Patent Classification Innovations

A method for accurately and efficiently classifying patents using artificial intelligence and natural language processing. The method involves training AI models to classify patents based on business language usage instead of the traditional hierarchical codes. The AI models learn from subject matter experts to understand how granted patents are actually used in industry. This allows more accurate classification of patents beyond just what they are. The business language classification system, called BVC, has other useful applications like patent audits for M&A.

Source

Patent early warning analysis method using natural language models to assist enterprises in analyzing patent information and improve efficiency. The method involves determining core classification phrases for patents using synonym chains and keyword segmentation. It then compares core phrases between patents to find repeated words and generates warning levels based on the number of repeated words. This allows identifying potential infringement risks between patents by comparing their core classification phrases.

Source

A method and system for predicting disruptive patents using intelligent models. The method involves identifying potentially disruptive technology themes from patent data using machine learning algorithms, scoring the themes using a disruptive technology measurement model, and selecting the top 10% with highest scores for further analysis. It combines SVM-LDA, indicator system construction, and cosine similarity algorithms to improve accuracy in identifying disruptive technologies in complex, uncertain environments.

Source

Patent classification method that improves accuracy by leveraging the interconnectedness between patents, inventors, and patent owners. The method involves extracting feature vectors for patents, inventors, and patent owners using separate networks. These feature vectors are then fused into a single representation vector and fed into a classification network. The separate view networks allow capturing distinct aspects of patent data, like text content, inventor expertise, and company background. The fusion step combines this multi-view information into a compact representation. This improves classification accuracy compared to single-view methods.

Source

Patent search tool using blockchain and AI to find more relevant prior art. The tool breaks down patent claims into limitations and compares them to patent text, link structures, and classifications to find the most relevant prior art. It uses a network of patents, citations, and classifications to identify prior art. The tool also has features like spam filtering and focused limitation searches.

Source

A method for efficiently classifying Chinese patent documents using a multi-feature fusion approach based on the TRIZ invention principle. The method involves dynamically representing Chinese patent texts using the ALBERT pre-trained language model. It then extracts local features using bidirectional convolutional neural networks and global contextual features using bi-directional GRUs with self-attention. The extracted features are fused to obtain a more comprehensive text representation for classification. This improves accuracy compared to traditional classification methods by capturing both local character and global contextual semantics of patent texts.

Source

A method for accurately and efficiently classifying scientific research projects into subject categories using machine learning. The method involves leveraging attention mechanisms to classify project content, and then using additional related information like project type and funding source to refine the classification. This two-stage classification process improves accuracy compared to just keyword matching. The method involves obtaining project details, passing the content through an attention-based network to get an initial classification, then feeding both the initial result and related info to a second network to refine the classification.

Source

Automated system for generating custom industry classifications for startups based on their descriptions. The system uses a hybrid deep learning-based hierarchical classifier to classify industries and products/services for startups using their descriptions. It leverages representation learning techniques to automatically convert industry and product descriptions into vector representations. Unsupervised matching is done between the vector representations to assign unmapped descriptions to known classifications. Supervised training is then done with startup descriptions to classify them into custom and standard industries.

Source

Method for accurately classifying the type of entities like companies and organizations even when they lack a unique identifier like a social security number. The method involves extracting features like keywords from the entity names and regular expressions from the coding categories to create a dataset for training a classification model. This allows classification of new entities without unique identifiers using the learned patterns from the known entities.

Source

Efficient patent classification method using semantic similarity to improve the speed and accuracy of patent classification compared to manual classification. The method involves extracting key features from patents and comparing them to pre-stored features at each classification level. If a feature matches, that level becomes the patent's classification. This leverages semantic similarity between features to classify patents without manual review.

Source

Multi-level patent text classification using graph neural networks to improve patent document categorization accuracy. The method involves a multi-stage classification process where the graph neural network extracts features from the patent text at each stage. In the first stage, it identifies broad classifications like technology areas. In subsequent stages, it further subdivides into more specific categories. This multi-level approach allows capturing both high-level and detailed classifications from the patent text.

Source

Semantic-based intellectual property text classification method using a trained appearance patent classification model. The method involves constructing a training set with patent names, drawings, and classifications. The model learns text, image, and fusion features. It adjusts parameters using loss functions based on the training data. This trained model is then used to classify new patent application texts using the learned features.

Source

Method for analyzing the subject content and popularity evolution of patented technologies by leveraging the International Patent Classification (IPC) system. The method involves using a topic modeling technique called Partial Latent Dirichlet Allocation (pLDA) to extract technical topics from patent documents. The pLDA model is modified to take into account the IPC classification level and abstract text of the patent. By setting the IPC level for topic mining, it allows fine-grained analysis of technical topics at different levels. In subject evolution analysis, word clouds are used to show the subject content under IPC classifications over time. The topic intensity and trend are calculated to determine the hotness trend of topics under IPC classifications.

Source

Patent evaluation method based on citation time difference and community classification to objectively rank patent importance. It constructs a patent citation network, performs unsupervised community classification on all nodes, counts the time difference between citations and cited patents, and assigns different weights to citation relationships based on community classification and citation time difference. The PageRank algorithm is used iteratively on the weighted network to obtain patent ranking. The method considers both citation quantity and quality, and factors like patent age and community context, to provide more objective patent evaluations compared to traditional indicators.

Source

Method for analyzing technology trends using patent data and classification systems to identify detailed technologies and their convergence, and objectively quantify technology development trends. The method involves extracting patent classification systems, configuring a matrix, clustering to form detailed technology groups, identifying technologies within each cluster, and applying time series analysis to track development trends.

Source

Patent recommendation method for assisting designers in expanding their knowledge space to improve conceptual design quality and efficiency. The method involves partitioning a network of patents into multiple domains based on their technical and semantic features. It then recommends patents related to design concepts by finding patents with similar concept combinations and evaluating their relevance and centrality in the domain network. This reduces the burden on designers to manually search and sift through many related patents to find specific ones.

Source

Analyzing businesses using patents and financial data to provide efficient and short-term business analysis when intangible assets like intellectual property are involved. The analysis involves extracting patent factors, weighting them, assigning scores to patent holders, extracting businesses based on financial data, weighting industries, clustering patents with industries, and outputting analysis results. This allows understanding patents and businesses as part of value chains.

Source

Automated patent document classification method that uses text processing techniques to improve efficiency and accuracy compared to manual classification. The method involves preprocessing patent documents by formatting and stop word removal. Then, initial classification by field is done based on patent classification numbers. Finally, core words are identified and classified using text algorithms and models based on their frequency in each document group.

Source

Keyword-based patent text classification to automate patent classification using machine learning instead of manual classification. The method involves training a patent classification model by extracting features from historical patents in a field, converting them into feature maps, and using those maps to train the model. When a new patent text is provided, it's analyzed using the trained model to classify it into the same field. This saves time compared to manual classification as the model can handle large numbers of patents more efficiently.

Source

An industry classification method based on patent big data to accurately and efficiently classify large numbers of companies across a wide range of industries. The method involves using patent data as a proxy to infer industry classifications for companies. It involves identifying sets of companies with similar patent profiles, then analyzing those sets to determine the industries represented. This allows scaling up classification beyond manual methods by leveraging the richness of patent data as an indicator of technical direction and business focus.

Source

Automatic processing of patent information to efficiently understand patent documents and their technical fields. The system uses artificial intelligence to extract synonyms for search terms, find similar patent documents containing those synonyms, and automatically classify the documents based on factors like filing year, country, and applicant nationality. This allows rapid classification and processing of patent information in a vast amount of documents.

Source

Automatic patent classification method that accurately classifies patents and predicts the probability of classification at each level. The method involves preprocessing patent text, running it through an independent classifier at each level, and combining the results to obtain overall classification with higher probability at upper levels. This leverages the hierarchical structure of patent classification.

Source

Automated system for classifying and visualizing patent research analysis data. The system provides a method for efficiently classifying and visualizing patent research analysis data using automation. It involves breaking down the process into components like project management, search, keyword extraction, and trend analysis. The components are integrated to allow users to start projects, search patents, extract keywords, and analyze trends. The automated system helps improve accuracy, accommodate diversity, and reduce manual effort compared to traditional methods.

Source

More accurate patent searching and analysis using AI, specifically natural language understanding (NLU) and machine learning (ML) techniques, to improve search results and avoid missing relevant patents. The method involves analyzing patent claims and relevant text from a reference document using AI systems to determine the meaning of claim elements and search for matches to rank patents based on how well their claims match the reference document.

Source

Method to classify patents as digital economy patents based on natural semantic processing. The method involves collecting patent text data, preprocessing the data, training a natural semantic processing model using labeled patent data, and then using the trained model to classify new patent documents as digital economy patents or non-digital economy patents. The natural semantic processing model analyzes the meaning and relationships of words in patent texts to identify those related to digital economy industries as defined by official classifications.

Source

Patent classification method using graph attention mechanism to accurately and efficiently categorize patents into multiple levels based on their text and label structure. The method involves segmenting patent text into chapters and paragraphs, extracting label definitions, constructing a graph with the segmented text and label definitions, applying graph attention mechanism to encode the text-label interactions, and classifying the labels using the encoded graph features. This leverages the internal structure of patents and label definitions to improve classification accuracy compared to whole-text methods.

Source

Automatic Chinese patent IPC classification using a bidirectional encoder representation from transformers (BERT) model called SBERT. The method involves encoding the patent text and IPC classification terms using BERT, and finding the similarity between the encoded vectors to determine the most relevant IPC categories. This leverages the BERT pre-training for semantic understanding and reduces calculation by encoding the patent text and classification terms separately.

Source

Automatically constructing learning data for training AI models to accurately classify patent documents based on their technological field. The method involves extracting company names from patent metadata, finding reference patents with those companies, finding similar patents to the references, calculating similarity between documents, and identifying patents with high similarity to the references as belonging to the extracted company's classification. This allows leveraging existing classification data for newly emerging patents by matching companies.

Source

A method for screening high-value patents using big data and AI to efficiently identify patents with potential for converting into valuable innovations. The method involves steps like setting up a patent database, cleaning and transforming the data, training AI models, and evaluating patents based on factors like patent family size, cited patents, litigation, claims, inventors, and innovation teams. This allows rapid screening of large numbers of patents to identify those with potential for significant innovation impact.

Source

Analyzing patent data using generalized additive models (GAM) to determine if a patent technology is sustainable. The method involves generating a patent matrix with keywords extracted from patent documents. GAM is used to calculate a significance probability (P-value) for each keyword. Regression plots are generated for keywords with low P-values. Slopes greater than zero indicate sustainable technologies. A diagram is generated to classify keywords based on slope.

Source

A machine learning-based patent evaluation method that provides detailed and quantitative patent value assessment using multiple categories subdivided based on split points. The method involves extracting quantitative information like bibliographic data and transfer fees from patents, embedding the textual data to convert it into numerical form, training patent classification models with the preprocessed data and split points, and using these models to evaluate the value of a target patent based on its embedded text and quantitative data. The models divide the patents into categories based on split points derived from sorting the training data by transfer fees.

Source

Automatically categorizing patent documents based on user input to improve consistency and reduce errors in patent analysis projects. The system allows users to manually categorize a first document and then automatically analyzes other documents to find similar ones and classify them with the same category. This leverages machine learning algorithms to identify documents with similar text characteristics to the initial categorized document. The system also provides a graphical user interface that displays project information, legacy analysis results, and interactive CPC hierarchies to reduce switching between views and improve productivity.

Source

Patent classification method, system and storage medium based on dual-channel feature fusion to improve the efficiency and accuracy of patent document classification. The method involves mapping patent abstract words into word2vec vectors and POS vectors, processing them with GRU sequences, concatenating the vectors, and further processing with fully connected layers. This dual-channel feature fusion of word and POS vectors enhances patent document classification compared to using just word vectors.

Source

Model training and patent knowledge map completion method to accurately and efficiently complete missing entities in patent knowledge graphs using semantic information from patent texts. The method involves training a patent knowledge graph completion model using intact patent texts with target completion entities. It extracts entities and relationships from the text, vectorizes them, clusters similar entities, and uses multi-label classification to train the model. To complete a knowledge graph, it uses the trained model to infer missing entities from a given entity and relationship. This leverages patent text semantics to avoid spreading abnormal entities like BERT language models.

Source

Improving the classification of multi-level patent texts using a BERT-ALMG model that captures contextual semantics of patent texts and hierarchical structure between labels. The BERT pre-training model analyzes patent texts for contextual semantics, and a label attention module maps text and label vectors to extract label-specific features. A multi-granularity feature extraction module extracts coarse and fine-grained features for classification. This allows better modeling of the hierarchical label structure compared to flat multi-label classification.

Source

Using machine learning to accurately predict patent values without relying on large amounts of labeled data or expensive computing resources. The method involves training a machine learning model using a sparse matrix of patent classification codes. The matrix has elements representing whether specific classification codes are present in a patent. The model is trained on patents with known values and then used to predict values for new patents based on their classification codes. This allows efficient patent valuation using minimal data.

Source

Patent classification method and system that provides fast, accurate, and low-cost patent text classification. The method involves a multi-module composite classification system that combines three modules: a rule matching module, a keyword matching module, and a BERT model. The modules are used sequentially, starting with the fastest, rule matching, then keyword matching, and finally the BERT model. This allows faster classification speed compared to using just BERT, while maintaining high accuracy. The modules are designed to complement each other, leveraging their respective strengths. The rule matching module uses predefined rules to quickly classify patents. The keyword matching module matches keywords against known classes. The BERT model provides fine-grained classification accuracy but is slower. By combining the modules in a specific sequence, the system balances speed, accuracy, and labor cost.

Source

Link prediction-based method for discovering element technologies that a company is likely to develop based on analysis of its existing patents. The method involves building a network of classification codes from granted patents, generating a company-specific network by reflecting its patents, and predicting missing connections using link prediction algorithms. This identifies candidate technologies that the company is likely to develop. The prediction accuracy is improved by calculating centrality and heterogeneity indices for the candidate technologies.

Source

A method to objectively assess patents using AI. The method involves obtaining information about the target patent and the corporation that owns it, generating input signals from the obtained data, feeding these signals into pre-trained neural networks to classify the patent and corporate, comparing the network outputs to stored comparison values to assess the patent and corporate, and providing the assessment results.

Source

Automatic patent classification system using deep learning models to improve accuracy and efficiency compared to manual classification. The system trains a Bert model for text classification and uses it to automatically classify new patent applications. It then performs a secondary verification by randomly extracting sections from the patent texts and rearranging them into new documents. These sub-texts are classified separately and the final classification is determined based on the original classification. This two-step verification improves the accuracy of the automatic classification.

Source

Automatically classifying patent documents using machine learning to reduce the effort required for manual classification. The method involves learning a user's patent classification criteria, predicting desired classification patterns, and then automatically classifying remaining unclassified documents using a basic classification model and predicted criteria. This allows users to easily manage patent durations and classification by leveraging machine learning to learn their unique classification habits.

Source

Using artificial intelligence to classify patents more accurately and consistently than manual methods. The AI models are trained to classify patents using a unique business-focused classification system called Business Language Codes (BVC). This involves training the AI to classify patents based on how they can be used in industry, rather than just categorizing the invention itself. The AI then uses this BVC classification to compare against the more traditional hierarchical CPC classification assigned by patent offices. This allows identifying variations and inconsistencies in patent classification. The AI can also leverage BVC classification to search for prior art and similar patents more effectively.

Source

Patent classification method using network representation learning and hierarchical label embedding to improve patent classification accuracy. It involves enhancing patent text features with label embedding and network propagation. The method preprocesses patent data, obtaining patent text, citations, inventor/right holder info, and labels. It embeds label descriptions and uses graph convolution to enhance label features based on hierarchical structure. Patent text is embedded too. Networks are built from inventor/right holder info. Features are learned from networks and fused with text features for final classification.

Source

A platform for conducting patent searches and analyzing patent value using a combination of AI, machine learning, and blockchain technology. The platform leverages natural language processing to improve patent search relevance by analyzing similarities between claim limitations and prior art specifications. It also uses AI algorithms to rank prior art based on relevance to specific claim elements. The platform further provides tools for highlighting, weighting, and customizing search parameters.

Source

Automated method of evaluating an attribute of patent contents using an artificial neural network to evaluate patents. The evaluation factors are selected based on the technology class of the patent. The attribute evaluation feature value is calculated using input scores and predetermined connection weights. The weights are limited based on class to avoid vanishing gradients. This allows differentiated automated patent evaluation for different technical fields.

Source

Training and using machine learning models for efficient and accurate overall classification of patents. The method involves labeling training samples with both patent features and overall planning categories. A model is trained using these labeled samples. For unlabeled patents, a technique combines feature matching and category comparison to preliminarily classify the patents into overall planning categories. This preliminary classification is then refined using the trained model. The refined classification provides more accurate and efficient overall category assignment for patents compared to manual labeling.

Source

Automatically classifying patents via machine learning to save time and resources. The method involves learning from a patent database to establish a basic classification model. When a user searches for patents, it uses their classification input to predict their personalized classification standard. This combines with the basic model to classify the remaining unclassified patents. It leverages transfer learning to customize the basic model using the user's classification pattern.

Source

A method for classifying patent documents related to financial innovation using a BERT (Bidirectional Encoder Representations from Transformers) model. The method involves preprocessing patent documents, feeding them into a BERT model trained on labeled financial innovation patents, and using the BERT model to classify new patent documents as either belonging to the financial innovation category or not. The BERT model is fine-tuned on labeled financial innovation patents to improve its classification performance.

Source

Automatically classifying patent documents into technology categories based on user-defined criteria with improved accuracy using a technique called supervised learning. The method involves clustering the patent documents into groups, selecting representative samples from each cluster, having the user define the technology categories for those samples, learning a classification model from the labeled samples, and using the learned model to classify the entire population of documents. This allows accurate classification of the entire population with a small amount of labeled data.

Source

Small sample patent classification method using feature maps that allows accurate and efficient patent document classification with limited training data. The method involves constructing feature maps representing patent documents using a vocabulary. These feature maps are used to train a neural network model for patent classification. The feature maps provide a compact representation of patent documents compared to raw text, which improves the neural network's ability to learn from limited training data.

Source