Automated Knowledge Graph Construction via Random Walk Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The construction of knowledge graphs from ever-growing data is labor-intensive and requires highly skilled personnel, as existing methods rely on domain-specific parsers and infrastructure maintenance, limiting their applicability and efficiency.
Innovation Solution
A method involving the use of existing knowledge graphs to sample random walks, determine embedding vectors, and train machine-learning models to predict sequences of terms, allowing for the automated generation of new knowledge graphs without the need for extensive human intervention, using a sequence-to-sequence machine-learning model to process documents and generate new knowledge graphs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If domain-specific parsers and infrastructure are used to construct knowledge graphs, then manufacturing precision (knowledge graph construction accuracy) is improved, but device complexity and ease of manufacture deteriorate due to difficulty in defining, coding and maintaining parsers
Solution Approach 1:
The patent replaces the mechanical system of hand-coded domain-specific parsers with a machine-learning-based system. Instead of writing custom parsing code for each knowledge domain, the system uses pre-trained ML models that automatically learn to extract knowledge from unstructured data, thereby reducing infrastructure complexity while maintaining construction accuracy
Solution Approach 2:
The patent employs pre-trained machine-learning models that have been copied and applied across different knowledge domains. These models are trained on general knowledge graphs and can be reused without redefining parsers for each specific domain, simplifying the infrastructure while maintaining precision through the models' learned patterns
2Manufacturing precision
If domain-specific parsers are developed for one knowledge domain, then manufacturing precision for that domain is improved, but adaptability deteriorates because parsers cannot be used in a one-to-one fashion for another corpus or knowledge domain
Solution Approach 1:
The patent creates universal machine-learning models that can handle multiple knowledge domains simultaneously. Instead of building separate parsers for each domain, the system trains a single multi-domain model that adapts to different knowledge graphs through transfer learning, enabling one model to serve multiple functions across diverse corpora and domains
Solution Approach 2:
The patent utilizes parameter adaptation in machine-learning models, where the same model architecture can be fine-tuned by adjusting parameters for different knowledge domains. This allows the model to maintain its general structure while adapting to domain-specific requirements, thereby achieving both precision and versatility without redefining the entire parser infrastructure
3Manufacturing precision
If traditional parser-based methods are used, then knowledge graph construction can be performed, but productivity deteriorates due to labor-intensive processes requiring highly skilled personnel
Solution Approach 1:
The patent implements self-service through automated machine-learning systems that perform knowledge graph construction independently without requiring highly skilled human intervention. The ML models automatically extract, relate, and validate knowledge from unstructured data, performing the entire construction process autonomously and significantly improving productivity while maintaining quality through the models' intelligent processing
Data Source
AI summary
A method for building a new knowledge graph may be provided. The method comprises providing an existing knowledge graph, sampling random walks through the existing knowledge graph, determining embedding vectors for vertices and edges of the sampled random walks, and training of a machine-learning model taking as input sequences of the embedding vectors of the random walks. Furthermore, the method comprise receiving a set of documents determining sequences of terms from phrases from the documents the documents, building sequences of embedding vectors from the determined sequences of terms from the phrases, and using the built sequences of embedding vectors from the determined sequences of terms from the phrases as input for the trained machine-learning model for predicting second sequences of terms. Finally, the method comprises merging the predicted second sequences of terms thereby building the new knowledge graph.