Patent-technology standard association matching method and system based on technology-efficacy label system

Abstract

The invention relates to a patent-technology standard association matching method and system based on a technology-efficacy label system. The method comprises the following steps: respectively constructing technology-efficacy two-level label systems corresponding to patents and technology standards; the method comprises the following steps: acquiring a patent document and a technical standard document to be matched, and extracting tag terms corresponding to the patent document and the technical standard document based on a technology-effect two-level tag system through text preprocessing and core chapter extraction; respectively carrying out semantic coding on tag terms corresponding to the patent document and the technical standard document by utilizing the embedded model to obtain a patent technical word vector, a patent efficacy word vector, a standard technical word vector and a standard efficacy word vector; calculating the semantic similarity between the patent technology word vector and the standard technology word vector, and the semantic similarity between the patent efficacy word vector and the standard efficacy word vector; and screening patent-standard technology word pairs and patent-standard efficacy word pairs to realize patent-technology standard association matching.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a patent-technology standard association matching method and system based on a technology-efficacy labeling system. Background Technology

[0002] Currently, new energy equipment technologies encompassing photovoltaics, wind power, energy storage, hydrogen energy, and grid integration are the core support for ensuring energy security and promoting green development. Their industrialization highly depends on the coordinated transformation of technical standards and patents. Integrating promising patented technologies into standards can regulate applications and reduce entry costs, but accurately identifying the matching relationship between the two is the core bottleneck. Existing technologies have the following shortcomings: (1) The existing labels are not closely related to the transformation goals. Most of the existing labels focus on literature management or a single dimension, and are not designed around whether the patent can be transformed into a standard. They lack key judgment elements such as technical composition and functional goals, which leads to the mis-association of patents with no transformation value with standards and the underutilization of analysis resources.

[0003] (2) The tag system lacks operability and systematicity. Some tags rely on subjective judgment, have no clear extraction basis, have poor consistency, and lack a progressive structure of macro screening and micro matching. It is impossible to first exclude irrelevant objects through macro tags and then accurately locate matching points, resulting in low efficiency and difficulty in being applicable to massive data scenarios.

[0004] (3) Weak text preprocessing and core information extraction: Most patents and standards are in general formats, and digital versions are prone to format confusion during extraction. Scanned versions lack noise reduction and skew correction after optical character recognition, resulting in low accuracy. Furthermore, core chapters cannot be automatically identified, and text is mixed with noise such as headers and footers, interfering with the accuracy of tag extraction.

[0005] (4) The tag extraction algorithm is too simple and cannot meet the needs of multiple requirements. There are two types of tags: structured technology tags and descriptive technology and efficacy tags. Existing technologies or traditional keyword matching rely on manual labor and cannot extract both types of tags at the same time. It is difficult to form a complete technology-efficacy system and the data reliability supporting collaborative transformation is insufficient.

[0006] In summary, existing technologies cannot meet the needs of efficient and accurate collaborative transformation in the field of new energy equipment technology, and lack a tag extraction solution that adapts to document characteristics, achieves technology-efficiency linkage, and is highly accurate and efficient. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a patent-technical standard association matching method and system based on a technology-efficacy tagging system. This addresses the problems that arise when patents and technical standards lack a structured two-level technology-efficacy tagging system under a unified semantic framework, and when the tag extraction process suffers from insufficient adaptability of text preprocessing to the two types of documents, low efficiency in locating core chapters, weak semantic understanding of technology dimension tags, and strong subjectivity of efficacy dimension tags. These issues prevent accurate association matching between the two, making it difficult to effectively identify the technical problems of patents' potential for conversion into technical standards.

[0008] The objective of this invention can be achieved through the following technical solutions: One aspect of the present invention provides a patent-technology standard association matching method based on a technology-efficacy labeling system, comprising: Construct separate two-level labeling systems for patents and technical standards, focusing on both technology and efficacy. Obtain the patent documents and technical standard documents to be matched, and extract the corresponding tag terms based on the two-level tag system of technology and efficacy through text preprocessing and core chapter extraction. Using an embedding model, semantic encoding is performed on the tag terms corresponding to patent documents and technical standard documents to obtain patent technology word vectors, patent efficacy word vectors, standard technology word vectors, and standard efficacy word vectors. Calculate the first semantic similarity between the patent technology word vector and the standard technology word vector, and the second semantic similarity between the patent efficacy word vector and the standard efficacy word vector; Based on preset technical term similarity thresholds and efficacy term similarity thresholds, patent-standard technical term pairs with a first semantic similarity greater than the technical term similarity threshold, and patent-standard efficacy term pairs with a second semantic similarity greater than the efficacy term similarity threshold are selected. Patent-technology standard association matching is achieved through screening.

[0009] As a preferred technical solution, the technology-efficacy two-level labels both include technology dimension labels and efficacy dimension labels. The technology dimension labels include the technical field as the first-level label and the technical means as the second-level label. The efficacy dimension labels include the core functional objectives as the first-level label and the specific efficacy realization as the second-level label.

[0010] As a preferred technical solution, the labeling terms include technical field terms, technical means terms, core functional target terms, and specific efficacy achievement terms.

[0011] As a preferred technical solution, the process of constructing the patent-related two-level labeling system for technology and efficacy includes: The relevant technical text is extracted from the patent abstract, the preamble of the claims, and the background section to generate the technical field as a first-level tag; the description of the method, component, or process of the technical solution is extracted from the detailed implementation section to generate the technical means as a second-level tag. Qualitative functional descriptions are extracted from the patent abstract and the invention purpose section to generate core functional objectives as primary tags; quantitative or perceptible efficacy manifestations are extracted from the experimental data and effect examples of specific implementation methods to generate specific efficacy realizations as secondary tags. The technical means label belongs to the corresponding technical field label, and the specific effect realization label corresponds to the core function target label.

[0012] As a preferred technical solution, the construction process of the two-level labeling system for technology and efficacy corresponding to the aforementioned technical standard includes: The technical scope descriptions are extracted from the scope of technical standards and the normative reference documents to generate the technical field as the first-level label; the operation procedures, algorithms or protocol rules are extracted from the technical requirements to generate the technical means as the second-level label. Extract the core value description from the introduction and overview chapters of the technical standards to generate the core functional objectives as the first-level labels; extract the performance parameters or effect requirements from the performance indicators and testing methods chapters to generate the specific efficacy implementation as the second-level labels; The technical means label belongs to the corresponding technical field label, and the specific effect realization label corresponds to the core function target label.

[0013] As a preferred technical solution, the generation of technical field labels is based on a BERT pre-trained model, while the technical means labels, core functional target labels, and specific efficacy realization labels are implemented using a generative large model. The generation process of the technical field labels includes: Using IPC classification numbers as a standard system for labeling technical fields, a pairing sample set of core chapter text and IPC classification numbers was constructed. A fully connected layer is added to the output layer of the BERT model, and the Sigmoid activation function is used to output the label prediction probability. The binary cross-entropy is used as the loss function, and the model is trained based on the paired sample set. Input the text of the technical field to be processed into the fine-tuned model, and output the IPC classification number that exceeds the preset probability threshold as the technical field label based on the preset probability threshold.

[0014] As a preferred technical solution, after achieving patent-technology standard association and matching, it also includes: Record the association mapping between patents and technical standards, and retain the association results for the following two mapping scenarios: Mapping scenario 1: There is a corresponding match between the technology-efficacy label of the patent and the technology-efficacy label of the technical standard, and the match meets the threshold requirements; Mapping scenario 2: The patent's technology-efficacy label exists, and there is a match that meets the threshold requirements with the blank technology-efficacy label field in the technical standard.

[0015] As a preferred technical solution, the core chapters of patent documents and technical standard documents are extracted using preset regular expressions. For the patent document, the chapters on scope, technical requirements and performance indicators are extracted; for the patent document, the chapters on abstract, claims, specific implementation and background technology are extracted.

[0016] As a preferred technical solution, the text preprocessing includes the following steps: Determine the document PDF type. If it is a digital PDF, extract the text directly. If it is a scanned PDF, perform noise reduction, binarization and skew correction preprocessing, and perform OCR recognition through the OCR engine. Combine the OCR results with a special dictionary in the field of new energy equipment technology to verify the OCR results. For the extracted text, headers, footers, page numbers, garbled characters, and XML / HTML tags are removed. The text encoding is unified, and sentence segmentation and Chinese word segmentation are performed. Meaningless words are removed based on the stop word list, and word form restoration is performed.

[0017] Another aspect of the present invention provides a patent-technology standard association matching system based on a technology-efficacy labeling system, used to implement the aforementioned patent-technology standard association matching method, the system comprising: The technology-efficacy two-level labeling system construction module is used to construct the technology-efficacy two-level labeling system corresponding to patents and technical standards, respectively. The tag terminology extraction module is used to obtain the patent documents and technical standard documents to be matched. Through text preprocessing and core chapter extraction, it extracts the tag terms corresponding to the patent documents and technical standard documents based on the two-level tag system of technology and efficacy. The word vector extraction module is used to semantically encode the tag terms corresponding to patent documents and technical standard documents using an embedding model, respectively, to obtain patent technology word vectors, patent efficacy word vectors, standard technology word vectors, and standard efficacy word vectors. The similarity calculation module is used to calculate the first semantic similarity between the patent technology word vector and the standard technology word vector, and the second semantic similarity between the patent efficacy word vector and the standard efficacy word vector. The similarity filtering module is used to filter patent-standard technical word pairs with a first semantic similarity greater than the technical word similarity threshold and a second semantic similarity greater than the efficacy word similarity threshold, based on preset technical word similarity thresholds and efficacy word similarity thresholds. The association matching module is used to achieve patent-technology standard association matching based on filtering.

[0018] Compared with the prior art, the present invention has at least one of the following beneficial effects: (1) Achieving alignment of technical and efficacy information in the same dimension: In response to the problem that the core information description dimensions of patents and technical standards are inconsistent, making it impossible to compare and associate them in the same semantic framework, this invention constructs a two-dimensional, two-level hierarchical tag system of technology and efficacy. The technology dimension includes a first-level tag "Technology Field" that defines the macro-technical scope and a second-level tag "Technical Means" that describes the specific implementation method. The efficacy dimension includes a first-level tag "Core Functional Objective" that clarifies the core value and a second-level tag "Specific Efficacy Achievement" that quantifies / concretizes the effect. This system is constructed for patent and technical standard documents respectively to ensure that the tag dimensions of the two are consistent, so that the core information of technology and efficacy are aligned in the same dimension, eliminating the association obstacles caused by the difference in description.

[0019] (2) Achieve high accuracy in extracting technical field tags: In view of the problem that existing technical field tag extraction methods are difficult to capture the deep semantic influence on accuracy, this invention achieves accurate tag extraction in multiple dimensions. For technical field tags, a BERT pre-trained model is used, and for other tags, a generative large model is used. This can accurately identify patents and standards that cross multiple technical fields, and the extraction of efficacy-related tags is free from manual dependence. Attached Figure Description

[0020] Figure 1 This is a flowchart of the patent-technology standard association matching method based on the technology-efficacy label system in the embodiment; Figure 2 This is a schematic diagram of the prompt words in the embodiment; Figure 3 This is a schematic diagram of the patent-technology standard association matching system based on the technology-efficacy label system in the embodiment. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, 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 should fall within the scope of protection of the present invention.

[0022] Example 1 To address the problems existing in the aforementioned technologies, this embodiment provides a patent-technology standard association matching method based on a technology-efficacy label system. It focuses on the entire process of constructing the training dataset, from data source selection and attribute label system design to label extraction algorithm implementation and sample matching rule formulation, forming a reusable dataset construction methodology. This provides high-quality data support for the training of subsequent models such as conversion potential assessment models and conversion type recommendation models.

[0023] See Figure 1 The method includes the following steps: Step S1: Construct a two-level labeling system for technology and efficacy corresponding to patents and technical standards.

[0024] This step aims to build a tagging system that can accurately characterize the core "technical composition" and "functional objectives" of patents and technical standards, enabling them to be compared within the same semantic framework, thereby identifying matching points where patented technologies meet technical standard requirements and have transformation potential.

[0025] Tag selection should revolve around the inherent logic of "technology-efficacy," while also considering the commonalities between patents and technical standards. This ensures that the tags accurately depict the core characteristics of both and provide actionable analytical dimensions for correlation matching, specifically including: (1) Starting from the “technical essence”, anchoring the core feature dimension The core components of technology are "means" and "ends," so the label must first cover these two dimensions.

[0026] The technical dimension focuses on the "composition of technology itself," namely the "technical field" (the macro-level scope to which the technology belongs, addressing "what field the technology belongs to") and the "technical means" (the specific methods, tools, or paths to implement the technology, addressing "how to implement it"). Efficacy dimension: Focus on "the goals and effects of technology", namely "core functional goals" (the core value that the technology wants to achieve, solving "what is the purpose") and "specific efficacy realization" (the specific manifestations or indicators of the implementation of functions, solving "how effective is it".

[0027] This approach ensures that the labels do not deviate from the essential attributes of the technology, providing underlying logical support for association matching.

[0028] (2) Starting from "matching efficiency", construct a "layered and progressive" tag system. First-level tags (technical fields, core functional objectives) are used for "quick filtering," addressing the question of "whether they belong to the same category"; second-level tags (technical means, specific effects achieved) are used for "precise matching," addressing the question of "whether the technical details match." This hierarchical approach, starting from the macro level and moving to the micro level, reduces invalid matching (such as the correlation between cross-domain technologies and standards), improves analysis efficiency, and aligns with the cognitive pattern of moving from coarse to fine.

[0029] The core objective of the labeling system constructed in this step is to accurately depict the essence and value of technical solutions and establish a matching bridge between patents and technical standards, thereby identifying patents with the potential to be transformed into standards. This system adopts a two-dimensional (technical dimension and efficacy dimension) and two-level (first-level macro, second-level specific) structural design.

[0030] From a technical perspective, the focus is on technical terms that highlight the technical characteristics of the implementation solution. The primary tag is the technical field, used to macroscopically define the major category or branch to which the technology belongs (e.g., "power electronic equipment control," "energy storage power station frequency regulation control," "power system operation situation awareness technology"). The secondary tag is the technical means, describing the specific technologies, methods, components, or protocols used to achieve the function within a specific technical field (e.g., "dynamic simulation modeling method," "layered energy coordination control strategy"). The technical means tag emphasizes its criticality, distinctiveness, and specificity, aiming to accurately capture the core technical contribution of the patent.

[0031] In terms of effectiveness, focus on the efficacy terms that the technical solution can achieve in application. The primary label is the core functional objective, which describes the fundamental problem the technical solution aims to solve or the core value it achieves (e.g., "increase transmission rate," "enhance security," "reduce power consumption," "improve reliability"). Closely align with the system objectives, user needs, or specification requirements clearly stated in the technical standard document to ensure that the label reflects the fundamental value orientation pursued by the standard. The secondary label is the specific effectiveness achieved, which refines and quantifies (or concretizes) the core functional objective, describing the specific, measurable (or perceptible) performance improvement or effect manifested in achieving the core objective, emphasizing its measurability and direct relevance to the core objective.

[0032] Through the above labeling system design, the labels must directly point to the core judgment element of "conversion possibility" and be strongly related to the core goal of "whether the patent can be converted into a technical standard" to avoid interference from irrelevant features.

[0033] Step S2: Obtain the patent documents and technical standard documents to be matched. Through text preprocessing and core chapter extraction, extract the corresponding tag terms for the patent documents and technical standard documents based on the two-level tag system of technology and efficacy.

[0034] In this embodiment, the acquisition of technical standard documents is based on the "State Grid Corporation of China 2024 Technical Standard System Table." By establishing a mechanism for focusing on key areas, verifying validity, and mapping substitution relationships, existing valid standards in key technical fields such as new energy power generation equipment and grid connection control devices are precisely selected. In the text acquisition stage, a dual-channel strategy is adopted, prioritizing publicly available channels and supplementing with restricted texts. Legally valid texts are obtained first from official channels such as the National Standards Information Public Service Platform and local and industry regulatory authority platforms. Standards subject to copyright or access restrictions are supplemented through document sharing platforms, with strict content verification and risk assessment.

[0035] For obtaining technical standard documents, this embodiment uses a commercial global patent database, focusing on the field of new energy equipment technology in the power industry. Through field refinement and keyword iterative optimization, a composite search query is constructed for batch data download. The search query design adopts a multi-field combination and multi-level logical nesting strategy to ensure coverage of core patents in the target technology field.

[0036] (1) Data preprocessing Data preprocessing is a fundamental step in achieving accurate tag extraction, and its quality directly determines the accuracy of the final keyword extraction. Especially when the source file is PDF, different processing methods are required depending on its type. It can be viewed as a processing pipeline integrating data cleaning and standardization. The core objective of this process is clear: to extract clean, well-organized, and structured plain text data from unstructured, disorganized raw text data through a series of ordered processing steps, thus laying a solid foundation for subsequent tag extraction.

[0037] The preprocessing in this embodiment includes the following steps: 1) PDF Text Extraction First, determine the file type. For PDF files, there are mainly two categories: digital PDFs (text-based PDFs) and scanned PDFs (image-based PDFs). Digital PDFs embed a text layer that can be directly selected and copied. They are usually generated by saving directly from office software such as Word or through the print function. Text extraction from these PDFs is relatively straightforward. On the other hand, scanned PDFs are essentially a collection of pages of images without an embedded text layer. They are generally generated by scanning paper documents. Text extraction from these PDFs requires optical character recognition (OCR) technology, making them far more difficult to process than digital PDFs.

[0038] In selecting text extraction methods, a differentiated strategy is adopted for different types of files. For digital PDFs, text can be extracted directly using a PDF parsing library. This method is not only highly efficient but also maximizes the integrity of the text information, avoiding information loss.

[0039] For scanned PDFs, OCR technology is essential to convert the text in the image into editable and processable text. This is the most complex step in the entire preprocessing workflow, mainly comprising three steps: image preprocessing, OCR recognition, and post-processing and proofreading. In the image preprocessing stage, the quality of the operation directly affects the accuracy of subsequent OCR recognition. This mainly includes noise reduction (removing blemishes, noise, and other interfering elements from the image surface), binarization (converting color or grayscale images to black and white images to enhance the contrast between text and background), skew correction (correcting tilted pages to ensure horizontal text alignment), and resolution improvement (increasing the image resolution to over 300 DPI to provide a clear recognition foundation for the OCR engine). In the OCR recognition stage, a suitable OCR engine must be selected based on actual needs. After OCR recognition is complete, post-processing and proofreading are necessary. Since OCR output is not absolutely accurate, especially when dealing with blurry text or special fonts, recognition errors are prone to occur. Therefore, spell checking tools or comparison with domain-specific dictionaries are required to correct any recognition errors and ensure text accuracy.

[0040] 2) Text cleaning and standardization After text extraction, the original text data has been obtained. However, there is still a lot of useless information and irregularities in these original texts. Therefore, it is necessary to further optimize the text quality through text cleaning and standardization to make it meet the requirements of subsequent analysis.

[0041] In removing useless characters and noise, the primary task is to remove non-text elements, including headers, footers, page numbers, and improperly filtered image titles. These elements have no practical significance for text analysis and will only interfere with the results. Secondly, special characters need to be cleaned up, requiring the removal or replacement of garbled text and XML / HTML tags (such as...). Special characters such as LaTeX commands, as well as excessive line breaks and spaces, should be avoided to prevent them from affecting the regularity of the text. At the same time, encoding issues need to be addressed by uniformly converting the text to the UTF-8 encoding format to solve the problem of garbled text caused by inconsistent encoding.

[0042] Text normalization is another important aspect of this stage, mainly covering several key processing steps: 1. Sentence segmentation, which divides continuous text content into a list of independent sentences, facilitating subsequent processing at the sentence level.

[0043] 2. Word tokenization, which further divides a single sentence into a list of words or phrases. For Chinese text, tokenization is particularly crucial.

[0044] 3. Stop word removal step, aiming to remove words that frequently appear in the text but have extremely low information content. For example, words like "的", "是", "在" in Chinese contribute less to the semantic expression of the text. Removing them can effectively reduce the text data volume and improve the efficiency of subsequent processing.

[0045] 4. Word form restoration and stemming. Word form restoration is to restore a word to its original form in the dictionary, which is more accurate but relatively time-consuming. Stemming, on the other hand, obtains the stem by roughly cutting off the word affixes, with faster processing speed but possible inaccuracies.

[0046] (2) Core section extraction In patent documents, the technical field is mainly extracted from the International Patent Classification (IPC), the patent title, the abstract, and the field descriptions in the claims, supplemented by the definition of the technical application scenario in the background art section of the specification. The core source of technical means is the specific implementation manner in the specification and the technical solutions in the claims, focusing on the specific methods to solve technical problems. The core functional objective labels are concentrated in the patent abstract and the invention objective section, extracting qualitative descriptions such as "improving data transmission reliability"; while the specific realization of efficacy needs to be mined from the experimental data, effect embodiments, and conclusions in the specification to reflect the technical effects.

[0047] For technical standard documents, the technical field needs to be identified from the relevant sections of the standard technical scope and the list of normative reference documents. Technical means are derived from the technical specification descriptions in the core clauses of the standard, and the operation processes, algorithms, or protocol rules specified in the standard need to be analyzed. The core functional objectives usually exist in the introduction or overview section of the standard; the specific realization of efficacy needs to be extracted from the performance index section and the test methods of the standard.

[0048] This step uses regular expressions to identify the core sections through chapter titles.

[0049] (1) Pattern analysis First, analyze the document structure to determine the keywords for the core chapter titles. For example, patent documents include abstracts, claims, background technology, detailed embodiments, experimental data, etc. Technical standard documents include scope, normative references, technical requirements, performance indicators, test methods, etc.

[0050] (2) Regular expression design Design flexible regular expression patterns based on common title formats (such as potential numbering, chapter prefixes, and variations in spaces). The patterns should ignore irrelevant formatting differences and focus on matching keywords.

[0051] Example of regular expression patterns in the core sections of technical standard documents: Example of regular expressions for core sections of a patent document: Regular expressions for core chapters of patent documents: Chapter titles in patent documents are usually more fixed and less often numbered, but there may be some variations. The following pattern directly matches keywords, allowing leading and trailing spaces and common prefixes. (3) Matching and Extraction Apply regular expressions to the document text to search for matching chapter titles. Once a title is found, extract the content of that chapter (usually from the beginning of the title to the next title or the end of the document). Clean up the extracted content, such as removing headers, footers, and extra spaces, and verify the accuracy of the extraction. If necessary, adjust the regular expression pattern according to the context to improve accuracy.

[0052] Step S3: Use the embedding model to perform semantic encoding on the tag terms corresponding to the patent document and the technical standard document respectively to obtain patent technology word vectors, patent efficacy word vectors, standard technology word vectors and standard efficacy word vectors.

[0053] The entire tag extraction process is divided into two stages, forming a phased and progressive pipeline. Each stage is closely connected and performs its own function. It relies on the BERT algorithm and large model to accurately extract technology-efficacy tags. Finally, it integrates to form a structured output of a two-level technology-efficacy tag system, fully presenting the core technology and value positioning of patents and standards.

[0054] (1) Extraction from the technical field The first stage involves extracting the technical field information, primarily utilizing the BERT model for zero-shot or few-shot classification and clustering. The essence of this task is to assign one or more predefined, structured labels (such as IPC classification numbers) to the input text blocks. The BERT series models, with their powerful semantic understanding capabilities, can accurately capture core information related to the technical field within the text. Furthermore, their high-efficiency processing characteristics meet the needs of batch text processing, making them the ideal technology choice for this stage, effectively achieving rapid localization and classification of the technical field within the text.

[0055] 1) Data preparation and tagging system construction Source text screening: Not all content in technical documents (patents, standards) directly reflects the core technology field. It is necessary to screen text fragments with high information density and clear technical focus to avoid redundant information interfering with model judgment. For patent documents, priority should be given to extracting the abstract (condensing the core technical solution and application scenarios), the preamble of the claims (clarifying the technical field to which the patent protection subject belongs), and the background technology (explaining the defects of the prior art and the technical positioning of this patent). For standard documents, the focus should be on extracting the "Scope" section (directly defining the technical field and objects to which the standard applies) and the "Normative References" section (indirectly assisting in determining the technical scope of the current standard through the reference of other technical standards).

[0056] Labeling System Establishment: To ensure the universality and authority of the labels, the internationally recognized technical classification system is directly adopted as the labeling system, avoiding the "classification confusion" problem caused by custom labels. The core advantages of choosing IPC / ICS as the labeling system are: first, it covers the vast majority of technical fields globally, and the classification granularity can be adjusted as needed (e.g., from a coarse classification at the "department" level to a fine classification at the "subgroup" level); second, it has clear official interpretations, and the correspondence between labels and technical fields can be obtained through the official websites of the World Intellectual Property Organization (WIPO) and the International Organization for Standardization (ISO), providing authoritative basis for data labeling.

[0057] To create text-label paired samples: Pair the filtered source text with the corresponding IPC / ICS labels one-to-one (or more-than-one) to build a structured dataset that the model can directly learn from. Two key points to note in this process are label mapping accuracy and sample diversity coverage.

[0058] 2) Model selection and training This step is the core of technical label extraction. It relies on the semantic understanding capabilities of the BERT pre-trained model and uses "fine-tuning" to adapt the model to the classification task of "technical text - IPC / ICS labels", ultimately enabling it to predict labels from new text.

[0059] The pre-training corpus of the BERT model directly affects its performance in specific tasks. Since the task deals with Chinese technical texts (patents and standards are mostly written in Chinese), it is necessary to choose a model pre-trained on Chinese scientific and technological corpora to reduce the "understanding cost" of the model for technical terms. This study adopts the BERT-Base-Chinese basic Chinese BERT model, whose pre-training corpus includes general Chinese texts and some scientific and technological texts, which is suitable for scenarios with small amounts of data and low requirements for model complexity.

[0060] The pre-trained BERT model only possesses general semantic understanding capabilities. To adapt it for multi-label classification tasks, it needs to be trained using labeled data and added as a task-specific layer. The specific fine-tuning process is as follows: First, add a classification layer (ClassificationHead). After the output layer of the BERT model (usually represented by a vector of [CLS] tokens, which integrates the semantic information of the entire text), add a fully connected layer (DenseLayer) and output the predicted probability of each label using the Sigmoid activation function (the Sigmoid function maps probabilities to the 0-1 range, adapting to the logic of "independent judgment for each label" in multi-label classification). Then, determine the loss function. Since the IPC / ICS label system is hierarchical and supports multiple labels (a technical text may involve multiple fields, such as a patent involving both "communication technology" and "computer technology," corresponding to two IPC labels), a loss function suitable for multi-label classification—the binary cross-entropy loss function—needs to be selected. This function can independently calculate the loss for each label (determining whether "the text belongs to this label"), avoiding the limitation of "only predicting one label" in multi-class classification. Finally, training parameter settings: To avoid model overfitting (only remembering training data and unable to generalize to new data), training parameters need to be set reasonably. For example, use a "small learning rate" (such as 2e-5~5e-5) to avoid excessive learning rate causing model parameter oscillations; set according to hardware computing power, usually 16 or 32; the number of training rounds is generally set to 3~5 rounds, and the model performance is monitored through a validation set (dividing 20%~30% of the total data as the validation set). If the validation set accuracy no longer improves, training is stopped early (EarlyStopping); use regularization strategies, add a Dropout layer (dropout rate set to 0.1~0.3), randomly "shut down" some neurons to reduce the model's excessive dependence on training data.

[0061] 3) Extraction from the technical field First, following the text filtering rules of the "data preparation stage", extract core text fragments such as the abstract, the preamble of the claims (patents) or the scope section (standards), and then input them into the trained model.

[0062] Text preprocessing: Perform preprocessing on the input text fragments consistent with the training data (such as removing special symbols, standardizing Chinese and English punctuation, and truncating or completing the text to the length required by the model) to ensure that the input format meets the model's expectations; Probability output: The model performs semantic encoding on the preprocessed text and outputs the predicted probability (range 0~1) of each IPC / ICS label through the classification layer. The higher the probability, the greater the likelihood that the text belongs to that label. Threshold filtering: Set a probability threshold (usually 0.5, which can be adjusted according to actual needs: if you need to reduce false negatives, you can lower the threshold to 0.3~0.4; if you need to reduce false positives, you can raise the threshold to 0.6~0.7), and use all tags whose predicted probabilities exceed the threshold as "valid technical field tags" for the text.

[0063] (2) Extraction of other tags The second stage of the process involves extracting tags for technical means, core functional objectives, and specific effects. Compared to the structured tag allocation in the first stage, this stage requires a deeper exploration of the technical details within the text. Therefore, a Generative Large Model (LLM) combined with a text generation method based on instructions is employed. This task requires the model not only to understand the text content but also to summarize and categorize the technical elements within it, generating descriptive tags in a non-fixed format. The creativity and powerful language generation capabilities of the Generative Large Model perfectly match this requirement, thereby enabling in-depth mining of the technical elements within the text.

[0064] 1) Source text preparation The technical means, core functional objectives, and specific effects are scattered across different chapters of patent and standard documents. To improve accuracy, it is necessary to filter out text that directly describes this information and avoid irrelevant information interfering with the model's judgment. Refer to the previous section for specific corresponding chapters.

[0065] 2) Large Model Selection Based on data sensitivity, deployment requirements, and budget, choose an adaptation solution from two categories: "Commercial APIs" and "Open Source Models." The core differences and applicable scenarios of these two types of models differ. Commercial APIs include OpenAIGPT-4 / GPT-3.5-Turbo, AnthropicClaude, Wenxin Yiyan, and Tongyi Qianwen. These models are powerful and easy to use, accessed via API calls. Open Source Models include ChatGLM3-6B, Qwen-7B-Chat, and Baichuan2-13B-Chat. These models can be deployed locally or on private servers, ensuring data security.

[0066] 3) Prompt word engineering Prompts are the "core instructions" for large models to extract labels. Their role, task objectives, output format, and example references must be clearly defined. Few-shot learning reduces model comprehension bias. Prompts need to meet four requirements: 1. Role Definition: Assign the model the role of a "label extraction expert," clearly defining its task boundaries and professional attributes; 2. Task Description: Clearly state the "input text type" and the "label definition to be extracted" (e.g., "efficacy labels refer to the specific functions, effects, or problems solved by the technology, and must have quantifiable or clearly describable features"); 3. Output Format: Specify that the model outputs in JSON format, including fields such as "patent ID," "technical keywords," "core efficacy labels," "specific efficacy descriptions," and "data support (if any)," ensuring direct parsing later; 4. Few-Shot Examples: Provide 1-2 paired "input text-output results" examples to help the model understand the granularity and expression specifications of the labels. Example of a prompt: Figure 2 As shown.

[0067] 4) Calling and Post-processing The constructed Prompt is sent to the large model via API or local calls. After obtaining the JSON string returned by the model, it is parsed using Python's json.loads() function to handle possible formatting anomalies (such as occasional annotation text added by the model, which needs to be cleaned up using regular expressions first). Finally, the parsed tag data is integrated with technology tags to form a two-level tag matrix of technology-efficacy for patents and standards.

[0068] Step S4: Calculate the first semantic similarity between the patent technology word vector and the standard technology word vector, and the second semantic similarity between the patent efficacy word vector and the standard efficacy word vector.

[0069] Step S5: Filter patent-standard technical term pairs whose first semantic similarity is greater than the technical term similarity threshold, and patent-standard efficacy term pairs whose second semantic similarity is greater than the efficacy term similarity threshold.

[0070] After constructing the multi-level technical efficacy matrices for patents and standards, semantic similarity correlation calculations are performed on the corresponding technical efficacy matrices of both. The correlation between patents and standards represents the rationality of the transformation between them. An embedding model is employed, such as the Zhiyuan General Embedding Model (BGE), to represent technical terms and efficacy terms as vectors. For technical terms, the BGE model analyzes their specific orientation, technical characteristics, and associations with other technical terms in the relevant technical field, thereby generating a vector that accurately represents the core information of the technical term. For example, for the technical term "artificial intelligence algorithm," the model comprehensively considers factors such as the technical scope of "artificial intelligence," the mathematical principles of "algorithm," and application scenarios to generate a corresponding vector. For efficacy terms, the model focuses on the semantic information of the function, effect, and role expressed by the word. For example, for the efficacy term "improve efficiency," the model captures the positive changes reflected by "improve" and the performance indicators involved in "efficiency," transforming them into corresponding vector representations. In this way, the semantic information of technical terms and efficacy terms is quantified into computable vectors, laying the foundation for subsequent similarity calculations.

[0071] Cosine similarity is used to characterize the semantic similarity between coordinate points. Cosine similarity is an index that measures the similarity of two vectors in direction, and its value ranges from [-1, 1]. When the two vectors are in the same direction, the cosine similarity value is 1; when the two vectors are in opposite directions, the cosine similarity value is -1; and when the two vectors are perpendicular, the cosine similarity value is 0. In this embodiment, since the vectors of technical terms and efficacy terms are generated based on semantic information, the closer the cosine similarity is to 1, the more similar the two terms are semantically. In this embodiment, two similarity thresholds are set, namely the technical term similarity threshold. Similarity threshold between efficacy terms This is used to filter out technical term pairs and efficacy term pairs that exceed a threshold in the calculation results. The vertical axis of the patent's technical efficacy matrix is... , ,…, ( (representing the number of points on the ordinate), and the x-coordinate is... , ,…, ( (This represents the number of points on the horizontal axis). The vertical axis of a standard technical effectiveness matrix is... , ,…, ( (representing the number of points on the ordinate), and the x-coordinate is... , ,…, ( (Indicates the number of points on the x-axis).

[0072] Each coordinate point of the patent's technical efficacy matrix is ​​compared one by one with the coordinates of the standard technical efficacy matrix to calculate the semantic similarity value, and coordinates exceeding a threshold are filtered out. Specifically, this involves dividing the vertical axis vector of the patent's technical efficacy matrix... one by one Calculate semantic similarity and return the associated vertical axis data; simultaneously, calculate the horizontal axis vector for each value. and Semantic similarity, providing feedback on horizontal and vertical data related to technical terms. semantic similarity value Efficacy terms semantic similarity value The calculation methods are shown below. In the formula: It usually represents a set of technical terms; This represents a set of efficacy terms; A set of efficacy terms representing the standard; A set of technical terms representing a standard; ,in This represents the number of patent terminology in a group after clustering; ,in This represents the number of standard technical terms in a cluster. The method for calculating the semantic similarity mapping of technical terms is shown below.

[0073] The calculation principle of efficacy terms is the same as that of technical terms, and will not be repeated here. After obtaining the relevant coordinate representations based on the above calculations, the coordinate points that are related to each other are deduced, and the relationship between the two points is determined by equation (7).

[0074] Coordinate point association The coordinates are not related.

[0075] In the formula, C is the similarity threshold. The calculated mapping relationship mainly falls into the following four categories: (1) Both patents and standards at both ends of the mapping have layouts; (2) Both patents at both ends of the mapping have layouts, but standards do not; (3) Both patents at both ends of the mapping have no layouts, but standards do; (4) Neither patents nor standards at both ends of the mapping have layouts. This embodiment mainly focuses on the analysis of existing patents and the construction of a dataset, so it only focuses on the first two categories.

[0076] Step S6: Based on the screening, achieve patent-technology standard association matching.

[0077] By matching patents with technical standards, the rationality of patent conversion into standards can be accurately identified based on the relationship between patents and technical standards. The dataset built based on these relationships can be used to train models such as conversion potential prediction models, conversion type recommendation models, and technology direction matching models, thus providing structured, high-quality training data for patent conversion research in the field of new energy equipment technology.

[0078] When a patent has related technical standards, it indicates that the patent is reasonable to transfer to a standard; conversely, if a patent does not have related technical standards, it indicates that the patent is less reasonable to transfer to a standard.

[0079] For example, this embodiment uses a dataset based on correlations, with 10,853 patent records in the field of new energy equipment technology as its core foundation, while also integrating relevant technical standard data in this field. This data covers the multi-dimensional characteristics of patents and the basic attributes of standards, laying a data foundation for the analysis and application of correlations.

[0080] (1) Ideas for constructing dataset features The features of the dataset are designed around the core objective of "accurately identifying the rationality of patent conversion into standards". The aim is to comprehensively and multidimensionally capture key information that may affect the association and conversion between patents and standards, thereby providing a solid data foundation for subsequent research.

[0081] Regarding patent characteristics, considering that the technical, legal, and market attributes of a patent all influence its relevance to standards and its potential for commercialization to varying degrees, multiple dimensions were selected. Basic identifying characteristics are the foundation for data traceability and differentiation, ensuring the uniqueness and identifiability of each patent. Technical characteristics reflect the patent's technical content, R&D background, and technological influence. These factors are closely related to whether a patent meets the technical requirements of a standard; the more mature the technology and the greater the influence of the patent, the more likely it is to be associated with a standard and possess commercialization potential. Legal characteristics reflect the scope and strength of patent protection. Sound legal protection provides a guarantee for the transformation of a patent into a standard and is also an important aspect of assessing its commercialization value. Market characteristics reflect the patent's market layout and influence; patents with high market recognition and wide market penetration are more easily accepted and promoted when transformed into standards.

[0082] The selection of standard features mainly considers their basic attributes. The standard number is used to uniquely identify the standard, ensuring that it can accurately point to a specific standard in data processing and analysis. The standard level reflects the authority and scope of application of the standard. Standards at different levels may have different requirements for the association and transformation of patents. Including this feature helps to analyze the relationship between patents and standards more accurately.

[0083] The correlation feature is the core feature that directly reflects whether there is a correlation between a patent and a standard. It is a key basis for judging the rationality of patent commercialization. Including it in the dataset can intuitively reflect the commercialization potential of a patent.

[0084] (2) Patent features 1) Basic Identification Features It includes the patent title and publication number, serving as a unique identifier for the patent and ensuring data traceability.

[0085] 2) Technical Dimension Features Number of inventors: reflects the scale of team collaboration in patent research and development, and indirectly reflects the complexity of technology research and development; The number of cited patents reflects the relevance of the patents to existing technologies; a higher number indicates a stronger technological foundation. IPC classification number: reflects the breadth of technical fields involved in the patent; a higher number of classifications means a wider range of technologies covered. Number of non-patent cited references: including journals, papers, etc., reflecting the degree of integration between patents and academic research; Number of cited patents: This reflects the technological influence of the patent; the more times a patent is cited, the higher its technological recognition.

[0086] 3) Legal Dimension Features Number of claims: This reflects the breadth of patent protection; the more claims, the more comprehensive the protection. Number of independent claims: This reflects the strength of protection for the core technology of the patent; the more claims there are, the more prominent the core technology.

[0087] 4) Market Dimension Characteristics Number of simple family patent applications: This reflects the geographical protection layout of the patents; the more applications there are, the more obvious the intention to expand into the market. Total number of cited patents in a simple family: reflects the overall market influence of the patent family; Number of simple patent-related countries / regions: This reflects the international market coverage of the patent; the more countries / regions covered, the wider the global reach.

[0088] (3) Standard features It mainly includes basic attributes such as standard number and hierarchy: Standard Name: The name should clearly reflect the core content and scope of the standard, facilitating a quick understanding of the standard's subject matter. Standard Number: Serves as a unique identifier for a standard, ensuring accurate location of standard information; Standard level: Reflects the scope and level of application of a standard, such as national standards, industry standards, local standards, etc., and embodies the authority and influence of the standard.

[0089] (4) Characteristics of Relationship Each patent data entry includes an association status identifier with the standard (association exists / no association exists), directly reflecting the rationality of the patent's standard conversion.

[0090] Example 2 Based on Example 1, this example provides a patent-technology standard association matching based on a technology-efficacy labeling system to implement the patent-technology standard association matching method of Example 1. See [link to example]. Figure 3 The system includes: (1) Technology-Efficacy Two-Level Label System Construction Module, used to construct the technology-efficacy two-level label system corresponding to patents and technical standards respectively.

[0091] (2) Tag term extraction module, used to obtain the patent documents and technical standard documents to be matched. Through text preprocessing and core chapter extraction, the tag terms corresponding to the patent documents and technical standard documents are extracted based on the two-level tag system of technology and efficacy.

[0092] (3) Word vector extraction module, which uses the embedding model to perform semantic encoding on the tag terms corresponding to the patent document and the technical standard document respectively, to obtain patent technology word vector, patent efficacy word vector, standard technology word vector and standard efficacy word vector.

[0093] (4) Similarity calculation module, used to calculate the first semantic similarity between the patent technology word vector and the standard technology word vector, and the second semantic similarity between the patent efficacy word vector and the standard efficacy word vector.

[0094] (5) Similarity filtering module, used to filter patent-standard technical word pairs with a first semantic similarity greater than the technical word similarity threshold and patent-standard efficacy word pairs with a second semantic similarity greater than the efficacy word similarity threshold based on a preset technical word similarity threshold and efficacy word similarity threshold.

[0095] (6) Association matching module, used to achieve patent-technical standard association matching based on screening.

[0096] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A patent-technology standard association matching method based on a technology-efficacy labeling system, characterized in that, include: Construct separate two-level labeling systems for patents and technical standards, focusing on both technology and efficacy. Obtain the patent documents and technical standard documents to be matched, and extract the corresponding tag terms based on the two-level tag system of technology and efficacy through text preprocessing and core chapter extraction. Using an embedding model, semantic encoding is performed on the tag terms corresponding to patent documents and technical standard documents to obtain patent technology word vectors, patent efficacy word vectors, standard technology word vectors, and standard efficacy word vectors. Calculate the first semantic similarity between the patent technology word vector and the standard technology word vector, and the second semantic similarity between the patent efficacy word vector and the standard efficacy word vector; Based on preset technical term similarity thresholds and efficacy term similarity thresholds, patent-standard technical term pairs with a first semantic similarity greater than the technical term similarity threshold, and patent-standard efficacy term pairs with a second semantic similarity greater than the efficacy term similarity threshold are selected. Patent-technology standard association matching is achieved through screening.

2. The patent-technology standard association matching method based on a technology-efficacy labeling system according to claim 1, characterized in that, The technology-efficacy two-level tags both include technology dimension tags and efficacy dimension tags. The technology dimension tags include the technical field as the first-level tag and the technical means as the second-level tag. The efficacy dimension tags include the core functional objectives as the first-level tag and the specific efficacy realization as the second-level tag.

3. The patent-technology standard association matching method based on a technology-efficacy labeling system according to claim 2, characterized in that, The terminology used in the labeling includes technical field terms, technical means terms, core functional objectives terms, and terms for achieving specific effects.

4. The patent-technology standard association matching method based on a technology-efficacy labeling system according to claim 3, characterized in that, The process of constructing a two-tiered labeling system for technology and efficacy corresponding to a patent includes: The relevant technical text is extracted from the patent abstract, the preamble of the claims, and the background section to generate the technical field as a first-level tag; the description of the method, component, or process of the technical solution is extracted from the detailed implementation section to generate the technical means as a second-level tag. Qualitative functional descriptions are extracted from the patent abstract and the invention purpose section to generate core functional objectives as primary tags; quantitative or perceptible efficacy manifestations are extracted from the experimental data and effect examples of the specific implementation methods to generate specific efficacy realizations as secondary tags. The technical means label belongs to the corresponding technical field label, and the specific effect realization label corresponds to the core function target label.

5. The patent-technology standard association matching method based on a technology-efficacy labeling system according to claim 3, characterized in that, The construction process of the two-level labeling system for technology and efficacy corresponding to the aforementioned technical standards includes: The technical scope descriptions are extracted from the scope of technical standards and the normative reference documents to generate the technical field as the first-level label; the operation procedures, algorithms or protocol rules are extracted from the technical requirements to generate the technical means as the second-level label. Extract the core value description from the introduction and overview chapters of the technical standards to generate the core functional objectives as the first-level labels; extract the performance parameters or effect requirements from the performance indicators and testing methods chapters to generate the specific efficacy implementation as the second-level labels; The technical means label belongs to the corresponding technical field label, and the specific effect realization label corresponds to the core function target label.

6. The patent-technology standard association matching method based on a technology-efficacy labeling system according to claim 2, characterized in that, The generation of technical field tags is based on a BERT pre-trained model. Technical means tags, core functional target tags, and specific efficacy achievement tags are implemented using a generative large model. The generation process of the technical field tags includes: Using IPC classification numbers as a standard system for labeling technical fields, a pairing sample set of core chapter text and IPC classification numbers was constructed. A fully connected layer is added to the output layer of the BERT model, and the Sigmoid activation function is used to output the label prediction probability. The binary cross-entropy is used as the loss function, and the model is trained based on the paired sample set. Input the text of the technical field to be processed into the fine-tuned model, and output the IPC classification number that exceeds the preset probability threshold as the technical field label based on the preset probability threshold.

7. The patent-technology standard association matching method based on a technology-efficacy labeling system according to claim 1, characterized in that, After achieving patent-technology standard association matching, it also includes: Record the association mapping between patents and technical standards, and retain the association results for the following two mapping scenarios: Mapping scenario 1: There is a corresponding match between the technology-efficacy label of the patent and the technology-efficacy label of the technical standard, and the match meets the threshold requirements; Mapping scenario 2: The patent's technology-efficacy label exists, and there is a match that meets the threshold requirements with the blank technology-efficacy label field in the technical standard.

8. The patent-technology standard association matching method based on a technology-efficacy labeling system according to claim 1, characterized in that, The core chapters of patent documents and technical standard documents are extracted using preset regular expressions. For the patent document, the chapters on scope, technical requirements and performance indicators are extracted. For the patent document, the chapters on abstract, claims, detailed implementation and background technology are extracted.

9. The patent-technology standard association matching method based on a technology-efficacy labeling system according to claim 1, characterized in that, The text preprocessing includes the following steps: Determine the document PDF type. If it is a digital PDF, extract the text directly. If it is a scanned PDF, perform noise reduction, binarization and skew correction preprocessing, and perform OCR recognition through the OCR engine. Combine the OCR results with a special dictionary in the field of new energy equipment technology to verify the OCR results. For the extracted text, headers, footers, page numbers, garbled characters, and XML / HTML tags are removed. The text encoding is unified, and sentence segmentation and Chinese word segmentation are performed. Meaningless words are removed based on the stop word list, and word form restoration is performed.

10. A patent-technology standard association matching system based on a technology-efficacy labeling system, characterized in that, For implementing the patent-technology standard association matching method as described in any one of claims 1-9, the system includes: The technology-efficacy two-level labeling system construction module is used to construct the technology-efficacy two-level labeling system corresponding to patents and technical standards, respectively. The tag terminology extraction module is used to obtain the patent documents and technical standard documents to be matched. Through text preprocessing and core chapter extraction, it extracts the tag terms corresponding to the patent documents and technical standard documents based on the two-level tag system of technology and efficacy. The word vector extraction module is used to semantically encode the tag terms corresponding to patent documents and technical standard documents using an embedding model, respectively, to obtain patent technology word vectors, patent efficacy word vectors, standard technology word vectors, and standard efficacy word vectors. The similarity calculation module is used to calculate the first semantic similarity between the patent technology word vector and the standard technology word vector, and the second semantic similarity between the patent efficacy word vector and the standard efficacy word vector. The similarity filtering module is used to filter patent-standard technical word pairs with a first semantic similarity greater than the technical word similarity threshold and a second semantic similarity greater than the efficacy word similarity threshold, based on preset technical word similarity thresholds and efficacy word similarity thresholds. The association matching module is used to achieve patent-technology standard association matching based on filtering.

Images