Intelligent system for improving patent classification quality

Abstract

The invention relates to the technical field of patent classification and artificial intelligence, in particular to an intelligent system for improving patent classification quality, which comprises a patent information extraction module, a patent information synchronization and retrieval module, a patent information extraction module and a patent classification module, all the modules work cooperatively and are in communication connection with the local patent database, the authoritative patent data source and the LLM. The patent information extraction module supports multi-format input and obtains complete patent information; the synchronization and retrieval module realizes patent information synchronization, storage and vector search; the information extraction module is combined with the optimization cue word to call LLM, and patent core key points are extracted in a specified length; and the classification module adopts a patentClassifier model subjected to patent data fine adjustment, and outputs a multi-classification label with possibility. According to the system, the problems of low recall rate and single input format caused by input length limitation in the prior art are solved, patent classification accuracy, integrity and efficiency are improved, and the system is suitable for various application scenes and high in practical value.

Description

Technical Field

[0001] This invention relates to the fields of patent classification and artificial intelligence technology, and in particular to an intelligent system for improving the quality of patent classification. Background Technology

[0002] In today's era of knowledge explosion, patent classification is of great significance and value. It not only allows for more precise matching of user needs but also reduces the time users spend searching for the technologies they require. With tens of thousands of new patents generated every year, how to classify patents quickly, accurately, and comprehensively has become a major challenge.

[0003] Currently, there are numerous studies relying on AI algorithms to automatically identify patents. These studies primarily utilize patent titles and abstracts, or claims from the patent text, feeding them into AI models like BERT, which excel at classification tasks, to obtain multi-label classification results. However, limited by the length constraints of the input data processed by AI models, the models only extract the first N tokens (words after segmentation) of the original input data. This easily leads to the loss of important information, making it impossible to classify patents comprehensively and accurately. Test results from related studies show that existing solutions have high precision but low recall. The fundamental reason is that the actual information input into the classification model is incomplete, resulting in the omission of key classification information. Furthermore, most existing systems lack support for multiple patent input formats, reducing the ease of use and flexibility of the system in practical applications.

[0004] Therefore, there is an urgent need for an intelligent patent classification system that can comprehensively extract patent information, accurately extract core points, and adapt to multiple input formats. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems of the prior art and provide an intelligent system to improve the quality of patent classification. This system effectively solves the technical problems of existing patent classification systems, such as limited input length leading to the omission of key information, a lack of adaptability in single input formats, and insufficient classification accuracy. Through multi-module collaborative work, combining the summarizing capabilities of LLM with a specially fine-tuned classification model, it achieves a dual improvement in patent classification quality and efficiency, while adapting to multiple input formats to meet the needs of different usage scenarios.

[0006] The above objectives are achieved through the following technical solutions: An intelligent system for improving the quality of patent classification includes a patent information extraction module, a patent information synchronization and retrieval module, a patent information refinement module, and a patent classification module that work in sequence and in collaboration. The patent information synchronization and retrieval module is communicatively connected to a local patent database and a remote authoritative patent data source. The patent information refinement module and the patent classification module are both communicatively connected to a Large Language Model (LLM). The patent information extraction module is used to receive at least one user input, including patent number, PDF / Word / TXT format patent file, and patent structured information, to obtain complete patent information; The patent information synchronization and retrieval module is used to obtain or periodically synchronize patent information from the authoritative patent data source to the local patent database, and provides a content-based vector search function. The patent information extraction module is used to receive structured patent information, input system prompts that are optimized and adapted to the CPC classification standard, and call the core patent points extracted by the large language model LLM. The core patent points include the patent's field of application, the key problems it solves, the innovative methods / technologies / tools proposed, and title information, which helps to identify the patent as comprehensively and accurately as possible. The patent classification module is used to call the complete patent information obtained by the patent information extraction module. If the patent information contains classification information and the user does not choose to regenerate the structured patent information, the classification result is returned directly. If the patent information is unstructured text, the patent structured information is extracted by the Large Language Model (LLM). The core points output by the patent information extraction module are then input into the PatentClassifier, a patent classification model based on BERT fine-tuning. Multiple classification labels and their corresponding probabilities are output according to the minimum probability of the labels specified by the user.

[0007] As a further optimization of this system, the system prompt words are specifically as follows: You are a patent expert who is very familiar with the CPC's patent classification standards. Based on the input structured patent information, please summarize the information that helps to correctly classify the patent. This information should include the field involved in the patent, the key problem solved, the innovative methods, technologies or tools proposed, title information, etc., which will help to determine the patent category as comprehensively and accurately as possible. The patent category should be distinguished into the CPC subclass. For example, A01B represents agricultural machinery, tools, and common components related to soil cultivation; the input patent structured information is in JSON format, as follows: { "title": <Patent title information> “abstract”: <Patent abstract> “claim”: <claims> "description": <Patent details> } In addition, the information in the summary must ensure that its maximum length does not exceed 500 characters, including punctuation marks.

[0008] As a further optimization of this system, the large language model LLM is an open-source large model deployed locally or an API service provided by a large model service provider. The open-source large models include Meta's LLaMA and Alibaba's Qwen, and the API services include those provided by Kimi and DeepSeek.

[0009] As a further optimization of this system, the PatentClassifier is a patent classification-specific model obtained by extracting key information from a large amount of labeled and classified patent data and then fine-tuning the original BERT model through multiple rounds of training.

[0010] As a further optimization of this system, the patent information extraction module obtains complete patent information in the following way: If the input is a patent number, the local patent database will be queried first. If the patent information is not found locally, the patent information synchronization and retrieval module will be called to obtain it from the remote authoritative patent data source and return the patent structured information in JSON format. The patent structured information may optionally include classification labels. If the input is a PDF / Word / TXT file, the complete patent text information will be extracted. If the input is structured patent information in JSON format, then a check for structural validity and completeness will be performed.

[0011] As a further optimization of this system, the fields of the JSON format structured information include "title" (patent title), "abstract" (patent abstract), "claim" (claims) and "description" (patent details).

[0012] As a further optimization of this system, the synchronization cycle of the patent information synchronization and retrieval module can be customized, and the synchronization content includes new patent information and updates to patent classification information.

[0013] As a further optimization of this system, in the patent classification module, after the unstructured text is processed by the Large Language Model (LLM) to extract structured information, the user can confirm that the information is correct before proceeding with subsequent processing. By default, subsequent processing is performed directly.

[0014] As a further optimization of this system, the content-based vector search function converts patent text information into vector representations and uses vector similarity algorithms to achieve in-depth patent content retrieval.

[0015] As a further optimization of this system, the classification labels output by the patent classification module are distinguished into CPC subcategories, and each classification label corresponds to a unique category meaning description.

[0016] This invention provides an intelligent system for improving patent classification quality. It achieves flexible and convenient application through multi-module collaboration and multi-format input adaptation; improves classification recall and precision by extracting core elements through LLM; enhances efficiency through automated processes and reuse of historical classifications; and upgrades practical value by refining classification labels and probability output. Specific beneficial effects are as follows: 1. Solves the problem of information omission caused by input length limitation in existing technologies: By calling LLM through the patent information extraction module, the core points of the patent are accurately extracted within the specified length, avoiding the loss of key information caused by the traditional model truncating the first N tokens, which significantly improves the recall rate of patent classification. At the same time, combined with the specially fine-tuned PatentClassifier model, the classification accuracy is maintained at a high level, achieving a dual improvement in classification accuracy and completeness.

[0017] 2. Adapts to multiple patent input formats: Supports multiple input types such as patent number, PDF / Word / TXT files, and JSON structured information, meeting the usage needs of different users in different scenarios, greatly improving the system's practical application flexibility and convenience, and lowering the user's usage threshold.

[0018] 3. Improve the efficiency of patent information acquisition and retrieval: Through the patent information synchronization and retrieval module, the local database and authoritative data sources are synchronized regularly to ensure data timeliness; the content-based vector search function allows users to quickly and accurately find relevant patents, solving the problems of low efficiency and insufficient accuracy of traditional retrieval methods.

[0019] 4. Construct an automated classification pipeline: Through the collaborative work of various modules, the entire process from patent information input, extraction, refinement to classification output is automated. At the same time, it supports the reuse of historical classification information, reduces repetitive calculations, replaces the heavy work of manual classification, significantly improves patent classification efficiency, and reduces labor costs.

[0020] 5. High practical value of output results: The classification results include multiple classification labels and corresponding possibilities, and the classification labels are distinguished down to CPC subcategories, providing users with more detailed and valuable classification information. This can assist users in subsequent work such as patent analysis and technology research, thereby enhancing the practical value of the system. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the structure of an intelligent system for improving the quality of patent classification as described in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. The described embodiments are merely some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, this solution provides an intelligent system for improving the quality of patent classification. The system includes a patent information extraction module, a patent information synchronization and retrieval module, a patent information refinement module, a patent classification module, a local patent database, and a communication interface with a remote authoritative patent data source. The modules work together to form a complete patent classification processing pipeline.

[0024] The core function of the patent information extraction module in this system is to obtain complete patent information based on the patent-related information input by the user. The specific processing method varies depending on the type of input information, and includes: 1. Input is a patent number: The system first queries the local patent database. If the local database contains complete information about the patent, it retrieves it directly from the local database. If the local database does not contain the patent information, the system calls the patent information synchronization and retrieval module to obtain the complete information from a remote, authoritative patent data source and stores the obtained patent information in the local patent database. At this point, the system returns structured patent information in JSON format to the user or subsequent modules, containing the following information: { "title": "<Patent Title>" “abstract”: “<Patent Abstract>” “claim”: “<claims>” “description”: “<Patent Details>” } If the database contains classification information for the patent, the structured information will also include an optional field "labels" (a list of classification labels, in the format [label1, label2, ..., labelN]), with the following structure: { "title": "<Patent Title>" “abstract”: “<Patent Abstract>” “claim”: “<claims>” “description”: “<Patent Details>” “labels”: [<label1>, <label2>, ... <labelN>] } The labels are optional and depend on whether the information is contained in the database. Users can then choose whether to use this classification information directly to avoid redundant classification operations.

[0025] 2. Input patent documents in PDF format: Through file parsing technology, complete patent text information is extracted from the PDF file, including all core information such as title, abstract, claims, and detailed patent content, ensuring that no information is omitted.

[0026] 3. Input is a patent document in Word format: The text extraction algorithm adapted to Word documents is used to accurately extract complete patent text information from the document, covering all core elements of the patent, ensuring the completeness and accuracy of information extraction.

[0027] 4. Input patent documents in TXT format: Directly read the complete patent text information from the file without a complex parsing process, improving information acquisition efficiency while ensuring the integrity of the text information.

[0028] 5. Input is structured patent information: This structured information is in JSON format, with fields including "title" (patent title), "abstract" (patent abstract), "claim" (claims), and "description" (patent details), structured as follows: { "title": "<Patent Title>" “abstract”: “<Patent Abstract>”, "claim": "<claims>" “description”: “<Patent Details>” } The patent information extraction module will perform validity and completeness checks on the structured information, verifying whether each field is missing and whether the format meets the requirements, to ensure that subsequent processing is based on complete and valid information.

[0029] The patent information synchronization and retrieval module described in this system is the core of the system's data support, and mainly implements the following three functions: 1. Patent Information Acquisition and Storage: Based on the patent number, complete patent information is obtained from remote authoritative patent data sources (such as the State Intellectual Property Office patent database, the European Patent Office (EPO) database, the United States Patent and Trademark Office (USPTO) database, etc.) and stored in the local patent database in a preset JSON structured format to ensure the integrity of the local database data.

[0030] 2. Regular patent information synchronization: Set a fixed synchronization cycle (such as daily or weekly) to synchronize the latest patent information from various authoritative remote patent databases to the local patent database, including new patent information and patent classification information updates, to ensure the timeliness of the local patent database and provide users with the latest and most accurate patent data support.

[0031] 3. Advanced Search Functions: Offers diverse patent search capabilities. In addition to conventional patent number searches, it also supports content-based vector search. By converting patent text information into vector representations and utilizing vector similarity algorithms (such as cosine similarity algorithms), it enables in-depth searching of patent content. Users can quickly find relevant patents based on technical keywords, innovative points, and other content, improving the accuracy and efficiency of the search.

[0032] The patent information extraction module in this system takes structured patent information as input. Its core objective is to extract the key points of the patent within a specified length limit, providing high-quality input data for the subsequent classification model. The specific implementation process is as follows: 1. Prompt Design: Carefully designed, tested, and optimized system prompts are used. The specific content of the prompt is: "You are a patent expert and you are very familiar with the CPC's patent classification standards. Based on the input structured patent information, please summarize information that helps in the correct patent classification. This information should include the patent's field, the key problem it solves, the innovative methods, technologies, or tools proposed, title information, etc., to help determine its patent category as comprehensively and accurately as possible. This category should be distinguished down to a CPC subclass, such as A01B, which represents 'agricultural machinery, tools, and general components related to soil cultivation'. The input structured patent information is in JSON format, as follows:" { "title": <Patent title information>, "abstract": <Patent abstract>, "claim": <claims> "description": <Patent details> } In addition, the information in the summary must ensure that its maximum length does not exceed 500 characters, including punctuation marks.

[0033] 2. LLM Invocation: Input the optimized prompts and structured patent information into the large language model LLM. Leveraging LLM's powerful summarization capabilities, it performs in-depth analysis and compression of core elements such as the patent title, abstract, claims, and detailed patent content, extracting concise key information. The LLM can be a locally deployed open-source large model, such as Meta's LLaMA or Alibaba's Qwen, or an API service provided by popular large model service providers, such as Kimi or DeepSeek. Users can flexibly choose according to their needs and deployment conditions.

[0034] 3. Output results: The patent key points output by LLM strictly follow the prompt word requirements, with a length not exceeding 500 characters (including punctuation marks), and contain the key information required for patent classification, which can accurately support the subsequent classification model to perform comprehensive and accurate classification.

[0035] The patent classification module described in this system is the core of the system's patent classification function. It is responsible for processing users' patent classification requests. The specific processing flow is as follows: 1. Information Preprocessing: Based on user input, the patent information extraction module is invoked to obtain complete patent information, and then targeted processing is performed according to the type of patent information: If the obtained information is structured patent information and it already contains patent classification information (i.e., the "labels" field contains valid data), and the user does not choose to regenerate the classification results, the patent classification information will be returned directly without further classification processing. If the user chooses to regenerate the classification results, or if the structured information does not contain classification information, the subsequent classification process will continue.

[0036] If the obtained patent text is unstructured and complete, the Large Language Model (LLM) is invoked to analyze it and extract structured patent information conforming to JSON format. In this case, the system provides two processing options: First, the structured information should be confirmed by the user before further processing is carried out; Secondly, it can directly perform subsequent processing. By default, the system will automatically perform subsequent processing, and users can flexibly choose according to their own requirements for information accuracy.

[0037] 2. Patent Summary Generation: For cases requiring further classification, the structured patent information is input into the patent information extraction module, which then generates a patent summary that meets the length and content requirements.

[0038] 3. Classification Label Output: The patent summary output by the patent information extraction module is used as input to the patent classification-specific model, PatentClassifier. This model is based on the original BERT model and has been fine-tuned through multiple rounds of training using a large amount of patent classification-related data (including structured information of patents with classification labels, core patent points, etc.). It is specifically optimized for patent classification tasks and has higher classification accuracy. PatentClassifier outputs multiple classification labels for the patent and the probability of each label based on the minimum probability of the labels specified by the user, providing the user with comprehensive and valuable classification results.

[0039] Example 1 Users need to classify a patent by entering the patent number "CN202310123456.7".

[0040] After receiving the patent number input, the patent information extraction module first queries the local patent database. If it finds that the local database does not store the patent information, it then calls the patent information synchronization and retrieval module.

[0041] The patent information synchronization and retrieval module obtains complete information about the patent, including title, abstract, claims, and detailed patent content, from the authoritative patent data source of the State Intellectual Property Office based on the patent number "CN202310123456.7". The patent information does not include the category tags "labels". The patent information is then stored in the local patent database in JSON structured format and fed back to the patent information extraction module.

[0042] The patent information extraction module transmits structured patent information in JSON format to the patent classification module. Since the structured information does not contain classification information, the patent classification module calls the patent information extraction module.

[0043] After receiving structured patent information, the patent information extraction module inputs the optimized system prompts along with the structured information into the locally deployed AlibabaQwen open-source big model. Based on the prompts, the Qwen model summarizes the patent's field of application, key problems solved, innovative methods, and title information within 500 characters (including punctuation) to generate a patent summary.

[0044] The patent information extraction module transmits the patent summary to the patent classification module, which then inputs the patent summary into the PatentClassifier model. The user has pre-specified the minimum probability of the label.

[0045] The PatentClassifier model analyzes the patent summary and outputs multiple classification labels and corresponding possibilities. The classification labels are all distinguished into CPC subcategories, such as G06F16 / 35 (representing clustering and classification in information processing), G06N3 / 04 (representing computer systems using neural network architecture), and G06Q50 / 18 (representing data processing systems or methods suitable for administrative, commercial, financial, management, supervision, or predictive purposes). The patent classification module feeds back the classification results to the user, completing the classification process.

[0046] Example 2 Users input patent documents in PDF format, which need to be categorized.

[0047] After receiving the PDF file, the patent information extraction module extracts the complete patent text information (unstructured) using PDF parsing technology.

[0048] After receiving unstructured patent text, the patent classification module calls the API service of Kimi Big Model to analyze the unstructured text and extract structured patent information in JSON format. The system then proceeds with the subsequent processing by default (unless the user selects manual confirmation).

[0049] The patent classification module examines the structured patent information and finds that it does not contain classification labels, so it calls the patent information extraction module.

[0050] The patent information extraction module inputs the optimized system prompts and structured patent information into the DeepSeek large model API to generate a patent summary of no more than 500 words (including punctuation). The summary covers the patent's field of application, key problems solved, innovative methods, and title information.

[0051] The patent classification module inputs a summary of patents into the PatentClassifier model, where the user pre-specifies the minimum probability of each label.

[0052] The PatentClassifier model analyzes the patent summary and outputs multiple classification labels and corresponding possibilities. The classification labels are all distinguished into CPC subcategories, such as H04L67 / 12 (representing "framework for distributed computing on a network") and G06F9 / 54 (representing inter-process communication in a program control device). The patent classification module feeds back the classification results to the user, completing the classification process.

[0053] Example 3 The user inputs structured patent information in JSON format, which includes classification labels: ["A01B1 / 02", "A01B2 / 00"]. The user chooses not to regenerate the classification results.

[0054] After receiving the structured information, the patent information extraction module checks its validity and completeness, verifying that fields such as "title" (patent title), "abstract" (patent abstract), "claim" (claims), and "description" (patent details) are complete and in valid format.

[0055] After receiving the structured information, the patent classification module detected that there were valid classification labels in the "labels" field, and the user explicitly chose not to regenerate the classification results. Therefore, it directly fed back the classification labels and their corresponding category meanings to the user: A01B1 / 02 (representing agriculture, forestry, animal husbandry, hunting, trapping, and fishing, specifically involving hand tools) and A01B2 / 00 (representing agriculture, forestry, animal husbandry, hunting, trapping, and fishing, specifically involving manual tools for farming). No further classification processing was performed, and the classification response was completed quickly.

[0056] The above description is merely illustrative of the embodiments of the present invention and is not intended to limit the present invention. For those skilled in the art, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent system for improving the quality of patent classification, characterized in that, It includes a patent information extraction module, a patent information synchronization and retrieval module, a patent information refinement module, and a patent classification module that work in sequence. The patent information synchronization and retrieval module is connected to a local patent database and a remote authoritative patent data source. The patent information refinement module and the patent classification module are both connected to a large language model (LLM). The patent information extraction module is used to receive at least one user input, including patent number, PDF / Word / TXT format patent file, and patent structured information, to obtain complete patent information; The patent information synchronization and retrieval module is used to obtain or periodically synchronize patent information from the authoritative patent data source to the local patent database, and provides a content-based vector search function. The patent information extraction module is used to receive structured patent information, input system prompts optimized to adapt to the CPC classification standard, and call the core patent points extracted by the large language model LLM. The core patent points include the patent's field of application, the key problems solved, the innovative methods / technologies / tools proposed, and title information. The patent classification module is used to call the complete patent information obtained by the patent information extraction module. If the patent information contains classification information and the user does not choose to regenerate the structured patent information, the classification result is returned directly. If the patent information is unstructured text, the patent structured information is extracted by the Large Language Model (LLM). The core points output by the patent information extraction module are then input into the PatentClassifier, a patent classification model based on BERT fine-tuning. Multiple classification labels and their corresponding probabilities are output according to the minimum probability of the labels specified by the user.

2. The intelligent system for improving patent classification quality according to claim 1, characterized in that, The system prompt words are specifically: You are a patent expert who is very familiar with the CPC's patent classification standards. Based on the input structured patent information, please summarize the information that helps to correctly classify the patent. This information should include the field involved in the patent, the key problem solved, the innovative methods, technologies or tools proposed, title information, etc., which will help to determine the patent category as comprehensively and accurately as possible. The patent category should be distinguished into the CPC subclass. The input patent structured information is in JSON format, as follows: { "title": <Patent title information> "abstract": <Patent abstract> "claim": <claims> "description": <Patent details> } The information in the summary must be no more than 500 characters long, including punctuation.

3. The intelligent system for improving patent classification quality according to claim 1, characterized in that, The large language model LLM is an open-source large model deployed locally or an API service provided by a large model service provider. The open-source large models include Meta's LLaMA and Alibaba's Qwen, and the API services include those provided by Kimi and DeepSeek.

4. The intelligent system for improving patent classification quality according to claim 1, characterized in that, The PatentClassifier is a patent classification model obtained by extracting key information from a large amount of labeled and classified patent data and then fine-tuning the original BERT model through multiple rounds of training.

5. The intelligent system for improving patent classification quality according to claim 1, characterized in that, The patent information extraction module obtains complete patent information in the following way: If the input is a patent number, the local patent database will be queried first. If the patent information is not found locally, the patent information synchronization and retrieval module will be called to obtain it from the remote authoritative patent data source and return the patent structured information in JSON format. If the input is a PDF / Word / TXT file, the complete patent text information will be extracted. If the input is structured patent information in JSON format, then a check for structural validity and completeness will be performed.

6. The intelligent system for improving patent classification quality according to claim 5, characterized in that, The fields of the JSON-formatted structured information include "title" (patent title), "abstract" (patent abstract), "claim" (claims), and "description" (patent details).

7. The intelligent system for improving patent classification quality according to claim 1, characterized in that, The synchronization cycle of the patent information synchronization and retrieval module can be customized, and the synchronization content includes newly added patent information and updates to patent classification information.

8. The intelligent system for improving patent classification quality according to claim 1, characterized in that, In the patent classification module, after the unstructured text is processed by the Large Language Model (LLM) to extract structured information, the user can confirm that the information is correct before proceeding with further processing. By default, the subsequent processing is performed directly.

9. The intelligent system for improving patent classification quality according to claim 1, characterized in that, The content-based vector search function converts patent text information into vector representations and uses vector similarity algorithms to achieve in-depth patent content retrieval.

10. The intelligent system for improving patent classification quality according to claim 1, characterized in that, The classification labels output by the patent classification module are distinguished into CPC subcategories, and each classification label corresponds to a unique category meaning description.

Images