutional networks to detect logical loops and inconsistent reasoning paths in the directed graph. A semantic conflict detection algorithm is used to identify contradictory sentence pairs in the text. The ratio of the number of contradictory sentence pairs to the total number of sentence pairs in the text is calculated to obtain a logical consistency score. The results of the detailed assessment of the technical solution are determined based on the completeness score of the technical elements, the sufficiency score of the implementation details, and the logical consistency score.

[0013] In some embodiments, the analysis of the document to be processed to obtain the evaluation results of the novelty and inventiveness of the document to be processed includes: The content of the file to be processed and the uploaded comparison file are subjected to paragraph-level semantic vectorization processing; Calculate the semantic similarity between paragraphs of the file to be processed and paragraphs of the comparison file, and determine the novelty evaluation result based on the semantic similarity; The core technical features are automatically identified and extracted from the file to be processed, and a structured list of technical features is generated. The structured list of technical features is matched and analyzed with the technical features in the comparison file to identify the differences between the technical features; An importance assessment analysis is performed on the identified differential features, and the contribution of each differential feature to the overall creativity is quantitatively calculated to obtain the evaluation result of the creativity.

[0014] According to a second aspect of this disclosure, a patent pre-examination evaluation apparatus is provided, comprising: The preprocessing unit is used to perform vectorized preprocessing on the file to be processed to obtain a structured sequence corresponding to the text content of the file to be processed. The determination unit is used to determine the patent subject matter corresponding to the structured sequence based on the patent law clause rule base and semantic determination model; The evaluation unit is used to evaluate the implementation details and logical relationships of the technical solution in response to the successful determination of the patent subject matter, and to obtain the evaluation result of the completeness of the technical solution. The first determining unit is used to analyze the document to be processed and obtain the evaluation results of the novelty and inventiveness of the document to be processed. The generation unit is used to generate a structured patent pre-examination report containing the pre-examination conclusion based on the patent subject matter, the technical solution detail assessment result, and the novelty and inventiveness assessment result, in response to the passing of the novelty and inventiveness assessment result.

[0015] In some embodiments, the apparatus further includes: The second determining unit is used to, after evaluating the implementation details and logical relationships of the technical solution and obtaining the technical solution completeness evaluation result, extract the technical keywords from the file to be processed in response to the passing of the technical solution completeness evaluation result, and perform IPC classification on the technical keywords based on a multi-label classification algorithm to obtain the target IPC classification. The generation unit is also used to generate the structured patent pre-examination report containing the pre-examination conclusion based on the patent subject matter, the technical solution detail assessment results, the target IPC classification, and the novelty and inventiveness assessment results.

[0016] The preprocessing unit includes: A conversion module is used to convert the file to be processed into plain text format data; The word segmentation module is used to perform word segmentation and standardization operations on the plain text format data to obtain the word segmentation results. The processing module is used to perform vectorization preprocessing on the word segmentation results to obtain the structured sequence corresponding to the text content.

[0017] In some embodiments, the processing module is further configured to: The word segmentation results are processed based on the whole word masking model to obtain an attention mask matrix of a preset sequence length. The attention mask matrix contains the positions of valid words and filler words that mark the word segmentation results. The attention mask matrix is ​​vectorized to obtain the structured sequence corresponding to the text content.

[0018] In some embodiments, the determining unit includes: The filtering module is used to classify and filter the text content in the file to be processed based on the rule matching engine and the patent law clause rule library, and to mark the filtering results as needing in-depth judgment. The encoding module is used to input the text content marked as depth determination into the semantic determination model for encoding processing to obtain the semantic determination result; The verification module is used to verify the legal compliance of the semantic determination result and obtain the legal compliance test result. The determination module is used to determine whether the document to be processed belongs to patent subject matter based on the semantic determination result and the regulatory compliance verification result.

[0019] In some embodiments, the encoding module is further configured to: The text content is input into the semantic determination model for encoding processing, and the hidden state vector corresponding to the target word of the text content is extracted; wherein, the hidden state vector contains the global semantic representation of the entire text sequence; The hidden state vector is classified based on a linear classification layer to obtain the classification result; The classification results are subjected to probability normalization to determine the semantic determination result that the document to be processed belongs to the patent subject matter.

[0020] In some embodiments, the determining module is further configured to: The semantic judgment results are similar to those obtained by performing similarity calculations based on a pre-defined knowledge graph database. The knowledge graph database stores entities and their relationships among at least one of legal provisions, violation cases, and keywords in a node-relationship format. The compliance verification result of the regulations is obtained based on the similarity calculation result.

[0021] In some embodiments, the evaluation unit is further configured to: The technical elements in the file to be processed are identified based on a preset technical element tagging system of the patent text; wherein, the technical elements constitute the technical solution; Based on the technical elements in the file to be processed, a technical element integrity score is determined; By constructing a sentence similarity graph and applying the PageRank algorithm to identify key sentences in the technical elements, and calculating the technical depth score based on the technical content and description depth of the key sentences to determine the sufficiency score of implementation details; The technical solution is converted into a directed graph, wherein the nodes of the directed graph represent technical steps, technical features or implementation elements, and the edges of the graph represent logical relationships between technical elements, including at least one of causal relationships, dependency relationships and temporal relationships. The directed graph is analyzed and processed. The relational representation between nodes is learned through graph convolutional networks to detect logical loops and inconsistent reasoning paths in the directed graph. A semantic conflict detection algorithm is used to identify contradictory sentence pairs in the text. The ratio of the number of contradictory sentence pairs to the total number of sentence pairs in the text is calculated to obtain a logical consistency score. The results of the detailed assessment of the technical solution are determined based on the completeness score of the technical elements, the sufficiency score of the implementation details, and the logical consistency score.

[0022] In some embodiments, the first determining unit is further configured to: The content of the file to be processed and the uploaded comparison file are subjected to paragraph-level semantic vectorization processing; Calculate the semantic similarity between paragraphs of the file to be processed and paragraphs of the comparison file, and determine the novelty evaluation result based on the semantic similarity; The core technical features are automatically identified and extracted from the file to be processed, and a structured list of technical features is generated. The structured list of technical features is matched and analyzed with the technical features in the comparison file to identify the differences between the technical features; An importance assessment analysis is performed on the identified differential features, and the contribution of each differential feature to the overall creativity is quantitatively calculated to obtain the evaluation result of the creativity.

[0023] According to a third aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in the first aspect above.

[0024] According to a fourth aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are configured to cause the computer to perform the method described in the first aspect above.

[0025] According to a fifth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method described in the first aspect above.

[0026] The patent pre-examination evaluation method and apparatus disclosed herein perform vectorized preprocessing on the document to be processed to obtain a structured sequence corresponding to the text content of the document to be processed; based on the patent law clause rule base and semantic judgment model, determine the patent subject matter corresponding to the structured sequence; in response to the patent subject matter being determined to be acceptable, evaluate the implementation details and logical relationships of the technical solution to obtain a detailed evaluation result of the technical solution; analyze the document to be processed to obtain an evaluation result of the novelty and inventiveness of the document to be processed; in response to the novelty and inventiveness evaluation result being acceptable, generate a structured patent pre-examination report containing a pre-examination conclusion based on the patent subject matter, the detailed evaluation result of the technical solution, and the evaluation result of the novelty and inventiveness. The method provided by the embodiments of this disclosure can realize the fully automated processing of the patent pre-examination process, significantly improve the pre-examination efficiency and result consistency, while reducing labor costs, and providing a scientific basis for patent examination and application decision-making.

[0027] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0028] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1A schematic flowchart illustrating a patent pre-examination evaluation method provided in some embodiments of this disclosure; Figure 2 A schematic flowchart illustrating a patent pre-examination evaluation method provided in some embodiments of this disclosure; Figure 3 A schematic flowchart illustrating a patent pre-examination evaluation method provided in some embodiments of this disclosure; Figure 4 A schematic flowchart illustrating a patent pre-examination evaluation method provided in some embodiments of this disclosure; Figure 5 A schematic flowchart illustrating a patent pre-examination evaluation method provided in some embodiments of this disclosure; Figure 6 A schematic flowchart illustrating a patent pre-examination evaluation method provided in some embodiments of this disclosure; Figure 7 A schematic flowchart illustrating a patent pre-examination evaluation method provided in some embodiments of this disclosure; Figure 8 A schematic diagram of the structure of a patent pre-examination evaluation device provided for some embodiments of this disclosure; Figure 9 A schematic diagram of the structure of a patent pre-examination evaluation device provided for some embodiments of this disclosure; Figure 10 A schematic block diagram of an example electronic device provided for some embodiments of this disclosure. Detailed Implementation

[0029] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0030] The following description, with reference to the accompanying drawings, describes a patent pre-examination evaluation method, apparatus, electronic device, and storage medium according to some embodiments of the present disclosure.

[0031] Figure 1 This is a flowchart illustrating a patent pre-examination evaluation method provided in some embodiments of this disclosure. Figure 1 As shown, the method includes the following steps: Step 101: Perform vectorization preprocessing on the file to be processed to obtain a structured sequence corresponding to the text content of the file to be processed.

[0032] Specifically, in some implementations, this step involves structured parsing of user-uploaded files (such as patent disclosure documents). Its core lies in using deep learning models to perform semantic recognition and chapter segmentation of the document content, thereby extracting key information and providing a data foundation for the subsequent intelligent pre-examination process. Technically, this step typically employs a Transformer-based natural language processing model (such as BERT, RoBERTa, or a domain-customized model). Through a combination of pre-training and fine-tuning, the model acquires the ability to recognize technical terms, chapter titles, and content structure in the patent text. In practice, the uploaded disclosure document is first parsed using OCR or PDF, converting it into a processable text format. Subsequently, a deep learning model is used to analyze the text segment by segment, identifying standard chapters such as "Technical Background," "Invention Content," "Technical Effects," and "Claims," ​​and performing entity extraction and semantic annotation of the technical elements within each chapter.

[0033] In terms of parameters, the model input is typically a standardized, preprocessed structured sequence, limited to a length of 512 tokens or less, to ensure computational efficiency and semantic integrity. The model output consists of chapter tags and their corresponding text content. The tagging system can be defined based on ISO / IEC 11179 or CNIPA patent document structure standards. During model training, a cross-entropy loss function can be used, the expression of which is... Where x represents the text feature vector input to the model, This serves as a measure of the difference between the predicted output and the true label. The training dataset can contain a large number of labeled patent disclosure samples covering different technical fields to enhance the model's generalization ability.

[0034] In terms of application scenarios, this step is widely used in patent intelligent pre-examination systems, especially when processing a large number of user-submitted disclosure documents, which can significantly improve information extraction efficiency.

[0035] From a technical perspective, this step achieves efficient structuring of unstructured text, reducing the workload of manual reading and annotation, and improving the automation level of the pre-examination process. Simultaneously, through the semantic understanding capabilities of deep learning models, the logical relationships within the technical content can be effectively identified, providing reliable data support for subsequent patent quality assessment and examination decisions.

[0036] Step 102: Based on the patent law clause rule base and semantic determination model, determine the patent subject matter corresponding to the structured sequence.

[0037] Step 103: In response to the successful determination of the patent subject matter, the implementation details and logical relationships of the technical solution are evaluated to obtain the technical solution detail evaluation result.

[0038] Specifically, this step involves a composite decision-making mechanism based on a patent law rule base and a semantic judgment model (such as a machine learning model) to perform patent subject matter eligibility assessment and detail verification. In some implementations, this mechanism integrates structured legal knowledge with unstructured text analysis capabilities to automatically identify whether the technical content in the patent disclosure meets the requirements for patent subject matter protection, and further verifies whether the technical disclosure is detailed and complete, thereby providing a preliminary assessment of the feasibility of the patent application.

[0039] At the technical implementation level, this step first constructs a rule base based on patent law provisions. Each rule in the rule base is stored in a structured format, including fields such as applicable conditions, judgment logic, and exceptions. Secondly, a pre-trained and fine-tuned Natural Language Processing (NLP) model, such as a classification model based on BERT or RoBERTa architecture, is introduced to perform semantic parsing and feature extraction on the technical content of the disclosure document. The model input is text paragraphs, and the output is tags related to patent law provisions, such as "whether it constitutes a technical solution" and "whether it possesses practical applicability."

[0040] In terms of parameters, the input text to the model needs to undergo standardized preprocessing, including word segmentation, stop word removal, entity recognition, and normalization. The confidence threshold of the model output can be set as follows: This is to ensure the reliability of the judgment results. The detail verification module calculates the disclosure density of technical features. , where x represents the frequency and distribution of key parameters in the text. The smaller the value, the more detailed the disclosure. The system can further combine disclosure standards in the rule base, such as the completeness requirements of technical principles, implementation methods, and application scenarios, to conduct a multi-dimensional evaluation.

[0041] In application scenarios, this step is suitable for intelligent pre-examination systems before patent applications. It can analyze user-uploaded disclosure documents in real time and output assessments of patent eligibility and completeness. Its technical advantages include significantly improving pre-examination efficiency, reducing the burden of manual examination, and assisting applicants in optimizing technical disclosure content through quantitative indicators, thereby increasing the patent application success rate.

[0042] Step 104: Analyze the document to be processed to obtain the evaluation results of the novelty and inventiveness of the document to be processed.

[0043] Specifically, in some implementations, analyzing the novelty and inventiveness of patent applications through a semantic retrieval system and an innovation point evaluation module is one of the core technical steps of this intelligent patent pre-examination system. This step, based on natural language processing (NLP) and knowledge graph technology, combined with semantic information from the patent database, enables automatic novelty judgment and inventiveness assessment of patent technical solutions.

[0044] At the technical implementation level, the semantic retrieval system first performs text preprocessing on the patent disclosure documents uploaded by users, including word segmentation, part-of-speech tagging, entity recognition, and semantic role annotation to extract technical features, invention themes, and key terms. Subsequently, the system vectorizes these features and uses a Transformer-based semantic encoder (such as BERT or RoBERTa) to map the text into embedded vectors in a high-dimensional semantic space. The innovation evaluation module constructs a patent knowledge graph, structurally representing technical solutions in existing patent documents, and performs semantic similarity matching with the technical features in the current disclosure document. The matching process can employ cosine similarity or semantic-based retrieval models (such as DPR, Dense Passage Retrieval) to identify any technological overlap or existing technical solutions.

[0045] At the parameter level, semantic retrieval systems typically set similarity thresholds. This is used to determine the degree of similarity between the current technical solution and existing patents. If the similarity is... If the novelty conflict is not met, then a conflict of interest is considered to exist. The creativeness assessment then introduces a technical difference index. This indicator is quantitatively evaluated based on dimensions such as the innovativeness of the combination of technical features, the significance of the technical effects, and the way technical problems are solved. In some implementations, the system can also combine IPC classification results to limit the search scope and improve the accuracy of the evaluation.

[0046] At the application level, this step is widely used in intelligent pre-examination systems before patent applications, especially in technology-intensive fields (such as artificial intelligence, communications, and biomedicine). It can effectively assist examiners in quickly identifying potential duplicate applications or solutions lacking inventiveness. The system can be deployed on the cloud or local servers, supports batch processing and real-time feedback, and is suitable for scenarios such as the State Intellectual Property Office, enterprise R&D departments, and patent agencies.

[0047] The technical effect of this step is that, through in-depth semantic analysis, it significantly improves the automation level and accuracy of judging patent novelty and inventiveness, reduces the workload of manual searches, and improves the efficiency and scientific nature of pre-examination, providing reliable technical support for the subsequent patent examination process.

[0048] Step 105: In response to the successful evaluation of novelty and inventiveness, a structured patent pre-examination report containing the pre-examination conclusion is generated based on the patent subject matter, the detailed evaluation result of the technical solution, and the evaluation result of novelty and inventiveness.

[0049] Specifically, in some implementations, this is a key output link in the intelligent patent pre-examination system. Its technical implementation is based on a data fusion and logical reasoning mechanism after multi-module collaborative processing. The core of this step lies in the unified integration of the processing results from sub-modules such as chapter content analysis, patent subject matter judgment, detail verification, and novelty and inventiveness analysis, and generating a structured pre-examination report that conforms to the standards of the State Intellectual Property Office based on preset report templates and logical rules.

[0050] At the technical implementation level, the system first extracts processing result data from each submodule through message queues or database transaction mechanisms. The output of each module is typically stored in JSON or XML format, containing key fields such as patent type, claim completeness, technical field tags, and novelty score. The system employs data validation and normalization techniques to ensure consistent data format and clear semantics across modules. Subsequently, the system calls the pre-examination report generation engine, which uses a template engine (such as Apache FreeMarker or Thymeleaf) to dynamically populate content, organizing and presenting the output results from each module according to the structural requirements of the pre-examination report.

[0051] At the parameter level, the system supports custom report generation strategies, such as setting confidence thresholds for preliminary review conclusions and allocating module weights. For example, the output of the novelty analysis module can be set as a comprehensive score. When the score is below a preset threshold In such cases, the system will automatically mark the patent application as "insufficient in novelty" and record the relevant supporting evidence. Furthermore, the system supports multilingual report output to meet the multilingual needs of international patent applications.

[0052] At the application level, this step is widely used in the patent pre-examination process of the State Intellectual Property Office, especially in high-throughput patent application scenarios, such as when companies submit patent disclosures in batches. The system can automatically generate a pre-examination report, significantly improving pre-examination efficiency and reducing the workload of manual review. At the same time, this structured report can serve as a reference for subsequent examination processes, supporting examiners to quickly identify key issues.

[0053] The technical effect of this step is that, through modular data integration and structured output, the patent pre-examination process is standardized and automated, improving the overall intelligence level and decision support capability of the system, and has significant practical value and prospects for promotion.

[0054] The patent pre-examination evaluation method disclosed herein includes: performing vectorized preprocessing on the document to be processed to obtain a structured sequence corresponding to the text content of the document; determining the patent subject matter corresponding to the structured sequence based on a patent law rule base and a semantic judgment model; in response to the successful determination of the patent subject matter, evaluating the implementation details and logical relationships of the technical solution to obtain a detailed evaluation result of the technical solution; analyzing the document to be processed to obtain an evaluation result of the novelty and inventiveness of the document; and in response to the successful evaluation result of the novelty and inventiveness, generating a structured patent pre-examination report containing a pre-examination conclusion based on the patent subject matter, the detailed evaluation result of the technical solution, and the evaluation result of the novelty and inventiveness. The methods provided in some embodiments of this disclosure can achieve fully automated processing of the patent pre-examination process, significantly improving pre-examination efficiency and result consistency, while reducing labor costs and providing a scientific basis for patent examination and application decisions.

[0055] Figure 2 This is a flowchart illustrating a patent pre-examination evaluation method provided in some embodiments of this disclosure. Figure 2 As shown, the method includes the following steps: Step 201: Perform vectorization preprocessing on the file to be processed to obtain a structured sequence corresponding to the text content of the file to be processed; Step 202: Based on the patent law clause rule base and semantic determination model, determine the patent subject matter corresponding to the structured sequence; Step 203: In response to the successful determination of the patent subject matter, the implementation details and logical relationships of the technical solution are evaluated to obtain the technical solution detail evaluation result; Step 204: Once the technical solution's level of detail assessment is passed, extract the technical keywords from the file to be processed, and perform IPC classification on the technical keywords based on a multi-label classification algorithm that combines patent semantic features and technical field features to obtain the target IPC classification.

[0056] Specifically, in some implementations, this step employs a multi-label classification algorithm, combining patent semantic features with technical field features to perform IPC classification on the patent text. Its technical implementation principle is based on the fusion of Natural Language Processing (NLP) and machine learning models, aiming to achieve multi-level, multi-category classification of patent technical content through multi-dimensional feature extraction and label prediction mechanisms. The specific operation includes: First, preprocessing the patent disclosure text, including word segmentation, part-of-speech tagging, entity recognition, and stop word filtering, to extract key semantic information; second, constructing feature vectors, where semantic features can be represented using word embeddings (such as Word2Vec, BERT, etc.), and technical field features are encoded based on the hierarchical structure in the IPC classification system (such as departments, major categories, minor categories, groups, subgroups); further, inputting the above features into a multi-label classification model, such as MLP, LSTM, Transformer structures based on neural networks, or ensemble learning methods (such as Binary Relevance, ClassifierChains, Label Powerset, etc.), to achieve joint prediction of multiple IPC classification labels.

[0057] At the parameter level, a classification threshold θ can be set during model training to control the confidence of label prediction. Typically, θ ∈ [0.5, 0.9], balancing recall and precision. Simultaneously, a loss function loss = f(x) can be introduced to measure the difference between the predicted and true labels, where x represents the input feature vector, and f(x) can be a binary cross-entropy or multi-label cross-entropy loss function. Model evaluation metrics can include macro-F1 score, micro-F1 score, and Hamming Loss to comprehensively measure classification performance.

[0058] At the application level, this step can be embedded in a patent intelligent pre-examination system as an automated classification module. It is used to perform preliminary technical field classification on user-uploaded patent disclosures, assisting examiners in quickly locating relevant technology categories and improving pre-examination efficiency. This method is suitable for patent texts involving multiple technical fields, and demonstrates good adaptability, especially in scenarios involving multiple IPC classification numbers.

[0059] The technical effect of this step is that by integrating semantic and technical features, it significantly improves the accuracy and coverage of IPC classification, reduces the workload of manual classification, provides structured input for subsequent novelty and inventiveness analysis, and enhances the intelligence level and practicality of the patent pre-examination system.

[0060] Step 205: Analyze the document to be processed to obtain the evaluation results of the novelty and inventiveness of the document to be processed; Step 206: In response to the successful evaluation of novelty and inventiveness, a structured patent pre-examination report containing the pre-examination conclusion is generated based on the patent subject matter, the technical solution detail evaluation result, the target IPC classification, and the novelty and inventiveness evaluation result.

[0061] Specifically, in some implementations, this is a key output link in the intelligent patent pre-examination system. Its technical implementation is based on a data fusion and logical reasoning mechanism after multi-module collaborative processing. The core of this step is to unify and integrate the processing results from sub-modules such as chapter content analysis, patent subject matter judgment, detail verification, IPC classification, and novelty and inventiveness analysis, and generate a structured pre-examination report that conforms to the standards of the State Intellectual Property Office based on preset report templates and logical rules.

[0062] For details on steps 201-203 and 205-206, please refer to [the relevant documentation / reference]. Figure 1 The detailed explanation is omitted here.

[0063] In some embodiments, when performing vectorization preprocessing on the file to be processed in step 101 to obtain a structured sequence corresponding to the text content, the following methods may be used, but are not limited to: Figure 3 As shown, it includes: Step 301: Convert the file to be processed into plain text format data.

[0064] Users upload PDF and Word format files (patent disclosure documents) to be processed via the system's web interface or API. Upon receiving the upload request in real time, the system automatically identifies the file type and initiates the data processing flow. The system's file receiving module, a file upload controller based on the Spring Boot framework, can automatically identify the file type, verify file integrity, and store the file in a temporary cache.

[0065] Once the file to be processed has been uploaded, the data preprocessing module performs format recognition and content extraction operations on the uploaded document. In some instances, this data preprocessing module first calls the Apache PDFBox framework to perform structured parsing of the PDF document, automatically recognizing and extracting the text content and image information in the document; for non-text content contained in the document, including but not limited to images, tables, charts, etc., the system calls the Tesseract OCR optical character recognition engine to perform text recognition and conversion processing, converting the text information in the image into a machine-readable text format; at the same time, the system automatically recognizes and filters auxiliary information unrelated to the technical content in the document, such as headers, footers, tables of contents, page numbers, etc.; for Word format documents, the system uses the Apache POI document processing framework to parse the document, extracting the main text content, table data, and formatting information to ensure the integrity and accuracy of the document content.

[0066] Through the above processing flow, the system can uniformly convert uploaded documents of different formats into a standardized plain text format, providing clean and structured text data input for subsequent natural language processing and semantic analysis.

[0067] After the document format conversion is complete, the text cleaning submodule performs cleaning and standardization operations on the extracted raw text data. This text cleaning submodule runs in the memory cache environment of the data preprocessing pipeline and uses a regular expression-based text processing algorithm to systematically clean the raw text.

[0068] In some embodiments, the text cleaning process includes, but is not limited to, the following processing steps: First, a regular expression algorithm is used to automatically identify and remove residual HTML tags and symbols from the text. The regular expression pattern is: The replacement function replaces the matched tag content with an empty string to ensure the purity of the output text; Secondly, redundant whitespace characters in the text are normalized, including multiple consecutive spaces, tabs, newlines, etc., using regular expression patterns. Replace multiple consecutive whitespace characters with a single space, and remove whitespace characters at the beginning and end of the text; Secondly, the punctuation marks in the text are standardized by converting full-width punctuation marks into their corresponding half-width punctuation marks to ensure consistency in punctuation mark format during subsequent processing. Then, based on text location information and a repetition pattern recognition algorithm, the algorithm automatically detects and deletes recurring header and footer content in the document. The recognition algorithm accurately locates and removes formatting information that is irrelevant to the technical content by analyzing the frequency and location distribution characteristics of text blocks in the document.

[0069] Finally, after the above cleaning process, the system outputs standardized plain text content without redundant information, providing high-quality data input for subsequent word segmentation processing and semantic analysis.

[0070] Step 302: Perform word segmentation and standardization operations on the plain text format data to obtain a word segmentation result.

[0071] After text cleaning is completed, the word segmentation processing module performs word segmentation and vocabulary standardization operations on the cleaned text data. This word segmentation processing module runs on a dedicated word segmentation service node and uses a word segmentation algorithm based on the jieba Chinese word segmentation engine to perform lexical segmentation on continuous text.

[0072] In some embodiments, the word segmentation processing flow adopts but is not limited to the following technical steps: First, the system loads the standard dictionary library of the jieba word segmentation engine and a specially constructed professional term dictionary for the patent field. The patent field dictionary covers professional vocabulary such as technical terms, legal terms, and patent-specific expressions to improve the accuracy of word segmentation and domain adaptability; Second, the system calls the cut() method of the jieba word segmentation engine to perform word segmentation on the text. The word segmentation mode is set to the exact mode to ensure accurate lexical boundary recognition and segmentation of the text, generating an ordered sequence of words; Third, the system performs词性标注processing on the word segmentation result. By calling the jieba.posseg module, corresponding词性 information is marked for each word segmentation result, including nouns, verbs, adjectives, prepositions, etc., providing词性 feature support for subsequent semantic analysis; Finally, the system filters the word segmentation result based on a preset stop word list. The stop word list contains 1500 common stop words, covering function words with low contribution to semantic analysis such as "的", "是", "在", "了", "和", etc. Filtering stop words improves the efficiency and accuracy of subsequent processing. The above example content in some embodiments of the present disclosure is not specifically limited.

[0073] Step 303: Perform vectorization preprocessing on the word segmentation result to obtain a structured sequence corresponding to the text content.

[0074] In some embodiments, when performing vectorization preprocessing on the word segmentation result to obtain a structured sequence corresponding to the text content, the following methods can be adopted but are not limited to: processing the word segmentation result based on a full-word masking model to obtain an attention masking matrix with a preset sequence length, where the attention masking matrix contains positions of valid word tokens and padding word tokens marking the word segmentation result; performing vectorization processing on the attention masking matrix to obtain a structured sequence corresponding to the text content. It should be noted that the term "词性标注" in the original text seems to be a Chinese term that might be a misspelling or an incomplete expression. I've left it as it is in the translation for the sake of maintaining the integrity of the original text. If it's supposed to be something else, please correct it in the original text for a more accurate translation.

[0075] After word segmentation, the word vector generation module performs vectorization encoding on the segmentation results. This module runs in a GPU-accelerated computing environment equipped with an NVIDIA Tesla V100 computing card, converting discrete word sequences into continuous numerical vector representations.

[0076] In some embodiments, the vectorization process may employ, but is not limited to, the following technical steps: First, the system loads the Chinese-bert-wwm (Chinese BERT with Whole Word Masking) language model, which is pre-trained on a large-scale Chinese corpus and has a powerful Chinese semantic understanding capability, capable of generating 768-dimensional context-related word vectors for each word; Secondly, the system sets the maximum processing length of the text sequence to 512 tokens. For text sequences that exceed this length limit, a truncation operation is performed, retaining the first 512 tokens for subsequent processing to ensure the consistency of the model input. Secondly, for text sequences with a length of less than 512 tokens, the system performs a padding operation, adding a special padding token to the end of the sequence to make the length of all input sequences uniformly 512. Finally, the system generates an attention_mask matrix, which is a binary matrix of the same length as the input sequence. This matrix is ​​used to identify the positions of valid tokens and padding tokens in the sequence, ensuring that the model ignores the padding part and only focuses on the valid text content when calculating attention weights.

[0077] After the above vectorization process, the system converts the original text into a standardized numerical representation suitable for deep learning models, providing high-quality input features for subsequent tasks such as semantic analysis and similarity calculation.

[0078] In some embodiments, when determining the patent subject matter corresponding to the structured sequence in step 102 based on the patent law clause rule base and semantic determination model, the following steps may be used, but are not limited to: Figure 4 As shown, it includes: Step 401: Based on the rule matching engine and the patent law clause rule library, classify and filter the text content in the file to be processed, and mark the filtering results to determine whether in-depth judgment is required.

[0079] After vectorization preprocessing, the rule matching engine immediately performs preliminary filtering on the text content. Built on the Drools rule engine framework, this engine runs in an in-memory rule base environment and is used to identify obvious non-patented object content in the text.

[0080] Specifically, the rule matching engine adopts a multi-level judgment rule system based on Article 25 of the Patent Law. This rule system includes 15 exclusion rules and 8 inclusion rules, and quickly classifies and filters text content through preset logical judgment rules. 1. First exclusion rule: When the text content contains keywords such as "game rules", "intellectual activity", or "psychological method", the system will automatically mark the text as non-patentable subject matter; 2. Second type of suspected exclusion rule: When the text content contains keywords such as "mathematical formula", "scientific discovery", or "abstract concept", the system marks the text as suspected non-patentable subject matter and requires further verification; 3. The third type of inclusion rule: When the text content contains technical keywords such as "technical solution", "device", "method", "system", etc., the system marks the text as needing to enter the deep semantic judgment process.

[0081] Through the initial screening using the aforementioned rule engine, the system can quickly eliminate content that clearly does not meet the requirements for patent protection, thereby improving the efficiency and accuracy of subsequent in-depth analysis.

[0082] Step 402: For the text sequence marked as depth determination, input the text content into the semantic determination model for encoding processing to obtain the semantic determination result.

[0083] In some embodiments, when executing a text sequence labeled as depth-based, the text content is input into the semantic determination model for encoding processing to obtain a semantic determination result, such as Figure 5 As shown, it includes: Step 4021: Input the text content into the semantic determination model for encoding processing, and extract the hidden state vector corresponding to the target word of the text content; wherein, the hidden state vector contains the global semantic representation of the entire text sequence.

[0084] When the rule engine marks text as "requiring deep semantic analysis," the deep semantic analysis module is triggered. This module is built on the BERT language model and runs in the model inference environment of the GPU inference server, performing deep semantic analysis on text content that the rule engine cannot explicitly determine.

[0085] The system inputs the preprocessed text sequence into the BERT model, which performs semantic encoding through the model's multi-layer Transformer encoder to extract the start position of the text sequence. Hidden state vector corresponding to special token This vector contains a global semantic representation of the entire text sequence; Step 4022: Classify the hidden state vector based on the linear classification layer to obtain the classification result.

[0086] The system will obtain The hidden state vector is input into a linear classification layer, which first processes the input vector... Regularization is applied to prevent overfitting, followed by a linear transformation to generate binary classification prediction results:

[0087] Step 4023: Perform probability normalization processing on the classification result to determine the semantic judgment result that the document to be processed belongs to the patent subject matter.

[0088] The system adopts The activation function normalizes the probability of the classification prediction results and calculates the probability that the text belongs to the patent subject matter:

[0089] Finally, the system sets the judgment threshold to 0.75 (this judgment threshold is an empirical value, and some embodiments of this disclosure are not limited thereto). When the calculated probability value is greater than the threshold, the text is determined to be within the scope of patent protection; otherwise, when the probability value is less than or equal to the threshold, the text is determined to be non-patentable.

[0090] Step 403: Perform regulatory compliance verification on the semantic determination result to obtain the regulatory compliance verification result.

[0091] In some embodiments, the step of verifying the legal compliance of the semantic determination result to obtain the legal compliance inspection result includes: performing similarity calculation on the semantic determination result based on a preset knowledge graph database to obtain a similarity calculation result; wherein, the knowledge graph database stores entities and their associations of at least one of legal provisions, violation cases, and keywords in the form of nodes-relationships; and obtaining the legal compliance inspection result based on the similarity calculation result.

[0092] After an object is determined to be "patentable," the regulatory compliance verification module immediately performs a compliance check on the determination result. This module is built on a compliance analysis engine. It runs in a graph database environment and is used to check whether the text content violates relevant laws, regulations, or social ethics.

[0093] Specifically, the compliance verification process includes the following technical steps: First, the system constructs a knowledge graph database covering 1,200 legal provisions, including the Patent Law, the Implementing Regulations of the Patent Law, and relevant judicial interpretations. This knowledge graph stores legal provisions, violation cases, keywords, and other entities and their relationships in the form of nodes and relationships. Secondly, the system uses a keyword matching algorithm combined with semantic similarity calculation to perform compliance retrieval of text content. By searching for relevant legal provisions and violation cases in the knowledge graph, potential compliance risks can be identified. Next, the system calculates the cosine similarity between the text content and known violation cases. When the similarity value exceeds the preset threshold of 0.75, the text is marked as having potential compliance risks. Finally, the system generates a compliance score of 0-100 based on the search results and similarity calculation results, and outputs a detailed list of potential risk points to provide a compliance reference for subsequent patent application decisions.

[0094] Step 404: Based on the semantic determination result and the regulatory compliance verification result, the determination result that the document to be processed belongs to the patent subject matter is obtained.

[0095] After the three sub-modules of initial screening by the rule engine, deep semantic judgment, and regulatory compliance verification have all completed their judgments, the judgment result integration module performs comprehensive analysis and fusion processing on the judgment results of each module. This module, as a decision fusion unit, runs in memory computing space and outputs the final patent subject matter determination conclusion.

[0096] Specifically, the result integration process employs a weighted voting mechanism to fuse results from multiple modules: The system assigns different weight coefficients to each judgment module: the judgment result of the rule engine ( The weight is 0.3, and the deep semantic model judgment result is ( The weight is 0.5, and the result of the regulatory compliance test ( The weight is 0.2; the system calculates the final judgment score according to the weighted fusion formula:

[0097] In some embodiments, when step 103 is performed in response to the successful determination of the patent subject matter, the implementation details and logical relationships of the technical solution are evaluated to obtain a result of the technical solution's comprehensiveness assessment, such as Figure 6 As shown, it includes: Step 501: Identify the technical elements in the file to be processed based on the preset technical element tagging system of the patent text.

[0098] After the patent subject matter determination is passed, the technical element identification module immediately performs technical element extraction and integrity analysis on the disclosure document. This module is built on named entity recognition technology and runs in a dedicated entity recognition service environment to automatically identify and extract the core technical elements in the disclosure document.

[0099] Specifically, the technical element identification process uses a deep learning model based on BiLSTM-CRF for named entity recognition processing: The system constructs a technical element tagging system specifically for patent texts. This tagging system adopts the BIO (Begin-Inside-Outside) annotation format and includes the following six types of tags: B-TECH tags are used to identify the starting words of technical features, and I-TECH tags are used to identify the internal words of technical features; B-IMPL tags are used to identify the starting words of implementation methods, and I-IMPL tags are used to identify the internal words of implementation methods; B-EFFECT tags are used to identify the starting words of technical effects, and I-EFFECT tags are used to identify the internal words of technical effects. This disclosure does not limit the specific format of the label in some embodiments.

[0100] Step 502: Determine the technical element integrity score based on the technical elements in the file to be processed.

[0101] Secondly, the system uses the BERT word vectors generated in the preprocessing stage as the input features of the BiLSTM-CRF model, captures contextual semantic information through a bidirectional long short-term memory network, learns the transition probability between labels through a conditional random field layer, and outputs the optimal BIO label sequence corresponding to each word. Furthermore, the system extracts complete technical element entities based on the annotation results, including key elements such as technical features, implementation methods, and technical effects, and counts the number and coverage of various technical elements. Finally, the system calculates the completeness score of technical elements using the following formula, where the number of standard technical elements is determined based on patent drafting guidelines and statistical standards for similar technical fields:

[0102] Step 503: By constructing a sentence similarity graph and applying the PageRank algorithm to identify key sentences in the technical elements, and calculating the technical depth score based on the technical content and description depth of the key sentences to determine the sufficiency score of implementation details.

[0103] After the technical elements are identified, the detailed analysis module immediately evaluates the sufficiency of the implementation details of the technical solution. This module is built on a semantic sufficiency analyzer and runs in a deep learning inference cluster environment to quantitatively evaluate the level of detail in the description of the technical solution.

[0104] The system uses the TextRank key sentence extraction algorithm to evaluate the depth level of technical descriptions. It constructs a sentence similarity map and applies the PageRank algorithm to identify key sentences in the technical descriptions. Based on the technical content and description depth of the key sentences, the system calculates the technical depth score.

[0105] Step 504: Convert the technical solution into a directed graph, wherein the nodes of the directed graph represent technical steps, technical features or implementation elements, and the edges of the graph represent logical relationships between technical elements, including at least one of causal relationships, dependency relationships and temporal relationships.

[0106] Step 505: Analyze and process the directed graph by learning the relationship representation between nodes through a graph convolutional network, detecting logical loops and inconsistent reasoning paths in the directed graph; identify contradictory sentence pairs in the text through a semantic conflict detection algorithm, calculate the ratio of the number of contradictory sentence pairs to the total number of sentence pairs in the text, and obtain a logical consistency score.

[0107] After the implementation details adequacy assessment is completed, the logic analysis engine immediately initiates a check and analysis of the internal logical consistency of the technical solution. This engine is built on a context consistency checker and runs in the knowledge graph reasoning engine environment to detect logical contradictions and inconsistencies in the description of the technical solution.

[0108] Specifically, the logical consistency verification process includes the following technical steps: First, the system converts the technical solution content into a directed graph data structure, where the nodes of the graph represent technical steps, technical features or implementation elements, and the edges of the graph represent the logical relationships between technical elements, including causal relationships, dependency relationships, temporal relationships, etc. Secondly, the system uses graph neural network algorithms to analyze and process the directed graph of the constructed technical solution. By learning the relationship representation between nodes through graph convolutional networks, it automatically detects logical loops and inconsistent reasoning paths in the graph structure. Finally, the system uses a semantic conflict detection algorithm to identify contradictory statement pairs in the text and calculates the ratio of the number of contradictory statement pairs to the total number of statement pairs in the text. Step 506: Determine the detailedness assessment result of the technical solution based on the completeness score of the technical elements, the sufficiency score of the implementation details, and the logical consistency score.

[0109] After completing the evaluations of the three sub-modules—technical element integrity analysis, implementation detail sufficiency assessment, and logical consistency verification—the comprehensive scoring calculator integrates and merges the evaluation results from each dimension. This calculator, acting as a multi-dimensional scoring fusion engine, runs within the numerical computation service environment to generate the final, detailed comprehensive score.

[0110] Specifically, the comprehensive scoring calculation process adopts a weighted linear fusion algorithm: The system assigns corresponding weight coefficients to each evaluation dimension: weight for completeness of technical elements. Set to 0.4, implementation detail sufficiency weight Set to 0.4, logical consistency weight The value is set to 0.2; specifically, some embodiments of this disclosure do not limit the setting of the weight.

[0111] The system calculates the total score for comprehensiveness based on the weighted fusion formula:

[0112] The system determines the level of detail based on the total score: when the total score is greater than or equal to 70, the system determines it as "sufficient detail"; when the total score is between 40 and 70, the system determines it as "moderate detail, it is recommended to supplement relevant technical details"; when the total score is less than 40, the system determines it as "insufficient detail, the technical solution description needs to be improved".

[0113] It should be noted that the scoring range for the total score of detail is for illustrative purposes only and is not a specific limitation.

[0114] In some embodiments, step 204 involves extracting technical keywords from the file to be processed, and classifying the technical keywords into IPCs based on a multi-label classification algorithm that combines patent semantic features and technical field features. The resulting target IPC classification can be achieved, but is not limited to, the following methods: After the comprehensive assessment is completed, the keyword extraction module immediately begins the automatic extraction process of technical keywords from the disclosure document. This module is built based on the TF-IDF feature extraction algorithm and runs in a text mining server environment to generate a list of core technical keywords and their corresponding weights.

[0115] Specifically, the keyword extraction process includes the following technical steps: First, the system uses the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm to calculate the importance weight of each word in the text. The calculation formula is as follows:

[0116] in Words In the document Frequency of occurrence in Indicates the total number of documents in the collection. Indicates included words The number of documents; Secondly, the system combines the TextRank graph ranking algorithm to extract key technical phrases, constructs a word co-occurrence graph by setting a sliding window with a window size of 5, and applies the PageRank algorithm to calculate the importance ranking of word nodes to identify representative technical phrases; Furthermore, the system utilizes the Word2Vec word vector model to semantically expand the extracted keywords, automatically identifying synonyms and semantically related words, expanding the vocabulary to 150% of the original number of keywords, thereby enhancing the completeness of keyword coverage. Finally, the system performs a general vocabulary filtering process, which removes non-professional general terms based on the dictionary of patent technology, retains technical terms, and finally outputs the top 20 core keywords by weight and their corresponding TF-IDF weight scores.

[0117] After the technical keywords are extracted, the IPC classification predictor immediately performs International Patent Classification (IPC) prediction analysis on the text content. This predictor is built on a specially fine-tuned BERT text classification model and runs in a deep learning classification environment on a GPU inference service to predict the most suitable IPC classification number and its probability distribution.

[0118] Specifically, the IPC classification prediction process includes the following technical steps: First, the system concatenates the extracted technical keywords with the original text content to construct a text sequence that conforms to the input format of the BERT model: Input sequence = ,in Indicates classification label, Indicates a sequence separator; Secondly, the system calls a fine-tuned BERT classification model trained on 2 million labeled patent-IPC correspondence data, which can cover the main classification range of the international patent classification system; Finally, the system passes The activation function converts the model output into a probability distribution and selects the top 3 IPC classification numbers with the highest probability values ​​as candidate prediction results, ensuring the accuracy and diversity of classification predictions.

[0119] After the IPC classification number is predicted, the business matching analyzer immediately calculates the degree of matching between the predicted classification and the enterprise's main business. This analyzer is built on a semantic similarity calculation engine and runs in the semantic matching environment of the vector retrieval service to quantitatively evaluate the relevance level between the technical solution and the enterprise's business.

[0120] Specifically, the business matching degree calculation process includes the following technical steps: First, the system builds a comprehensive dictionary covering the enterprise's business domains. This dictionary contains 500 keywords related to the enterprise's main products, core technology areas, and main application scenarios, providing basic knowledge support for business matching analysis. Secondly, the system uses a GTE (General Text Embeddings) pre-trained model to perform vector encoding processing on the standard description text of IPC classification and the enterprise business description text, respectively, converting the text content into a 512-dimensional dense vector representation; Next, the system calculates the cosine similarity between the IPC classification vector and the enterprise business vector. The calculation formula is as follows:

[0121] Where θ represents the angle between the two vectors, and the cosine similarity range is [-1, 1]; Finally, the system determines the business matching level based on the calculated similarity scores:

[0122] After the business matching degree is calculated, the classification verification module immediately verifies and optimizes the IPC prediction results. This module is built on an expert rule validator and runs in the rule verification environment of the expert knowledge base system to identify and correct obviously erroneous classification prediction results.

[0123] Specifically, the classification verification and optimization process includes the following technical steps: First, the system loads the IPC classification verification rule base built by patent classification experts. This rule base contains 500 verification rules, covering common classification error patterns and correction logic. Secondly, the system performs rule-based anomaly detection. For example, when the technical keywords contain keywords such as "electronic devices" and "semiconductors," and the IPC prediction result is "mechanical engineering," the system automatically marks the prediction result as a classification anomaly. Furthermore, the system employs an ensemble learning approach to fuse the prediction results of multiple independent classification models, and integrates the probability outputs of each model through a soft voting mechanism to improve the stability and accuracy of classification prediction. Finally, for the detected abnormal classifications, the system recalculates the probability ranking of the top 5 candidate classifications and, in conjunction with the business matching degree analysis results, selects the classification with the highest degree of matching with the enterprise's business as the final prediction result.

[0124] In some embodiments, when performing analysis on the document to be processed in step 104 to obtain the evaluation results of the novelty and inventiveness of the document to be processed, the following methods may be used, but are not limited to: Figure 7 As shown, it includes: Step 601: Perform paragraph-level semantic vectorization processing on the content of the file to be processed and the uploaded comparison file.

[0125] After the user uploads prior art application documents for comparison (pre-existing documents), the system performs paragraph-level semantic vectorization processing on the patent disclosure content and the uploaded pre-existing documents. This step is performed based on a text encoder built using a pre-trained language model, converting the text content into a numerical vector representation.

[0126] Specifically, the semantic vectorization process includes the following technical steps: First, the system performs paragraph segmentation on the disclosure document and the comparison document respectively. It adopts natural sentence segmentation rules based on periods and semicolons, combined with technical terminology boundary recognition algorithm, to decompose long documents into semantically complete paragraph units. The character length of a single paragraph is controlled within the range of 200-500 characters, with no specific limit. Secondly, the system calls the GTE-large text encoding model based on the Transformer architecture to perform vector encoding processing on each text segment. This model contains 24 Transformer encoder layers with a hidden layer dimension of 1024, generating a dense vector representation of 512 dimensions. Next, the system performs L2 normalization on the generated vector using the normalization vector formula:

[0127] Ensure that the Euclidean norm of all vectors is 1 to eliminate the influence of vector magnitude differences on similarity calculation; Finally, the system stores the vector data in a structured format: It also establishes a memory index structure based on NumPy arrays to support fast vector retrieval operations.

[0128] Step 602: Calculate the semantic similarity between paragraphs of the document to be processed and paragraphs of the comparison document, and determine the evaluation result of the novelty based on the semantic similarity.

[0129] After semantic vectorization, the system calculates the semantic similarity between paragraphs in the disclosure document and paragraphs in the comparison document. This step uses a vector similarity calculation algorithm to quantify the degree of semantic association between paragraphs through numerical calculation.

[0130] Specifically, the similarity calculation process includes the following technical steps: First, the system uses a nested loop algorithm to traverse all paragraph pair combinations: for each paragraph vector in the disclosure document... , compared with the vector of each paragraph in the comparison file Perform pairwise calculations; Secondly, the system uses the cosine similarity algorithm to calculate the semantic similarity between vectors. The calculation formula is as follows:

[0131] Finally, the system makes a classification based on the similarity score:

[0132] Step 603: Automatically identify and extract core technical features from the file to be processed, and generate a structured list of technical features.

[0133] After the novelty analysis is completed, the technical feature extractor immediately and automatically identifies and extracts the core technical features from the disclosure document. This extractor, as a feature recognition and extraction module, runs in the feature extraction environment of the NLP feature processing service, generating a structured list of technical features.

[0134] Specifically, the technical feature extraction process includes the following technical steps: First, the system uses a conditional random field model to perform technical term boundary identification and uses the BIO annotation system: B-TECH identifies the starting words of technical features, I-TECH identifies the words inside technical features, and O identifies non-technical feature words. The technical term boundaries are accurately located through sequence annotation. Secondly, the system combines dependency parsing algorithms to extract modification and dependency relationships between technical features, constructs a syntactic relationship graph of technical features, and identifies the hierarchical structure and logical connections between features; Secondly, the system applies predefined rule templates to match key technology description patterns, such as "achieved through... method...", "using... technology...", "based on... principle..." and other typical technical description sentences, to improve the accuracy of feature recognition; Then, the system calculates the importance weight of technical features based on the TF-IDF algorithm and positional weight information. The positional weight is assigned different weights according to the position of the feature in the document (summary, technical field, technical solution, etc.). Next, the system uses the Word2Vec semantic similarity calculation method to deduplicatize and classify the extracted technical features, merge semantically similar duplicate features, and improve the accuracy of the feature list. Finally, the system outputs a standardized list of technical features, each feature containing: Structured information, etc.

[0135] Step 604: Match the structured list of technical features with the technical features in the comparison file to identify the differences between the technical features.

[0136] After the core technical features are extracted, the feature matching engine immediately performs a precise matching analysis between the technical features in the disclosure document and the technical features in the comparative literature. This engine is built on a semantic feature matcher and runs in the feature comparison environment of a distributed feature matching service to identify the same, similar, and different relationships between technical features.

[0137] Specifically, the feature matching process includes the following technical steps: First, the system constructs a high-dimensional technical feature vector space, and encodes each identified technical feature into a 768-dimensional dense vector representation through a BERT pre-trained model to ensure that the semantic information of the feature is fully preserved. Secondly, the system uses the cosine similarity algorithm to calculate the semantic matching degree between the feature vector of the disclosure document and the feature vector of the comparison document, and quantifies the similarity level between features; Secondly, the system performs feature matching and classification based on similarity scores:

[0138] when When, it indicates the same technical features; when When, it indicates similar technical features; when similar When, the symbol indicates some related technical features; when , indicating differentiated technical features; Then, the system generates a global matching matrix between the features of the disclosure document and the features of the comparison documents. This matrix displays the similarity relationship between all feature pairs in the form of a two-dimensional table. Finally, the system identifies and marks unique technical features in the disclosure document, which may constitute the innovative points and creative contributions of the technical solution.

[0139] Step 605: Perform importance assessment analysis on the identified difference features, quantify the contribution of each difference feature to the overall creativity, and obtain the evaluation result of the creativity.

[0140] After feature matching is complete, the feature weight analyzer immediately performs an importance assessment of the identified differential features. This analyzer, as a feature importance assessment module, runs within the weight calculation environment of the machine learning feature analysis service, quantifying the contribution of each differential feature to the overall creativity.

[0141] Specifically, the differential feature weight analysis adopts a multi-factor comprehensive weight calculation model: The system calculates the importance weight of features based on their frequency of occurrence in the document. Location importance and syntax dependencies Determined by a combination of factors:

[0142] The system further combines the innovativeness, practicality, and technical difficulty of the technical features for comprehensive evaluation, generating a final creative contribution score for each difference feature, providing a quantitative basis for subsequent creative analysis.

[0143] The system integrates the outputs of various modules and automatically generates a structured pre-examination report, including object assessment, detail scoring, IPC classification, novelty and inventiveness analysis, and provides export functionality. This module is triggered after all analyses are completed and is executed by the pre-examination report generation module, which generates the pre-examination report through a background algorithm.

[0144] Some embodiments disclosed herein can achieve the following beneficial effects: First, the methods described in some embodiments of this disclosure enable automated generation of pre-examination reports: The semantic comparison results, regulatory compliance verification conclusions, and technical element integrity analysis results are uniformly integrated into the templated report generation module. Unlike existing technologies that require manual summarization, some embodiments of this disclosure automate the entire process from document parsing, feature comparison, compliance verification, to report generation, reducing manual intervention and improving review efficiency.

[0145] Furthermore, some embodiments of this disclosure improve the accuracy of semantic understanding and comparison in complex technical fields: This paper employs a BERT vectorization model combined with a BiLSTM-CRF entity recognition model to perform deep modeling and extraction of keywords, functional features, and semantic relationships in the technical solution. Compared with comparative documents that rely solely on shallow feature or keyword matching, some embodiments of this disclosure can more accurately identify the differences and similarities between technical features, thereby improving the accuracy of judging novelty and inventiveness.

[0146] Corresponding to the aforementioned patent pre-examination evaluation method, this invention also proposes a patent pre-examination evaluation device. Since the device embodiments of this invention correspond to the aforementioned method embodiments, details not disclosed in the device embodiments can be referred to the aforementioned method embodiments, and will not be repeated here.

[0147] Figure 8 This is a schematic diagram of the structure of a patent pre-examination evaluation device provided in some embodiments of this disclosure, such as... Figure 8 As shown, it includes: Preprocessing unit 61 is used to perform vectorization preprocessing on the file to be processed to obtain a structured sequence corresponding to the text content of the file to be processed; Determination unit 62 is used to determine the patent subject matter corresponding to the structured sequence based on the patent law clause rule base and semantic determination model; Evaluation unit 63 is used to evaluate the implementation details and logical relationships of the technical solution in response to the patent subject matter determination, and to obtain the technical solution detail evaluation result. The first determining unit 64 is used to analyze the document to be processed and obtain the evaluation results of the novelty and inventiveness of the document to be processed. The generation unit 65 is used to generate a structured patent pre-examination report containing the pre-examination conclusion based on the patent subject matter, the technical solution detail assessment result, and the novelty and inventiveness assessment result, in response to the passing of the novelty and inventiveness assessment result.

[0148] The patent pre-examination evaluation apparatus provided in this disclosure performs vectorized preprocessing on the document to be processed to obtain a structured sequence corresponding to the text content of the document; based on the patent law clause rule base and semantic judgment model, it determines the patent subject matter corresponding to the structured sequence; in response to the patent subject matter being determined to be acceptable, it evaluates the implementation details and logical relationships of the technical solution to obtain a technical solution detail evaluation result; it analyzes the document to be processed to obtain a novelty and inventiveness evaluation result; in response to the novelty and inventiveness evaluation result being acceptable, it generates a structured patent pre-examination report containing a pre-examination conclusion based on the patent subject matter, the technical solution detail evaluation result, and the novelty and inventiveness evaluation result. The methods provided in some embodiments of this disclosure can achieve fully automated processing of the patent pre-examination process, significantly improving pre-examination efficiency and result consistency, while reducing labor costs and providing a scientific basis for patent examination and application decisions.

[0149] Furthermore, in one possible implementation of some embodiments of this disclosure, such as Figure 9 As shown, the device further includes: The second determining unit 66 is used to, after evaluating the implementation details and logical relationships of the technical solution and obtaining the technical solution detail evaluation result, in response to the passing of the technical solution detail evaluation result, extract the technical keywords in the file to be processed, and perform IPC classification on the technical keywords based on a multi-label classification algorithm to obtain the target IPC classification; The generation unit 65 is further configured to generate the structured patent pre-examination report containing the pre-examination conclusion based on the patent subject matter, the technical solution detail assessment results, the target IPC classification, and the novelty and inventiveness assessment results.

[0150] Furthermore, in one possible implementation of some embodiments of this disclosure, such as Figure 9 As shown, the preprocessing unit 61 includes: Conversion module 611 is used to convert the file to be processed into plain text format data; The word segmentation module 612 is used to perform word segmentation and standardization operations on the plain text format data to obtain the word segmentation result. The processing module 613 is used to perform vectorization preprocessing on the word segmentation results to obtain the structured sequence corresponding to the text content.

[0151] Furthermore, in one possible implementation of some embodiments of this disclosure, such as Figure 9 As shown, the processing module 613 is further configured to: The word segmentation results are processed based on the whole word masking model to obtain an attention mask matrix of a preset sequence length. The attention mask matrix contains the positions of valid words and filler words that mark the word segmentation results. The attention mask matrix is ​​vectorized to obtain the structured sequence corresponding to the text content.

[0152] Furthermore, in one possible implementation of some embodiments of this disclosure, such as Figure 9 As shown, the determination unit 62 includes: The filtering module 621 is used to classify and filter the text content in the file to be processed based on the rule matching engine and the patent law clause rule library, and to mark the filtering results as needing in-depth judgment. Encoding module 622 is used to input the text content marked as depth determination into the semantic determination model for encoding processing to obtain semantic determination result; The verification module 623 is used to perform regulatory compliance verification on the semantic determination result to obtain the regulatory compliance verification result; The determination module 624 is used to determine whether the document to be processed belongs to the patent subject matter based on the semantic determination result and the regulatory compliance inspection result.

[0153] Furthermore, in one possible implementation of some embodiments of this disclosure, such as Figure 9 As shown, the encoding module 622 is further used for: The text content is input into the semantic determination model for encoding processing, and the hidden state vector corresponding to the target word of the text content is extracted; wherein, the hidden state vector contains the global semantic representation of the entire text sequence; The hidden state vector is classified based on a linear classification layer to obtain the classification result; The classification results are subjected to probability normalization to determine the semantic determination result that the document to be processed belongs to the patent subject matter.

[0154] Furthermore, in one possible implementation of some embodiments of this disclosure, such as Figure 9As shown, the determining module 624 is further configured to: The semantic judgment results are similar to those obtained by performing similarity calculations based on a pre-defined knowledge graph database. The knowledge graph database stores entities and their relationships among at least one of legal provisions, violation cases, and keywords in a node-relationship format. The compliance verification result of the regulations is obtained based on the similarity calculation result.

[0155] Furthermore, in one possible implementation of some embodiments of this disclosure, such as Figure 9 As shown, the evaluation unit 63 is further used for: The technical elements in the file to be processed are identified based on a preset technical element tagging system of the patent text; wherein, the technical elements constitute the technical solution; Based on the technical elements in the file to be processed, a technical element integrity score is determined; By constructing a sentence similarity graph and applying the PageRank algorithm to identify key sentences in the technical elements, and calculating the technical depth score based on the technical content and description depth of the key sentences to determine the sufficiency score of implementation details; The technical solution is converted into a directed graph, wherein the nodes of the directed graph represent technical steps, technical features or implementation elements, and the edges of the graph represent logical relationships between technical elements, including at least one of causal relationships, dependency relationships and temporal relationships. The directed graph is analyzed and processed. The relational representation between nodes is learned through graph convolutional networks to detect logical loops and inconsistent reasoning paths in the directed graph. A semantic conflict detection algorithm is used to identify contradictory sentence pairs in the text. The ratio of the number of contradictory sentence pairs to the total number of sentence pairs in the text is calculated to obtain a logical consistency score. The results of the detailed assessment of the technical solution are determined based on the completeness score of the technical elements, the sufficiency score of the implementation details, and the logical consistency score.

[0156] Furthermore, in one possible implementation of some embodiments of this disclosure, such as Figure 9 As shown, the first determining unit 64 is further configured to: The content of the file to be processed and the uploaded comparison file are subjected to paragraph-level semantic vectorization processing; Calculate the semantic similarity between paragraphs of the file to be processed and paragraphs of the comparison file, and determine the novelty evaluation result based on the semantic similarity; The core technical features are automatically identified and extracted from the file to be processed, and a structured list of technical features is generated. The structured list of technical features is matched and analyzed with the technical features in the comparison file to identify the differences between the technical features; An importance assessment analysis is performed on the identified differential features, and the contribution of each differential feature to the overall creativity is quantitatively calculated to obtain the evaluation result of the creativity.

[0157] It should be noted that the foregoing explanation of the method embodiments also applies to the apparatus of some embodiments of this disclosure, and the principle is the same. Therefore, some embodiments of this disclosure are no longer limited.

[0158] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0159] Figure 10 A schematic block diagram of an example electronic device 700 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0160] like Figure 10 As shown, the electronic device 700 includes a computing unit 701, which can perform various appropriate actions and processes based on a computer program stored in ROM (Read-Only Memory) 702 or loaded from storage unit 708 into RAM (Random Access Memory) 703. The RAM 703 can also store various programs and data required for the operation of the electronic device 700. The computing unit 701, ROM 702, and RAM 703 are interconnected via a bus 704. An I / O (Input / Output) interface 705 is also connected to the bus 704.

[0161] Multiple components in electronic device 700 are connected to I / O interface 705, including: input unit 706, such as keyboard, mouse, etc.; output unit 707, such as various types of displays, speakers, etc.; storage unit 708, such as disk, optical disk, etc.; and communication unit 709, such as network card, modem, wireless transceiver, etc. Communication unit 709 allows electronic device 700 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0162] The computing unit 701 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 701 include, but are not limited to, CPUs (Central Processing Units), GPUs (Graphics Processing Units), various special-purpose AI (Artificial Intelligence) computing chips, various computing units running machine learning model algorithms, DSPs (Digital Signal Processors), and any suitable processor, controller, microcontroller, etc. The computing unit 701 performs the various methods and processes described above, such as patent pre-examination evaluation methods. For example, in some embodiments, the patent pre-examination evaluation methods can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 708. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 700 via ROM 702 and / or communication unit 709. When the computer program is loaded into RAM 703 and executed by the computing unit 701, one or more steps of the methods described above can be performed. Alternatively, in other embodiments, the computing unit 701 may be configured to perform the aforementioned patent pre-examination evaluation method by any other suitable means (e.g., by means of firmware).

[0163] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application-Specific Standard Products), SOCs (System-on-Chips), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0164] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0165] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, RAM, ROM, EPROM (Electrically Programmable Read-Only Memory) or flash memory, optical fiber, CD-ROM (Compact Disc Read-Only Memory), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0166] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (Cathode-Ray Tube) or LCD (Liquid Crystal Display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0167] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include LANs (Local Area Networks), WANs (Wide Area Networks), the Internet, and blockchain networks.

[0168] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. A server can be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service system that addresses the shortcomings of traditional physical hosts and VPS (Virtual Private Server) services, such as high management difficulty and weak business scalability. Servers can also be servers for distributed systems or servers incorporating blockchain technology.

[0169] It's important to note that artificial intelligence (AI) is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. AI software technologies primarily include computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.

[0170] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0171] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for evaluating patent pre-examination, characterized in that, include: The file to be processed is preprocessed into vectorization to obtain a structured sequence corresponding to the text content of the file to be processed; Based on the patent law clause rule base and semantic determination model, the patent subject matter corresponding to the structured sequence is determined; In response to the successful determination of the patent subject matter, the implementation details and logical relationships of the technical solution are evaluated to obtain the results of the technical solution's comprehensiveness assessment. The document to be processed is analyzed to obtain the evaluation results of its novelty and inventiveness. In response to the successful evaluation of novelty and inventiveness, a structured patent pre-examination report containing the pre-examination conclusion is generated based on the patent subject matter, the detailed evaluation results of the technical solution, and the evaluation results of novelty and inventiveness.

2. The method according to claim 1, characterized in that, After evaluating the implementation details and logical relationships of the technical solution to obtain the results of the technical solution's comprehensiveness assessment, the method further includes: In response to the passing of the technical solution detail evaluation result, the technical keywords in the file to be processed are extracted, and the technical keywords are classified into IPCs based on a multi-label classification algorithm to obtain the target IPC classification; Based on the patent subject matter, the results of the detailed assessment of the technical solution, and the results of the novelty and inventiveness assessment, a structured patent pre-examination report containing a pre-examination conclusion is generated, including: Based on the patent subject matter, the results of the technical solution detail assessment, the target IPC classification, and the results of the novelty and inventiveness assessment, a structured patent pre-examination report containing the pre-examination conclusion is generated.

3. The method according to claim 1, characterized in that, The vectorization preprocessing of the file to be processed to obtain the structured sequence corresponding to the text content includes: Convert the file to be processed into plain text format data; The plain text data is segmented and standardized to obtain the segmentation results. The word segmentation results are preprocessed into vectors to obtain the structured sequence corresponding to the text content.

4. The method according to claim 3, characterized in that, The step of performing vectorization preprocessing on the word segmentation results to obtain the structured sequence corresponding to the text content includes: The word segmentation results are processed based on the whole word masking model to obtain an attention mask matrix of a preset sequence length. The attention mask matrix contains the positions of valid words and filler words that mark the word segmentation results. The attention mask matrix is ​​vectorized to obtain the structured sequence corresponding to the text content.

5. The method according to claim 1, characterized in that, The determination of the patent subject matter corresponding to the structured sequence based on the patent law rule base and semantic determination model includes: Based on the rule matching engine and the patent law clause rule library, the text content in the file to be processed is classified and filtered, and the filtering results are marked as to whether in-depth judgment is required. For a text sequence marked as depth-based, the text content is input into the semantic determination model for encoding processing to obtain the semantic determination result; The semantic determination result is then subjected to regulatory compliance verification to obtain the regulatory compliance verification result. Based on the semantic determination result and the regulatory compliance verification result, the determination result of whether the document to be processed belongs to patent subject matter is obtained.

6. The method according to claim 5, characterized in that, For the text sequence marked as depth-based, the text content is input into the semantic determination model for encoding processing to obtain the semantic determination result, including: The text content is input into the semantic determination model for encoding processing, and the hidden state vector corresponding to the target word of the text content is extracted; wherein, the hidden state vector contains the global semantic representation of the entire text sequence; The hidden state vector is classified based on a linear classification layer to obtain the classification result; The classification results are subjected to probability normalization to determine the semantic determination result that the document to be processed belongs to the patent subject matter.

7. The method according to claim 5, characterized in that, The step of performing regulatory compliance verification on the semantic determination result to obtain the regulatory compliance verification result includes: The semantic judgment results are similar to those obtained by performing similarity calculations based on a pre-defined knowledge graph database. The knowledge graph database stores entities and their relationships among at least one of legal provisions, violation cases, and keywords in a node-relationship format. The compliance verification result of the regulations is obtained based on the similarity calculation result.

8. The method according to claim 1, characterized in that, In response to the successful determination of the patent subject matter, the implementation details and logical relationships of the technical solution are evaluated to obtain the results of the technical solution's comprehensiveness assessment, including: The technical elements in the file to be processed are identified based on a preset technical element tagging system of the patent text; wherein, the technical elements constitute the technical solution; Based on the technical elements in the file to be processed, a technical element integrity score is determined; By constructing a sentence similarity graph and applying the PageRank algorithm to identify key sentences in the technical elements, and calculating the technical depth score based on the technical content and description depth of the key sentences to determine the sufficiency score of implementation details; The technical solution is converted into a directed graph, wherein the nodes of the directed graph represent technical steps, technical features or implementation elements, and the edges of the graph represent logical relationships between technical elements, including at least one of causal relationships, dependency relationships and temporal relationships. The directed graph is analyzed and processed. The relational representation between nodes is learned through graph convolutional networks to detect logical loops and inconsistent reasoning paths in the directed graph. A semantic conflict detection algorithm is used to identify contradictory sentence pairs in the text. The ratio of the number of contradictory sentence pairs to the total number of sentence pairs in the text is calculated to obtain a logical consistency score. The results of the detailed assessment of the technical solution are determined based on the completeness score of the technical elements, the sufficiency score of the implementation details, and the logical consistency score.

9. The method according to claim 1, characterized in that, The analysis of the document to be processed to obtain the evaluation results of the novelty and inventiveness of the document to be processed includes: The content of the file to be processed and the uploaded comparison file are subjected to paragraph-level semantic vectorization processing; Calculate the semantic similarity between paragraphs of the file to be processed and paragraphs of the comparison file, and determine the novelty evaluation result based on the semantic similarity; The core technical features are automatically identified and extracted from the file to be processed, and a structured list of technical features is generated. The structured list of technical features is matched and analyzed with the technical features in the comparison file to identify the differences between the technical features; An importance assessment analysis is performed on the identified differential features, and the contribution of each differential feature to the overall creativity is quantitatively calculated to obtain the evaluation result of the creativity.

10. An evaluation device for patent pre-examination, characterized in that, include: The preprocessing unit is used to perform vectorized preprocessing on the file to be processed to obtain a structured sequence corresponding to the text content of the file to be processed. The determination unit is used to determine the patent subject matter corresponding to the structured sequence based on the patent law clause rule base and semantic determination model; The evaluation unit is used to evaluate the implementation details and logical relationships of the technical solution in response to the successful determination of the patent subject matter, and to obtain the evaluation result of the completeness of the technical solution. The first determining unit is used to analyze the document to be processed and obtain the evaluation results of the novelty and inventiveness of the document to be processed. The generation unit is used to generate a structured patent pre-examination report containing the pre-examination conclusion based on the patent subject matter, the technical solution detail assessment result, and the novelty and inventiveness assessment result, in response to the passing of the novelty and inventiveness assessment result.

Images