Neural Open Schema Construction Pipeline for Knowledge Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing natural language processing (NLP) technologies face limitations in constructing knowledge graphs from unstructured text due to reliance on grammar-based information and pre/post-processing methods that are not domain-agnostic, leading to noise and incomplete information extraction.
Innovation Solution
A modular deep learning-based end-to-end pipeline for open schema construction (OSC) that performs coreference resolution, open information extraction, and entity canonicalization using hierarchical agglomerative clustering, enabling the creation of a domain-agnostic knowledge graph from unstructured text.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If grammar-based NLP algorithms are used for information extraction, then the processing is simpler and faster, but the extraction precision and recall are insufficient leading to noise and incomplete information
Solution Approach 1:
The patent replaces grammar-based mechanical NLP algorithms with deep learning-based neural networks that automatically learn patterns from data. The system uses pre-trained models (BERT, RoBERTa, GPT-3) that can process unstructured text and extract meaningful information without relying on rigid grammatical rules, thereby improving extraction precision while maintaining processing efficiency through optimized neural architectures.
2Measurement precision
If domain-specific models are used for knowledge graph construction, then the extraction is more accurate for specific domains, but the system becomes less adaptable to different domains and requires expert annotation
Solution Approach 1:
The patent implements a universal NLP pipeline that can process multiple domains simultaneously using pre-trained neural language models. The system uses domain-agnostic pre-trained models (BERT, RoBERTa, GPT-3) that have been trained on vast amounts of general text data, enabling them to adapt to various domains without requiring domain-specific retraining or expert annotation. The pipeline includes modular components for coreference resolution, open information extraction, and entity canonicalization that work across different domains.
Solution Approach 2:
The system dynamically adjusts processing parameters and model configurations based on the input domain. The pipeline can adapt to different domains by modifying extraction thresholds, entity type definitions, and relationship patterns while using the same core neural architecture. This allows the system to maintain high accuracy across diverse domains without requiring fundamental model redesign.
3Device complexity
If existing knowledge graphs are used for entity canonicalization, then the process is simpler, but noise is introduced due to non-domain specific information not being cleaned
Solution Approach 1:
The patent extracts and removes non-domain specific information and noise from the knowledge graph construction process. The system uses domain-specific filters and validation mechanisms that identify and eliminate irrelevant entities and relationships that would otherwise be introduced from general knowledge graphs. The pipeline includes a dedicated entity canonicalization module that cleans data by removing duplicates, resolving conflicts, and validating entities against domain-specific criteria before incorporating them into the final knowledge graph.
4Measurement precision
If hand-crafted schemas are used for knowledge graph construction, then the property definitions are more precise, but the model extractors become inflexible and require expert annotation
Solution Approach 1:
The patent implements dynamic schema construction where the knowledge graph schema is automatically adapted based on the input data and domain requirements. The system uses learnable entity type and relationship definitions that can evolve based on patterns detected in the data, rather than relying on fixed hand-crafted schemas. The pipeline dynamically generates schema elements during processing, allowing the model extractors to remain flexible and adaptable while maintaining precise property definitions through data-driven schema evolution.
Data Source
AI summary
Disclosed is a natural language processing pipeline that analyzes and processes a corpus of textual data to automatically create a knowledge graph containing the corpus entities such as subjects and object and their relationships such as predicates or verbs. The pipeline is configured as an end-to-end neural Open Schema Construction pipeline having a coreference resolution module, an open information extraction (OIE) module, and an entity canonicalization module. The processed textual data is input to a graph database to create the knowledge graph displayable through a graphical user interface. In operation, the pipeline modules serve to create a single term for all entity mentions in the corpus that reference the same entity through coreference resolution, extract all subject-predicate-object triplets from the coreference resolved corpus through OIE, and then canonicalize the corpus by clustering each entity mention to a canonical form for mapping to the knowledge graph and display.